Citation
Understanding the expansion and effects of Colorado's concurrent enrollment program

Material Information

Title:
Understanding the expansion and effects of Colorado's concurrent enrollment program
Creator:
Dickhoner, Brenda Bautsch ( author )
Place of Publication:
Denver, Colo.
Publisher:
University of Colorado Denver
Publication Date:
Language:
English
Physical Description:
1 electronic file (163 pages) : ;

Thesis/Dissertation Information

Degree:
Doctorate ( Doctor of philosophy)
Degree Grantor:
University of Colorado Denver
Degree Divisions:
School of Public Affairs, CU Denver
Degree Disciplines:
Public affairs

Subjects

Subjects / Keywords:
Education -- Colorado ( lcsh )
Education ( fast )
Colorado ( fast )
Genre:
bibliography ( marcgt )
theses ( marcgt )
non-fiction ( marcgt )

Notes

Review:
One of the prominent approaches among states to improve college access and success is concurrent enrollment, which provides high school students the opportunity to enroll in a college course for which they may receive both high school and college credit. This study set out to understand, first, what factors lead some schools to adopt concurrent enrollment more quickly and implement the program more intensely as compared to other schools. The study also sought to evaluate how effective concurrent enrollment is at improving college access and success for all students, including low-income and minority students. The dissertation finds that fiscal capacity, organizational capacity, school type and prior program offerings are key predictors of the adoption and implementation of concurrent enrollment programs. Additionally, participation in concurrent enrollment in high school results in positive gains in college enrollment rates, first-year grade point averages, and college persistence rates, and results in a decrease in the need for remedial education. While concurrent enrollment, on average, improves college outcomes for all students, low-income students experience a greater positive impact on their outcomes than higher income students. Moreover, Hispanic students who take concurrent enrollment courses see a greater impact on their likelihood of going to college than white students who participate in the program.
Bibliography:
Includes bibliographical references.
System Details:
System requirements: Adobe Reader.
Statement of Responsibility:
by Brenda Bautsch Dickhoner.

Record Information

Source Institution:
University of Colorado Denver
Holding Location:
Auraria Library
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
on10123 ( NOTIS )
1012399978 ( OCLC )
on1012399978
Classification:
LD1193.P86 2017d D53 ( lcc )

Downloads

This item has the following downloads:


Full Text
UNDERSTANDING THE EXPANSION AND EFFECTS OF
COLORADO'S CONCURRENT ENROLLMENT PROGRAM
by
BRENDA BAUTSCH DICKHONER B.A., Duke University, 2006
A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment of the requirements for the degree of Doctor of Philosophy Public Affairs Program
2017


This thesis for the Doctor of Philosophy degree by
Brenda Bautsch Dickhoner has been approved for the Public Affairs Program by
Todd Ely, Chair Paul Teske Kelly Hupfeld Matt Gianneschi
Date: May 13, 2017


Dickhoner, Brenda Bautsch (PH.D., Public Affairs Program)
Understanding the Expansion and Effects of Colorado's Concurrent Enrollment Program Thesis directed by Assistant Professor Todd Ely
ABSTRACT
One of the prominent approaches among states to improve college access and success is concurrent enrollment, which provides high school students the opportunity to enroll in a college course for which they may receive both high school and college credit. This study set out to understand, first, what factors lead some schools to adopt concurrent enrollment more quickly and implement the program more intensely as compared to other schools. The study also sought to evaluate how effective concurrent enrollment is at improving college access and success for all students, including low-income and minority students. The dissertation finds that fiscal capacity, organizational capacity, school type and prior program offerings are key predictors of the adoption and implementation of concurrent enrollment programs. Additionally, participation in concurrent enrollment in high school results in positive gains in college enrollment rates, first-year grade point averages, and college persistence rates, and results in a decrease in the need for remedial education. While concurrent enrollment, on average, improves college outcomes for all students, low-income students experience a greater positive impact on their outcomes than higher income students. Moreover, Hispanic students who take concurrent enrollment courses see a greater impact on their likelihood of going to college than white students who participate in the program.
The form and content of this abstract are approved. I recommend its publication.
Approved: Todd Ely


IV
DEDICATION
For Blair, whose love and support means everything. And for GraysonI hope you always pursue your dreams no matter how long the road ahead seems.


V
ACKNOWLEDGEMENTS
I am extraordinarily indebted to Dr. Todd Ely, who provided advice, guidance and mentorship over the past six years. Dr. Ely has an enviable aptitude for statistics and challenged me to explore various quantitative methods in an effort to carry out a rigorous and respectable research design. I learned more than I could have imagined, thanks to the patient facilitation of Dr. Ely. He even made the process enjoyableas much as such a process can be enjoyed. Drs. Paul Teske and Kelly Hupfeld lent their public affairs and education policy expertise to provide valuable feedback, particularly in the beginning stages as I was preparing what would be the roadmap for my research. I am grateful that Dr. Teske, Tanya Heikkla, Chris Weible, and Peter deLeonalong with many othershave created such a wonderful and welcoming PhD program for practitioners. The School of Public Affairs faculty encourages the blending of theory with practical application and warmly accepts practitioner students such as myself into their scholarly sphere.
I am incredibly grateful that Dr. Matt Gianneschi served on my committee as my outside reader. Dr. Gianneschi helped author the legislation that created Colorado's concurrent enrollment program and has a wealth of knowledge about education policy through his roles in state government, in the policy sector and as a college leader. Dr. Gianneschi was also one of the individuals who helped me land on the topic of concurrent enrollment; without him and Dr. Beth Bean I might still be wandering the doctoral wilderness in search of worthy topic. I am appreciative of Dr. Bean for not only helping me find a topic and a rich data set, but also for providing moral support as I worked for her at the Colorado Department of Higher Education (CDHE). Maggie Yang, Michael Vente and all of the CDHE staff were tremendously helpful and patient with my multiple data requests. Michelle Camacho Liu, who was the state's concurrent enrollment administrator while I was at CDHE, shared an abundance of knowledge with me to help inform the background,


VI
context and discussion portions of this dissertation. Michelle also happens to be a dear friend, and I am so grateful for her support and friendship in addition to the concurrent enrollment insight.
I owe a great deal of gratitude to Dr. Julie Bell at the National Conference of State Legislatures, who was my boss when I had the crazy idea to enter a Ph.D. program. Dr. Bell encouraged me to apply and supported my acceptance into the program with her letter of recommendation. She also permitted me to work a flexible schedule as I completed my coursework. Thank you, Dr. Bell, for your belief in meI truly would not be at this milestone without you.
Alyssa Pearson at the Colorado Department of Education has been an amazing friend and supervisor this past year as I have completed and defended my dissertation. Thank you, Alyssa, for your unwavering support, overflowing optimism, delicious baked goods, generosity of spiritand for being an inspiration to all! You are an excellent role model for public administrators everywhere.
Last, although certainly not least, I want to acknowledge my family and friends who have supported me on this long road. My parents instilled in me a love for education at a young age. My dad made it possible for me to attend the college of my dreams, and my mom made it possible for me to persist through graduate school. She helps take care of Grayson and moe, brings me food any time I need it (and when I don't), and is always there for me to lean on. My mom once said she would never be satisfied until I earned my doctorateso I am pleased to finally meet her lofty expectations. Thank you, Mom and Dad, for all you do and for valuing education so much!
My husband, Blair, has been my rock and is the reason I've made it to the finish line. While this process has been long and grueling at times, one positive outcome is that Blair was able to pursue multiple hobbies while I worked, including guitar playing, marathon running and beekeeping. I will be the proud recipient of a doctoral diploma and homemade honey! Finally, I am lucky enough to have many friendstoo many to namewho grabbed a drink with me when I needed one and understood when I had too much work to go get a drink. Thank you, all!


vii
TABLE OF CONTENTS
I: INTRODUCTION.........................................................................1
Problem Significance..................................................................3
Concurrent Enrollment Policy Landscape................................................6
Colorado's Concurrent Enrollment Programs Act........................................14
Contributions to the Field...........................................................17
Summary & Research Questions.........................................................22
II: LITERATURE REVIEW AND HYPOTHESES....................................................24
Policy Diffusion & Innovation Theory.................................................24
Education Theory.....................................................................30
Summary...............................................................................40
III: DATA & METHODS.....................................................................41
Data Sources and Collection..........................................................41
Research Design......................................................................44
Data & Methods: Summary..............................................................69
IV: POLICY DIFFUSION FINDINGS & DISCUSSION..............................................71
Descriptive Statistics...............................................................71
Event History Analysis...............................................................77
OLS Fixed Effects Regression Analysis................................................80
Dynamic Panel Data Model.............................................................83
Conclusion & Discussion..............................................................87
V: POLICY EVALUATION FINDINGS...........................................................93
Descriptive Statistics
93


viii
Effects of Concurrent Enrollment Participation on College Outcomes......................95
Concurrent Enrollment Effects for Low-Income Students and Minority Students............104
Effects of Concurrent Enrollment Credit Hour Levels on College Outcomes................112
Conclusion & Discussion................................................................117
VI: CONCLUSION............................................................................120
Key Findings...........................................................................122
Implications for Research and Practice.................................................123
Limitations and Future Research........................................................135
REFERENCES................................................................................141


IX
LIST OF TABLES
Table 1. Thematic Analysis of Model State Policy Elements and Standards......................10
Table 2. Summary of Research Questions and Hypotheses........................................40
Table 3. Concept Measurement Summary: Policy Diffusion.......................................45
Table 4: Variable Descriptions and Sources...................................................46
Table 5: Concurrent Enrollment Adoptions and Survivor Functions, by School Year..............52
Table 6. Concept Measurement Summary: Policy Evaluation......................................58
Table 7. Descriptions of Pre-College Independent Variables and College Outcome Variables.....59
Table 8: Methodological Approaches with Associated Research Questions........................69
Table 9: Descriptive Statistics for All High Schools, Beginning and End of Study.............73
Table 10: Comparison of Variable Means, by High School Adoption Year.........................74
Table 11: Cox Proportional Hazards Model Results.............................................78
Table 12. Predictors of Student Participation Rates in Concurrent Enrollment (CE)............81
Table 13. Dynamic Panel Data Model using Maximum Likelihood for Concurrent Enrollment (CE)
Participation Rates in High Schools.........................................................84
Table 14: Summary of Statistically Significant Results across Methods and Hypotheses........86
Table 15: Descriptive Statistics for Overall Sample and by Concurrent Enrollment (CE) Participation94
Table 16. Propensity Score Matching Average Treatment Effects...............................98
Table 17. Progression of Logistic Regression Models Estimating the Effect of................101
Table 18. Average Treatment Effects.........................................................102
Table 19. Comparison of Average Treatment Effects...........................................104
Table 20. Progression of Logistic Regression Models Estimating the Effect of Concurrent Enrollment Participation on College Matriculation......................................................106


X
Table 21. Regression Models Estimating the Interaction Effects of Concurrent Enrollment
Participation on College Outcomes...............................................................108
Table 22. Credit Hours Descriptive Statistics for Concurrent Enrollment Students...............113
Table 23. Sample Means of Key College Outcomes by Concurrent Enrollment Credit Hours...........113
Table 24. Progression of Regression Models Estimating the Effect of Concurrent Enrollment
Participation on College Matriculation..........................................................114
Table 25. Average Treatment Effects of Credit Hours Levels on College Outcomes.................115
Table 26. Comparison of Statewide Evaluations Assessing Effect of Dual Enrollment Programs on College Matriculation...........................................................................133


XI
LIST OF FIGURES
Figure 1. Number of Adopted Bills Pertaining to Dual Enrollment Programs across the U.S., by Year. 9
Figure 2. Distribution of Propensity Score Across Treatment and Comparison Groups...................66
Figure 3. Adoption of Concurrent Enrollment Programs from the 2010-11 School Year to the 2014-15
School Year, by School Districts and High Schools...................................................72
Figure 4: Average Percentage of High School Students Participating in Concurrent Enrollment (CE)
within High Schools, by Adoption Year Cohort from 2010-11 to 2014-15................................75
Figure 5. Maps of Colorado high schools and Concurrent Enrollment (CE) participation rates by covariates of interest...............................................................................76
Figure 6. Cox Proportional Flazards Regression Smoothed Flazard Functions for Charter Schools and
College Matriculation Rates......................................................................80
Figure 7. Participation in Concurrent Enrollment, by Graduation Year, Gender and Race/Ethnicity.. 95 Figure 8. Standardized bias differences (%) across all covariates in original and matched samples... 97 Figure 9. Probability of College Matriculation, by Concurrent Enrollment Participation and Free or
Reduced-Price Lunch (FRL) Status and Race/Ethnicity (Flispanic or white).........................109
Figure 10. Probability of College Remediation, by Concurrent Enrollment Participation and Free or
Reduced-Price Lunch (FRL) Status.................................................................110
Figure 11. Probability of College Persistence, by Concurrent Enrollment Participation and Free or
Reduced-Price Lunch (FRL) Status
112


1
CHAPTER I INTRODUCTION
In today's economy, higher education is increasingly necessary to have a productive career and earn family-sustaining wages (Carnevale, Smith & Strohl, 2013). Access to a high-quality K-12 education that prepares students for postsecondary education, however, is not a guarantee in America's school system. On average, low-income and minority students consistently have lower levels of academic achievement than their peers at all points along the education pipeline, including high school graduation, college enrollment, and college degree attainment (Bettinger & Long, 2005; Darling-Hammond, 2010; Kahlenberg, 2004; Terenzini, Cabrera, & Bernal, 2001; U.S. Department of Education [USDOE], 2006). States across the country have implemented countless policies to better prepare students for life after high school, but achievement gaps persist.
Colorado, which has the second largest gap in the country in the college degree attainment between majority and minority students (NCHEMS, 2013), is no exception. Several laws passed by the Colorado legislature in the last decade have targeted improving the transition from high school to college.1 The question remains, though, as to how effective are those laws at improving educational outcomes, particularly when policies create voluntary programs for schools and students. Colorado's concurrent enrollment law, for example, was specifically designed to improve college readiness for traditionally-underserved students by bolstering access to rigorous, college-level coursework (C.R.S. §22-35-101). Under the policy, qualified students in grades 9 through 12 can take tuition-free college courses at their high school, a postsecondary institution, online, or in a hybrid format and simultaneously earn high school and college credits (CDE, 2010). This law creates
1 See, for exampleSB08-212: Preschool to Postsecondary Education Alignment Act (Colorado Achievement Plan for Kids); SB 09-256: Individual Career and Academic Plans; HB09-1319: Concurrent Enrollment Programs Act; HB07-1118: High School Graduation Requirements; SB09-163: The Education Accountability Act; HB 12-1155: Supplemental Academic Instruction.


2
the operational frameworkthe funding mechanism, participation requirements, and oversight for the concurrent enrollment program. It is a voluntary initiative, howeverschools can choose whether or not to adopt the program.
Proponents of concurrent enrollment argue that it increases academic preparation for college and provides momentum toward degree attainment by giving students the opportunity to enter college with credits already accumulated (An, 2013; Hoffman, 2005). Prior research has found positive associations between concurrent enrollment participation and college access and success outcomes (Allen & Dadgar, 2012; An, 2013; Giani et al., 2014; Taylor, 2015). Often the previous research has focused on small-scale, institution-specific programs and used imperfect methods. Consequently, rigorous, empirical analyses of state-wide programs are still needed (Allen & Dadgar, 2012; Bailey & Karp, 2003; Blanco, 2006; Giani, Alexander, & Reyes, 2014; Hoffman, 2012; Rutschow & Schneider, 2011). Colorado's concurrent enrollment program provides fertile ground for such research. The purpose of this study is to examine the effects of Colorado's concurrent enrollment program on college access and success, as well as to analyze decisions by high schools to offer concurrent enrollment programs and by students to enroll in them.
To address these questions, the dissertation begins with an introduction to the problems under investigation and background on how concurrent enrollment state policies purport to solve those problems, both nationwide and in Colorado. This introductory chapter concludes with a summary of contributions the study will make to research and practice and sets forth formal research questions. Chapter Two provides a review of relevant literature from the public affairs and education domains and presents testable hypotheses. Chapter Three includes a description of the data collection, an explanation of variables and measures, and detailed review of the various methods employed to answer the research questions. Chapters Four and Five present findings from the empirical research, with Chapter Four focusing on an analysis of factors that influence the


3
adoption of concurrent enrollment programs at the school level; Chapter Five focuses on an analysis of the effects of participating in concurrent enrollment on college matriculation and success at the student level. Chapter 6 summarizes the study and its implications for research and practice.
Problem Significance
Achievement Gaps
Low-income and minority students, on average, lag behind their peers on nearly every important education milestone (An, 2012; Bettinger & Long, 2005; Darling-Hammond, 2010; Kahlenberg, 2004; Oakes, 2005). Children from low-income families, for example, are more likely to have lower reading abilities by the third grade than high-income students (Hernandez, 2011). Achievement data from the National Assessment of Educational Progress (NAEP) shows that black and Hispanic students, on average, score two grade levels below white students when taking the NAEP exam in 4th and 8th grades (USDOE, 2009, 2011). Low literacy levels in early grades have been linked to diminished achievement in later years, including decreased high school graduation rates (Hernandez, 2011). Early indicators are important to measure because low-income students are about five times as likely to drop out of high school as high-income students (Kahlenberg, 2004).
In high school, disparities in curriculum offerings and quality of instruction remain a significant problem, with low-income and minority students disproportionally receiving lower-quality instruction and fewer advanced course options. Oakes (1993, 2005) found that even after controlling for test scores, white and Asian students are far more likely to be placed into honors courses than their peers. High-achieving Latino students who scored at the 90th percentile on standardized tests had just a 56 percent chance of being assigned to a college preparatory class, as compared to a 97 percent chance for Asian students and 93 percent chance for white students
scoring in the same percentile (Oakes, 1993, 2005).


4
Due to a variety of factors, including lack of access to consistently high-quality instruction and rigorous curriculum, achievement gaps that can be observed as early as pre-Kindergarten persist for many children throughout their entire educational careers. The transition from high school to college is no exceptionlow-income and minority students are less likely to enroll in or graduate from college than their white, affluent peers (Adelman, 2006; An, 2012; Kahlenberg, 2004). For low-income, minority students who do attend college, they tend to be less academically prepared than their peers; studies on the relationship between income and race/ethnicity and college remediation rates indicate persistent achievement gaps (see, e.g., Bettinger & Long, 2005).
In Coloradothe focus of this study82 percent of African American students and 70 percent of Hispanic students need remediation at community colleges, as compared to 50 percent of white students (Colorado Department of Higher Education, 2016). Also, in Colorado, 53.4 percent of low-income students are not ready for college-level courses in at least one content area, as compared to 31.4 percent of wealthier students (Colorado Department of Higher Education, 2016).
The fact that half of all white high school graduates who immediately attend a community college are not academically prepared is indicative of systemic challenges in readying our young adults for postsecondary education. That statistic already discounts the numerous students who dropped out of high school or those who graduated high school but chose not to matriculate to college. Further, while the remedial education rate for white students is concerning in and of itself, having remedial education rates that are 20 to 30 percentage points higher for minority students is an alarming trend.
Returns to Education
Closing achievement gaps, particularly around college access and success, remains a significant imperative for society from an equity perspective, as well as from an economic perspective. If achievement gaps persist, then the U.S. society and economy will continue to


5
experience negative externalities stemming from lower individual quality of life. Research has time and again found that individuals without a college credential are far more likely to face severe challenges throughout life including joblessness, welfare, incarceration, family instability and health problems (Hout, 2012; Kingston et al., 2003). These challenges are costly and burdensome to the taxpayers who subsidize prisons, social support systems and healthcare. Researchers, however, have also long questioned the notion of whether education causes better outcomes or simply reflects advantages bestowed upon certain individuals as a matter of chance.
Nonetheless, there is substantial empirical evidence that education provides positive returns on investment for individuals. The literature, for example, on wage premiums for attending college has consistently found that individuals accrue increased earnings for additional years of education using a variety of statistical approaches to control for selection bias, including instrumental variables and natural experiments (Angrist & Krueger, 1992; Hausman & Taylor, 1981; Hout, 2012, Kane & Rouse, 1995). More recent research has also found that the benefits of higher education are greater for those who are less likely to attend and graduatethat is, students who typically perform somewhere in the middle of the spectrum of academic ability (Attewell & Lavin, 2007; Brand & Xie, 2010; Hout, 2012; Maurin & McNally, 2008). While students of higher ability may graduate from college at higher rates and earn higher wages, their education has a lesser effect on their success than students of lower academic ability who gain greater wage premiums from higher education (Brand & Xie, 2010; Hout, 2012). This strand of literature has important implications for policymakers in that it supports continued efforts by states to expand higher education access to students who are at risk of not attending.
There is also empirical evidence that societal and economic benefits accrue when higher education completion rates increase. Some studies have found that increasing the number of
college graduates in a labor market raises the productivity levels of less-educated workers and may


6
also increase their wages (Moretti, 2012; Mas & Moretti, 2009). Researchers have also linked college graduates with higher rates of volunteerism and positive views of civil liberties and minorities (e.g. Brand, 2010; Kingston et al., 2003). Putnam (1995, 672), for example, declares that "education is by far the strongest correlate that I have discovered of civic engagement in all its forms."
From the economic perspective, labor economists project that jobsin particular, those that provide family-sustaining wageswill increasingly require postsecondary credentials. The labor demand for college educated workers is projected to surpass supply by 2020, which could stymie economic growth (Carnevale, Smith & Strohl, 2013). As stated in a 2010 report by Georgetown's Center on Education and the Workforce:
Essentially, postsecondary education or training has become the threshold requirement for access to middle-class status and earnings in good times and in bad. It is no longer the preferred pathway to middle-class jobsit is, increasingly, the only pathway. (Carnevale, Smith & Strohl, 2013, 13)
As this short review indicates, there is a compelling case for expanding higher education opportunities to more students. Policymakers often understand this and thus have turned their attention in recent years to expanding college access through concurrent enrollment. The following section provides an overview of the national policy landscape surrounding concurrent enrollment.
Concurrent Enrollment Policy Landscape Concurrent enrollment is a term used in 15 states, including Colorado, to refer to opportunities for high school students to enroll in a college course for which they may receive both high school and college credit. Unlike other accelerated learning options such as Advanced Placement (AP), students earn college credit if they receive a passing grade in the coursejust as a college student wouldrather than by earning a certain score on an end-of-course exam (Allen,
2010). This provides a stronger guarantee that the course credit will count toward the student's


7
college degree. Forty states use the terms "dual enrollment" or "dual credit" to refer to the same arrangement;2 the terms are used interchangeably in the following section.
Concurrent enrollment programs have been available in public high schools for at least the last half-century, mostly as an enrichment opportunity for academically-advanced students. Programs have grown exponentially since the early 2000s when certain policymakers began expanding concurrent enrollment opportunities to students who are traditionally underserved, including students of color and low-income students, as well as to students who are not high academic performers (Hoffman, Vargas, & Santos, 2008a).
In the 2001-02 school year, public high schools across the country reported approximately 1.2 million enrollments in dual credit courses (Kleiner & Lewis, 2005). That number is a duplicated student countit is inclusive of each course enrollment during the school year. A decade later, during 2010-11 school year, dual enrollment participation at public high schools increased to just over 2 million (Thomas, Marken, Gray & Lewis, 2013). In 2001-02, 71 percent of high schools had dual enrollment programs; by 2010-11 that figure increased to 82 percent.
Concurrent Enrollment Promises and Challenges
Concurrent enrollment is promising to policymakers and practitioners because it is seen as a way to expose more students to rigorous curriculum that high schools may be lacking. Providing students exposure to college is thought to be a strategy for developing metacognitive skills3, readying students for the demands of college life, and increasing college aspirations. Policymakers are also drawn to concurrent enrollment as a way to increase college affordability by offering college courses at low or no cost to families.
2 In some instances, multiple terms are used within states.
3 Students with well-developed metacognitive learning skills will be able to manage their time effectively, think critically, navigate college resources, maintain study routines, have self-awareness of their strengths and weaknesses, analyze and interpret information, and have the confidence to overcome challenges (Conley, 2010, 2013).


8
The challenges to policy implementation, however, are also multifold. While state policies around concurrent enrollment have proliferated, expanding access to low-income students and students of color still remains a challenge. Further, while many states are attempting to increase access by ensuring there are no costs to students, state budgets are continually under constraint, leaving little dedicated funding available for concurrent enrollment. Even in states where students do not shoulder tuition costs, school districts and colleges still need to establish a financially viable model for operating the program. Cash-strapped states, districts and colleges increasingly have to find creative ways to fund concurrent enrollment programs or risk scaling back access (Borden et al, 2013; Zinth, 2014b, 2015b).
Another challenge is ensuring course rigor and quality when courses are taught at a high school or online, as opposed to on the college campus. Offering courses in a high school setting greatly expands access and eases the logistical hurdles of transportation and scheduling for off-campus courses, but it requires more oversight to ensure consistency of rigor (Borden et al, 2013; Lowe, 2010; Zinth, 2015a). It is also challenging to find high school teachers with the necessary qualifications to teach concurrent enrollment courses, especially in rural areas (Zinth, 2014a). These challengesand promiseshave spurred a great deal of legislative activity in recent years.
State Policy
According to the Education Commission of the States (ECS), as of 2016, 47 states have statutes and/or regulations in place governing dual enrollment programs (ECS, 2017). However, a great deal of variation among the 47 state policies exists regarding funding, eligibility, course type, instructor qualifications, general oversight and monitoring, and credit transferability (Borden et al., 2013). Further, state policies continue to evolve as states make modifications to their programs in these areas. According to data collected by ECS, over the past five years alone, 143 bills were adopted by state legislatures concerning dual enrollment programs; in the last ten years, states


9
passed a total of 243 bills (ECS, 2017). Figure 1 displays the number of bills signed into law by states across the country per calendar year. The chart shows low points (2008 and 2010) and high points (2013), but depicts that the number of adopted bills has remained near the average of 24 bills in most years over the past decade. Legislative changes have focused on clarifying or expanding funding streams, integrating career and technical education opportunities, promoting options that increase the number of qualified instructors, modifying student eligibility requirements, and implementing provisions to help ensure dual credit courses are as rigorous as traditional college courses.
38
z 0
2007 2008 2009 2010 2011 2012 2013 2014 2015 2016
Figure 1. Number of Adopted Bills Pertaining to Dual Enrollment Programs in the U.S., 2007-2016.
Data collected from the Education Commission of the States (ECS) State Policy Database, retrieved February, 11, 2017.
Model Policy Elements
With the high level of legislative activity around dual enrollment, policy researchers have delved into the numerous state policies and, based on other research and best practices, have identified key components that states should include in their dual enrollment policies. This section reviews three prominent sets of model policy elements and program standards, which are synthesized in Table 1. ECS and Jobs for the Future (JFF) have issued specific guidance for policymakers. The National Alliance of Concurrent Enrollment Partnerships (NACEP) issued guidance
focused on program oversight and ensuring academic rigor.


Table 1. Thematic Analysis of Model State Policy Elements and Standarc s
Themes Jobs for the Future "Model State Policy Elements" NACEP "Accreditation Standards" Education Commission of the States "Model Components of State-Level Policies"
Program Quality: Course rigor, instructor qualifications and course credit Quality Assurance States should ensure that college courses offered to high school students use the same syllabi and exams as comparable courses taught on a college campus The postsecondary institution conferring credit should set the qualifications for faculty teaching dual credit courses Dual enrollees earn both high school and college credits upon successfully completing courses Curriculum College courses offered in the high school are of the same quality and rigor as the courses offered on-campus at the college/university Faculty Concurrent enrollment instructors meet the academic requirements for faculty and instructors teaching in the sponsoring postsecondary institution and are provided discipline-specific professional development Assessment Students enrolled in concurrent enrollment courses are held to the same standards of achievement as students in on-campus courses, including grading standards and assessment methods Ensuring Course Quality Courses meet same level of rigor as traditional college courses Instructors meets same expectations as college faculty and receive support Transferability of Credit Postsecondary Institutions accept dual enrollment credit as transfer credit, provided measures for quality are ensured
Financial Provisions $$$ Sustainable Funding and Finance States should develop funding policies that: Allow high school students to take college courses free of tuition and non-course-related charges Permit both districts and postsecondary institutions to claim per pupil funding allocations to support the cost of offering dual credit courses Finance Responsibility for tuition does not fall to parents Districts and postsecondary institutions are fully funded or reimbursed for participating students


Table 1 (cont.)
Themes Jobs for the Future "Model State Policy Elements" NACEP "Accreditation Standards" Education Commission of the States "Model Components of State-Level Policies"
Student Access and Support Eligibility and Access A state's eligibility requirements are determined by the secondary and postsecondary sectors together Students have multiple ways to demonstrate readiness, including a combination of tests, end-of-course grades, teacher recommendations, and work portfolios. Academic and Social Supports States should require that districts colleges specify/document key roles and responsibilities in memoranda of understanding or cooperative agreements, including the provision of a college liaison for student advisement and support States should provide support and funding for programs serving students who are overage and under-credited and youth who have dropped out of high school Students Students officially register with a college Students meet the college's course pre-requisites The concurrent enrollment program provides students with a handbook of rights/responsibility of college students Access All eligible students may participate, based on demonstration of ability to access college-level content Caps on the maximum number of courses allowed should not be overly restrictive Students earn high school and college credit for successful completion of approved postsecondary courses All students and parents are annually provided wish program information Counseling is made available to students and parents before and during program participation
Reporting and Program Evaluation System for Accountability States should report annually on dual enrollment participation and impact and develop administrative structures to support program leaders and dual enrollment partnerships. States should also designate a state board or governing body as having the authority and responsibility to guide dual enrollment policy. Aligned Data Systems States should develop unit-record statewide data systems that identify dual enrollees by demographic characteristics and monitor student progress longitudinally across the K-12 and higher education systems Program evaluation Concurrent enrollment programs display greater accountability through required impact studies, student surveys, and course and program evaluations Ensuring Course Quality (cont.) Districts and institutions publicly report on student participation and outcomes Programs undergo evaluation based on available data
Sources: NACEP, 2011; Ward & Vargas, 2012; Zinth, 2014b.


12
NACEP "works with state legislators, agencies, and college and university systems to develop quality concurrent enrollment partnerships and hold them accountable to high standards" (NACEP, 2017). The organization administers the only national set of quality standards for concurrent enrollment programs, which it uses to accredit individual postsecondary institutions that offer concurrent enrollment programs across the country. NACEP advocates for states to use the standards as a quality measure in statewide concurrent enrollment programs. There are currently 17 states that have modeled their quality standards (as set in statue or regulation) on the NACEP standards, including Colorado. The standards are categorized around curriculum, faculty, students, assessment and program evaluation and are geared towards ensuring that courses taught by high school teachers, in particular, are as rigorous and high-quality as courses taught by postsecondary faculty on college campuses (NACEP, 2011).
ECS identifies 13 policy components organized under the categories of access, finance, ensuring course quality and transferability of credit (Zinth, 2014b). The guidance to policymakers notes that the policy components were selected because they "may increase the likelihood that a more diverse group of students successfully participates in high-quality dual enrollment courses and receives credit that will be transferable to other public postsecondary institutions in the same state" (Zinth, 2014b, 4).
Jobs for the Future undertook their policy scan with a lens similar to that used by ECS, but focused more on the key policy components needed to close achievement gaps. The organization posits that state policies have the potential to facilitate meaningful partnerships between high schools and colleges that result in a seamless transition into higher education for students who might not otherwise attend. Of the 47 statewide policies JFF reviewed, however, they found that "only a few have established sufficient mechanisms to ensure that all students, including those underrepresented in higher education, have access to these vital pathways to college" (Ward &


13
Vargas, 2012, 4). The six mechanisms JFF identified as important are categorized under quality assurance, eligibility and access, academic and social supports, systems for accountability, aligned data systems and sustainable funding and finance.
After analyzing the different policy elements and standards among ECS, JFF and NACEP, four themes were identified that provide a coherent grouping of the elements: 1) program quality, 2) student access and support, 3) reporting and program evaluation, and 4) financial provisions. Given NACEP's focus on program quality, its standards are concentrated under that first theme, but they do also address student access and program evaluation. The ECS and JFF model policy elements include more guidelines around program evaluation and financial provisions.
In terms of program quality, some states require or encourage their postsecondary institutions to seek NACEP accreditation as a way to ensure concurrent enrollment courses are rigorous. Other states defer to local control and leave it to individual colleges to monitor concurrent enrollment course quality.
There are two model components regarding financial provisions that are recommended by JFF and ECS. The first concerns keeping costs away from students and families so that the program is open and affordable to all. The second model policy element focuses on keeping costs low for districts and colleges. There are a variety of funding approaches across states; JFF and ECS recommend that states cover the full costs of concurrent enrollment, or, at a minimum, allow both K-12 and higher education systems to collect per-pupil funding for student enrollments to offset costs. The latter method is referred to as "double dipping," although in this case the term is used positively as it ensures both systems have the means and incentive to participate (Floffman, 2005; Lerner & Brand, 2006; Ward & Vargas, 2012; Zinth, 2014b).
An additional theme identified concerns ensuring students have adequate support
throughout the processincluding before, during and after the concurrent enrollment course takes


14
placeand equitable access to the program. JFF advocates for more open access wherein students may demonstrate readiness for college-level coursework through portfolios, end-of-course grades, and teacher recommendations. In many cases, readiness is demonstrated through a course placement assessment. Lastly, a theme across all three organizations was the importance of tracking student data, reporting outcomes, and evaluating the effectiveness of the program in meeting its intended goals. With these model policy elements in mind, Colorado's concurrent enrollment legislation is explored in more detail in the next section.
Colorado's Concurrent Enrollment Programs Act Prior to the passage of Colorado's Concurrent Enrollment Programs Act, there were dual enrollment opportunities available to Colorado high school students, but there was no state-level coordination of the programs, which resulted in little accountability or attention to quality and low participation rates, particularly for low-income and minority students (CDE, 2010; CRS §22-35-102(d)). In 2007, Governor Ritter convened a P-20 Education Coordinating Council to develop policies that would foster a seamless education system in which all students receive a high-quality education from pre-school through graduate school and enter the workforce prepared to meet the demands of today's economy (Lopez, 2011). One of the forces driving the creation of the P-20 Council was the "Colorado Paradox," which refers to the fact that Colorado is one of the most highly educated states due to imported talent, but Colorado's own K-12 students are not persisting to and through college at high rates (Lopez, 2011; NCHEMS, 2013). Postsecondary access and success was a focal point of the council's work, and in 2009, at the recommendation of the council, legislative leaders introduced the bipartisan Concurrent Enrollment Programs Act (House Bill 09-1319 and Senate Bill 09-285). The legislation passed unanimously in both chambers of the legislaturea rare
feat.


15
Policy Goals
The concurrent enrollment program was specifically created to reach traditionally underserved populations. As the legislative declaration of the Concurrent Enrollment Programs Act states:
Historically, the beneficiaries of concurrent enrollment programs have often been high-achieving students. The expanded mission of concurrent enrollment programs is to serve a wider range of students, particularly those who represent communities with historically low college participation rates. (CRS §22-35-102(d))
The program is also seen as a way to fulfill state goals of halving the high school dropout rate and doubling the number of postsecondary credentials earned by Coloradans (Lopez, 2011; CRS §22-35-102). To reach those goals, the legislation was designed to broaden access to concurrent enrollment courses and to improve the quality of the programs. Legislation also specifically permits students to take concurrent career and technical education (CTE) courses, which fits with the program intent of accelerating students to a credential through multiple pathways.4 Key Policy Features
Colorado's legislation is seen as model for other states looking to expand concurrent enrollment. JFF closely evaluated every statewide policy against their six model policy elements, and they identified Colorado as one of five "exemplar" states, along with Florida, New Mexico, Texas and Utah (Jobs for the Future [JFF], 2012).
One key feature of the legislation is that it establishes a transparent funding process that shares costs between high schools and colleges, while keeping costs low for families. The funding mechanism permits both districts and colleges to collect state funding for students in concurrent enrollment to help defray costs (CRS §22-35-101 et al.). As mentioned in the previous section, this
4 The Concurrent Enrollment Programs Act also creates the 5th year ASCENT program for students retained by school districts to receive instruction beyond the senior year. The focus of this dissertation will be on the 9th-12th grade Concurrent Enrollment program; ASCENT students will be excluded from the analysis.


16
funding mechanism is a model policy element according to both JFF and ECS (Ward & Vargas, 2012; Zinth, 2014b). School districts use per pupil revenue (PPR) to cover tuition costs for concurrent enrollment students. Districts pay tuition to the postsecondary institutions directly on behalf of students. Previously in Colorado, families would pay for tuition costs and would possibly be reimbursed by the district later. That process, however, can be prohibitive to low and middle-income students and reduce access. Partnering postsecondary institutions are allowed to include concurrent enrollment students in its determination of enrollment numbers for funding purposes. Lastly, students apply for and authorize the institution to collect the Colorado Opportunity Fund stipend to pay that portion of the tuition (C.R.S. 22-35-105 (2)). While students and families do not pay any tuition costs, they may be responsible for books, transportation, technology or fees, depending on local financial arrangements. Further, if students do not complete the course and do not have the permission of their principal for a non-completion, they may be required to reimburse the school for the tuition costs (C.R.S. 22-35-105(4)). Students and parents fill out formal paperwork to apply for concurrent enrollment, and the terms for repayment, if any, should be specified in the application (CDE, 2016).
Districts are required by statute to notify families of concurrent enrollment opportunities and, if any schools within the district want to concurrently enroll students, the district must enter into a "cooperative agreement" with a postsecondary institution. As set forth in the law, cooperative agreements must, at a minimum, include the following elements:
The amount of academic credit to be granted for successfully completed course work by concurrently enrolled students;
A requirement that concurrent enrollment course work qualifies as academic credit towards a certificate or degree, or basic skills credit;
A requirement that the local education provider (i.e. school district, charter school or Board of Cooperative Services) pay tuition for courses completed by a student, according to the negotiated amount;
A requirement that the local education provider and the postsecondary institution establish an academic plan of study for concurrently enrollment students, and a plan for the district to provide ongoing counseling and career planning;


17
Confirmation by the district of the student's unique State Assigned Student Identifier (SASID) for funding and enrollment purposes;
Authorization for payment of the College Opportunity Fund on behalf of the student;
Consideration and identification of ways for concurrent enrollment students to remain eligible for interscholastic high school activities; and
Additional financial provisions. ((C.R.S. 22-35-104(6))
The cooperative agreements set forth the basic ground rules for the partnership between high schools, districts and postsecondary institutions. Often included in the agreements, in addition to the components listed above, are the specific fiscal and operational arrangements regarding course location and instructors. The concurrent enrollment classes must be offered by an eligible institution of higher education, but can be delivered on the high school campus, college campus, online, or in a hybrid format. If the courses are taught by high school teachers they must be credentialed as college adjunct faculty.
The concurrent enrollment program rules specify that all qualified students in the ninth grade or higher in a public school may take courses for both high school and college credit. To determine if a student is qualified, institutions of higher education use the same course prerequisites they use with all other postsecondary students seeking to enroll in the same class on their campus (CRS § 22-35-104 (4)(a)). High schools and colleges have to collaborate to ensure that students are properly assessed and meeting prerequisite requirements for course placement. Colleges are ultimately responsible for the course content and the quality of instruction, even if the course takes place on a high school campus taught by a high school instructor (who has been approved as an adjunct faculty member).
Contributions to the Field
Since Colorado is seen as having a model state concurrent enrollment policy (JFF, 2012; Lopez, 2011), this study uses Colorado as a case study and begins by exploring the factors that led some schools in the state to adopt concurrent enrollment more quickly and implement it more


18
widely than other schools. After understanding the key conditions at the school level, the author analyzes student-level behaviors by exploring what types of students are choosing to participate in the program and what the effects are of taking concurrent enrollment courses on college access and success. The study considers, in particular, if concurrent enrollment improves postsecondary outcomes among traditionally-underserved students.
The findings of this study will be valuable to practitioners, policymakers, and other researchers because state policy continues to be heavily relied upon as a lever for changing educational outcomes, yet, there is not a clear understanding of whether, or how, state policy affects behaviors at the institutional and student levels. Policy diffusion behavior is especially informative at the sub-state level in Colorado because the state has a strong local control culture, and policies and behaviors can vary by locality. While Colorado provides an appropriate case study for the questions at hand, other states are also experimenting with education reforms under similar conditions. Therefore, the findings of this study can be generalized to other states and other related education policy areas.
Policy Diffusion and Innovation Research
This study also seeks to contribute to policy diffusion and innovation theory. The theory is most often applied to state governments (Berry & Berry, 2007). There have been studies conducted of local governments, but the body of research is much smaller and focuses on municipalities (Shipan & Volden, 2008). Thus, this research will contribute to the continual exploration of the theory by applying it to a unique unit of analysishigh schools. There is no apparent study on the diffusion of concurrent enrollment across high schools.
Additionally, the vast majority of the studies conducted using policy innovation and diffusion have focused on the adoption of a policy without considering what occurs after adoption in the
implementation stage. Scholars have identified this gap in the literature and have called for studies


19
to apply policy diffusion analysis beyond a simple dichotomous measure of adoption to measures of policy implementation (Shipan & Volden, 2012). This study will seek to fill this gap in the literature by conducting an analysis of the factors that influence policy implementation, as measured by the share of students taking concurrent enrollment courses within a high school.
Lastly, more recent diffusion research has focused on the importance of the characteristics of the policy itself in terms of salience, complexity and compatibility to the diffusion process (Boushey, 2010, Makse & Volden, 2011, Nicholson-Crotty, 2009). These policy attributes are theorized to affect how quickly policies are adopted among states and municipalities. Makse and Volden (2011), for example, analyzed the diffusion of criminal justice laws across states and found that compatible policiesthose that fit seamlessly into current practicesare quicker to diffuse than complex policies that require a great shift in the status quo. Given that this is a relatively newer stream of diffusion research, this study will provide a modest contribution to the literature on policy characteristics by conducting a diffusion analysis of Colorado's concurrent enrollment policy, which could be considered a compatible policy according to the typology of policy characteristics (Makse & Volden, 2011, Shipan & Volden, 2012).
Education Research
This study also seeks to contribute to the field's understanding of whetherand to what extentstudents participating in concurrent enrollment see improvement in educational outcomes in terms of college access and college readiness (whether students are prepared to academically succeed once in college). Education researchers often struggle with controlling for selection bias due to limitations on available data and analytical methods, and this is true for prior research on concurrent enrollment (Allen & Dadgar, 2012; An, 2012; Le, Casillas, Robbins, & Langley, 2005). This study will contribute to the education research field by attempting to better control for selection
bias to more precisely isolate the effects of this particular intervention.


20
Concurrent enrollment programs have been around for decades, but, until recently, studies of program effectiveness were limited in number and rigor. Karp et al. (2007) found dual enrollment students in New York City and Florida in career and technical education programs were more likely to enroll in college, persist to the second year, and have higher GPAs and higher credit accumulation. Martin (2013) found that dual enrollment students at one North Carolina community college had higher college grades than non-dual enrollment peers. Allen and Dadger (2012) evaluated the dual enrollment program at the City University of New York and found that dual enrollees earned higher GPAs and more credits once in college. Those studies, while finding positive outcomes, were narrow in focusinvestigating particular colleges or programsand often employed methods that did not adequately control for selection bias (Giani, Alexander, & Reyes, 2014; Taylor, 2015, USDOE, 2017). There is one study that meets rigorous quasi-experimental design standards and is broad in scopeAn (2013) used a national dataset and found that dual enrollment programs increase degree attainment rates for first-generation students (USDOE, 2017).
Very recently researchers have published quasi-experimental evaluations of statewide concurrent enrollment programs. These studies were possible due to the recent expansion of statewide longitudinal data systems. Cowan and Goldhaber (2014) used Washington's data system to analyze the statewide "Running Start" dual enrollment program and found positive effects on college enrollment, particularly for students who are lower-performers academically. Taylor (2015) followed Illinois' graduating class of 2003 to track college entrance and completion rates for dual enrollment students and found positive effects overall, though the effect sizes were smaller among low-income students and students of color. Haskell (2016) analyzed 2008 and 2009 high school graduates in Utah and found reduced time to college degree completion and potential financial savings to families and the state. Giani, Alexander and Reyes (2014) use the statewide longitudinal data system in Texas to track 2004 high school graduates into college. Their study found greater


21
enrollment, persistence and completion among dual credit students as compared to non-dual credit students. Importantly, dual credit students enjoyed greater postsecondary benefits as compared to students taking other forms of advanced coursework such as Advanced Placement and International Baccalaureate courses (Giani et al., 2014).
All but one of the above-mentioned studies (Allen & Dadgar, 2012; An, 2013; Cowan & Goldhaber, 2014; Giani et al., 2014, Karp et al., 2007, Martin, 2013, Taylor, 2015) were recently evaluated by the What Works Clearinghouse (WWC), and only two met the WWC design standards with reservations: Giani et al.'s (2014) evaluation of Texas's dual enrollment program and An's (2013) nationally representative study (USDOE, 2017).5
While the Illinois, Utah, Texas and Washington studies indicate that concurrent enrollment students participating in state-wide programs have more positive postsecondary outcomes than their non-participating peers, each study is set in its own state policy context. The Texas program design is substantially different than Colorado's. The Texas and Illinois studies both use graduating cohorts from the earlier part of the 2000s, which allows them to follow students further into higher education, but also negates the ability to identify more recent trends. With the exception of Giani et al. (2014), none of the studies uses an intensity measure of concurrent enrollment participation (e.g. number of credit hours taken). And, Giani et al.'s (2014) study does not include data on Texas high school graduates who attend college out-of-state in their college matriculation model, which could bias their results. Further, there is value in determining whether concurrent enrollment outcomes are consistent across states. The concluding chapter of this dissertation considers how Colorado's results compare to the findings in these other state studies.
In summary, additional research beyond the emergent state studies is needed for the field to gain confidence in concurrent enrollment as an effective college readiness intervention,
5 Haskell (2016) was not reviewed by the WWC.


22
particularly give the uniqueness of each state's policy and the uncertainty regarding consistency of findings across states. While prior research has found dual enrollment programs result in benefits for the average student, it is unclear to whom those benefits accrue and to what extent. Few studies have examined the effects on traditionally underserved populations, and of those that have, the results are inconsistent. Taylor (2015) found minority and low-income students saw smaller gains in postsecondary outcomes when compared to their peers in Illinois, while An (2013) found higher effect sizes for students from disadvantaged backgrounds. It is evident that with 47 states having statutes governing concurrent enrollment programs much can still be learned about the effectiveness of these policies.
Summary & Research Questions
Colorado's concurrent enrollment policy was enacted in 2009, and within five years, 91 percent of the state's high schools offered concurrent enrollment to some degree. Given the rapid diffusion of the program, this study will seek to identify the local variables and conditions that affect the decision to adopt concurrent enrollment programs in high schools in an effort to uncover any best practices that could be applied to other states trying to scale up similar programs. The first research question is stated as follows.
RQ 1: What factors influence whether or not high schools adopt concurrent enrollment
programs?
Additionally, because Colorado's state policy is voluntary, there is ample variance and room for innovation at the local level in regards to whether and how the program is implemented. High schools may adopt concurrent enrollment to add another option to an already existing portfolio of college readiness or credit accrual programs (e.g. Advanced Placement courses, International Baccalaureate program, honors courses, etc.). Alternatively, a high school may launch concurrent
enrollment as a way to provide access to college-level courses to all or nearly all upper-classmen. A


23
rural school may, for example, enroll all seniors in a concurrent enrollment college-level math course. Offering a concurrent enrollment math course to a classroom of seniors at a large high school would only constitute a small percentage of the total senior class, whereas at a small, rural school it may comprise the majority of the school's seniors. This potential for variation in the degree to which students are participating in concurrent enrollment within a high school leads to a subquestion:
RQ la: What factors influence the extent to which concurrent enrollment programs are utilized by students within high schools?
To date, there is no apparent empirical examination into the school- or district-level characteristics that lead to faster or deeper program adoption at certain high schools as compared to others. This study also seeks to understand if students participating in Colorado's concurrent enrollment program see improvement in educational outcomes in terms of both their participation in college and their success once in college. The author considers if the program has positive effects for Colorado's traditionally-underserved students in particular. Accordingly, the second and third research questions are as follows:
RQ2: How does participation in concurrent enrollment affect the college-going rates of Colorado's high school students?
RQ3: How does high school participation in concurrent enrollment affect the college performance and persistence of students?
These two research questions combined with the first question are collectively important because in order for state-facilitated, voluntary policies to significantly improve educational outcomes, the policy needs to be both widely diffused in schools and impactful on individual students. Answers to these questions will be beneficial to policymakers, practitioners and
researchers.


24
CHAPTER II
LITERATURE REVIEW AND HYPOTHESES
Several different literature streams inform the research questions set forth in the previous chapter. The first research question concerning differences in school-level adoption of concurrent enrollment is informed by policy innovation and diffusion theory. The second and third research questions, which are concerned with the individual educational outcomes of participating in concurrent enrollment, rely on different strands of education theory. Hypotheses are drawn from the literature and presented throughout the chapter.
Policy Diffusion & Innovation Theory
Policy diffusion and innovation theory submits that political, economic and social factors, along with competitive and emulative pressures, influence whether or not policy change is adopted (Mokher & McLendon, 2009, 251). The theory's roots reside in Walker's (1969) seminal article on the diffusion of innovation among American states in which he sought to understand why some states adopt innovative policies quicker than others. Walker (1969) defined innovation "simply as a program or policy which is new to the states adopting it" (881). His work was groundbreaking because he focused not only on the importance of a state's internal factors (drawn from organizational innovation literature) but also on the role of competitive and emulative pressures among states. Walker (1969) observed that national professional communities served as learning opportunities where ideas were spread among state policy makers and administrators. He also noted that some states were seen as leaders, and sought to determine if other states were more likely to emulate policies of the leader states (Walker, 1969).
In the years following, numerous scholars sought to test and refine Walker's (1969) propositions (Berry, 1994b; Shipan & Volden, 2012). Some scholars focused solely on the internal
factors, or determinants, that lead states to be early innovators, while other scholars focused their


25
research on regional diffusion and national interaction patterns (Berry, 1994b; Berry & Berry, 2007; Mokher & McLendon, 2009). In the 1990s, research by Berry & Berry (1990, 1992) significantly advanced the research genre by offering a methodologyevent history analysisthat provided a way to empirically test the effects of both internal determinants and external diffusion factors in one model. Since then, policy diffusion and innovation theory has been applied using event history analysis to a wide range of substantive topics, such as health care (Stream, 1999; Volden 2006), hate crime laws (Soule & Earl, 2001), electricity deregulation (Ka & Teske, 2002), and education (Mintrom, 1997; Wong & Shen, 2002). The majority of empirical applications of policy diffusion focus on state governments, but the theory has also been applied to local municipalities (e.g. Bingham, 1977; Hoyman & Weinberg, 2006; Lubell et al., 2002). These studies, along with others, considered different mechanisms that drive the diffusion process, including policy entrepreneurs (Balia 2001; Mintrom 1997), learning that occurs from effective policies (Gilardi, Fuglister, & Luyet, 2009; Volden 2006), competition (Baybeck et al., 2011, Berry & Berry, 1990) and coercive forces (Karch, 2006; Welch & Thompson, 1980). More recent diffusion research also has focused on the importance of the characteristics of the policy itself in terms of salience, complexity and compatibility to the diffusion process (Boushey, 2010, Makse & Volden, 2011, Nicholson-Crotty, 2009). Makse and Volden (2011), for example, found that compatible policiesthose that fit seamlessly into current practicesare quicker to diffuse than complex policies that require a greater shift in the status quo.
While the policy diffusion and innovation literature is wide and varied, the theory is generally focused on the following overarching factors: "a polity's motivation to innovate, the resources that it has to innovate, the obstacles that stand in the way of this innovation, other cocurrent policies that a polity is pursuing and the influence of the external environment" (Hoyman & Weinberg, 2006, 98-99). These factors often comprise the central elements of empirical models


26
used to understand and predict why certain public entities are quicker to adopt innovative policies than others. Given the lack of research on what affects policy implementation beyond the initial state of policy adoption, the same theoretical grounding will be used to explore both adoption and implementation effects in this study (Shipan & Volden, 2012).
Motivation to Innovate
The policy innovation literature theorizes that if problem severity is high, governmental entities are more likely to be motivated to adopt new policies (Berry & Berry, 1990, 2007; Hoyman & Weinberg, 2006; Mohr, 1969). Berry and Berry (2007), in their review of the literature on policy innovation, state that "problem severity can influence the motivation of state officials to adopt a policy directly by clarifying the need for the policy, or indirectly by stimulating demand for the policy by societal groups" (Berry & Berry, 2007, 235). Empirical evidence for this proposition has been found in several studies covering topics such as welfare, school choice and health care reform (Allard, 2004; Mintrom & Vergari, 1998; Stream, 1999).
Concurrent enrollment programs can be viewed as a tool for increasing student achievement levels and improving college readiness (Hoffman, 2005). Thus, districts struggling with low academic achievement levels may have more incentive to innovate and improve outcomes and may be more likely to turn to concurrent enrollment as a potential solution. The severity of low academic achievement in districts and schools should motivate them to adopt concurrent enrollment, whether that pressure to improve comes from internal leadership, the state (via performance ratings), or concerned parents. Likewise, if a school has an urgency to improve the achievement of its students, it could be hypothesized that the school will encourage higher levels of participation in concurrent enrollment courses. That is, if a high school adopts the program as an improvement strategy, it would follow that the school would actively encourage and recruit students to participate in it. A hypothesis on the motivation to innovate is proposed, as follows.


27
Hypothesis 1.1: High schools that have lower academic achievement levels are more likely to adopt concurrent enrollment and have higher student participation rates.
Resources to Innovate
Policy diffusion and innovation theory also points to the importance of resources, which are needed both for the innovation itself and to overcome any obstacles to innovate (Berry, 1994b). There are two types of resources that the literature on policy diffusion and innovation identifies that are particularly relevant to this study: fiscal capacity and organization size. Generally, past studies have found that a public entity's probability of adopting innovative policies is positively related to the resources at its disposal (Berry & Berry, 2007).
In regards to fiscal capacity, the literature on policy innovation draws from the broader literature on organizational innovation, which has consistently found that financially-secure organizations are more likely to innovate than organizations in fiscal trouble or with fewer slack resources (Berry, 1994a; Bingham, 1977; Cyert and March, 1963; Rogers, 1983). Policy innovation theory points to the fact that many new policies and programs require extensive funds to be implemented, and thus agencies with abundant resources may be more inclined to adopt such programs (Berry & Berry, 2007) and may have a greater capacity to widely implement them. Therefore, drawing from the literature and theory on financial resources and innovation, the following hypothesis is proposed. Hypothesis 1.2: High schools with greater fiscal capacity are more likely to adopt concurrent enrollment and have higher student participation rates.
Theory and research on organizational innovation has long considered the size of an organization as another key explanatory variable that is positively associated with the likelihood of innovation (Baldridge and Burnham, 1975; Berry, 1994a; Cyert & March, 1963; Mohr, 1969; Rogers, 2003). Size is considered an important element because it facilitates the presence of other factors
that may affect innovation such as the availability of slack, or surplus, resources and specialized staff


28
(Rogers, 2003). Larger organizations tend to have more resources, and they are more likely to have the capacity and structure to hire specialized staff. Having administrators whose job it is to stay attuned to the latest research and innovations in their specific program area makes it more likely that schools will be aware of new programs and have the capacity to implement them (Baldridge, 1975; Berry, 1994a). As school size increases, however, it may be more likely that the proportion of eligible students who participate in innovative programs decreases. This is particularly true if larger schools have a broad array of programs from which students may select. Larger organizations may have greater numbers of participants compared to smaller organizations, but in terms of participation rates it seems likely that organizational size may be inversely related to overall participation rates. Based on the organizational innovation literature and general reasoning, the following resource hypothesis is proposed. Hypothesis 1.3: The larger a school, the more likely it is to adopt concurrent enrollment, but the share of students participating may be lower.
External Factors
Policy diffusion scholars have devoted much attention to how competitive and emulative pressures among states affect policy innovation (Berry, 1994b; Berry & Berry, 2007; Walker, 1969). Policy innovation and diffusion theory points primarily to two avenues through which external pressure for policy change occurs: through regional diffusionthat is, through competition with or emulation of neighboring governmentsor through national interaction, which is the idea that policy ideas spread through networks of policymakers (Berry & Berry, 2007; Walker, 1969). Empirical tests of diffusion models have found varying results. A national study of the diffusion of concurrent enrollment policies among states, for example, found that regional diffusion pressures were not a statistically significant factor (Mokher & McLendon, 2009). On the other hand, a recent study of the diffusion of charter school legislation among states found the regional diffusion variable to have a statistically significant and substantive influence on policy adoption (Lee, 2014). Some scholars


29
argue that the geographic focus of policy adoption is an outdated concept given that policymakers and leaders can learn about innovative programs not just from their geographic neighbors, but from others across the statesor increasingly across the world (Shipan & Volden, 2012; Volden, Ting & Carpenter, 2008).
Nonetheless, as an open enrollment state, schools in Colorado have an incentive to compete with one another for students because those students come with revenue attached. If a high school sees that a neighboring school is offering advanced or enhanced programming, it seems likely that competitive and emulative pressures could influence policy innovation decisions and may also lead to more robust implementation reflected by higher participation rates Thus, although research and theory is unclear on the influence of regional pressures, this study hypothesizes that geographic proximity will be influential. Hypothesis 1.4: High schools nearby other schools that have already adopted concurrent enrollment are more likely to offer the program and have higher participation rates.
Another external factor that could be relevant to this research is a high school's proximity to a community college. There is precedence in the literature for focusing on this factor; the influence and presence of community colleges was used in the previously-mentioned study of national dual enrollment policy diffusion (Mokher & McLendon, 2009). The nearness of a community college to a high school makes the implementation of the concurrent enrollment policy easier in terms of accessibility to courses delivered at a college campus, credentialing of high school instructors or the provision of community college instructors (in cases where college faculty teach courses at a high school's campus). While some concurrent enrollment is provided by four-year institutions, the large majority of course enrollments are through community colleges. In Colorado during the 2015-16 school year, for example, 88 percent of concurrent enrollment students took courses through community colleges (Colorado Department of Higher Education [CDHE], 2017b). Moreover,


30
community colleges are proponents of concurrent enrollment because it is an immediate revenue generator, as well as a recruitment strategy (Crooks 1998; Morest & Karp 2006). Community colleges may exert pressure on schools to offer concurrent courses, in which case, if a community college is geographically close to a particular high school, the policy may diffuse more readily and, perhaps, more deeply, which would be evidenced by a greater share of students within a high school taking concurrent enrollment courses. Hypothesis 1.5: High schools with greater proximity to a community college will be more likely to adopt concurrent enrollment and have higher participation rates.
Education Theory
Because education is such a broad field, it is necessary to first situate the research through a specific theoretical lens. Some researchers take an institutional rational choice lens, for example, and place emphasis on how institutional arrangements constrain and aggregate individual choices resulting in organizations of different types and quality (Chubb & Moe, 1990). Other researchers focus on how manipulating inputs within existing organizational and institutional arrangements can alter performance outputs. These scholars, which devote their time to identifying problems and evaluating reforms that occur in and around schools, could loosely be grouped under a theoretical framework of school improvement.
The theoretical reasoning behind concurrent enrollment is grounded in the viewpoint that there are factors within the control of schools that can be manipulated to improve education outcomes. Empirical research has linked a variety of factors to the likelihood that students will attend and be successful in college. Some factors cannot be altered by schoolssuch as socioeconomic statusbut other factors, including academic preparation and metacognitive skill development, can be manipulated. Thus, the school improvement framework is an applicable lens
for this inquiry. This section of the literature review provides an overview of education theory and


31
research related to improving student achievement, generally, and college access and success, specifically.
School Improvement Framework
There are a significant number of education researchers who rely on a theoretical lens that focuses fundamentally on manipulating inputs within existing school arrangements to alter performance outputs (Hanushek, 2003; Purkey & Smith, 1983). Scholars coming from the field of economics refer to these types of studies as education production functions, which are used to relate "observed student outcomes to characteristics of the students, their families, and other students in the school, as well as characteristics of schools" (Hanushek, 1979, 354). Education and economics scholars could be loosely grouped together under a theoretical framework of school improvement, which asserts that the problems that plague student achievement are "found in and around the schools, and the schools can be 'made' better by relying on existing institutions to impose the proper reforms" (Chubb & Moe, 1990, 3). This school of thought emerged in the wake of the 1966 Coleman Reportformally known as Equality of Educational Opportunity (Chubb & Moe, 1990; Hanushek, 1979, 2003; Purkey & Smith, 1983). The Coleman Report (1966), which was the product of a substantial study covering over 600,000 students in 4,000 schools, found that once the family background of individual students and the overall racial composition of schools were taken into account, school characteristics had little effect on achievement levels (Coleman et al.). The school characteristics included in the Coleman study were numerous and included such factors as school funding, classroom size, teacher and principal salaries, education levels of teachers, number of free textbooks, and extracurricular offerings (Coleman et al., 1966).
In a backlash against the report's findings, researchers spent the following decades attempting to prove that school-based elements do matter (Hanushek, 2003; Purkey & Smith, 1983). One rationale for such research is that factors related to family background are not easily changed


32
at least in the short termand, therefore, the many factors that can be manipulated at the classroom, school, or district-level must continually be examined in the effort to improve schools (Purkey & Smith, 1983). Some researchers chose to focus on organizational and cultural attributes not included in the Coleman Report; this group of scholars contributed to what is known as the "effective schools research" (Chubb and Moe, 1990; Edmonds & Frederiksen, 1979; Hersh et al., 1981; Purkey & Smith, 1983). While encompassing a diverse set of studies and findings, overall, the effective schools research conducted in the 1970s and early 1980s emphasized the following elements as key to improving student achievement: "high staff expectations and morale, a considerable degree of control by the staff over instructional and training decisions in the school, clear leadership from the principal or other instructional figure, clear goals for the school, and a sense of order in the school" (Purkey & Smith, 1983, 438). Empirical evidence that these cultural and organizational elements could affect student achievement (as measured typically through assessment scores) was heralded as proof that schools matter (Edmonds, 1979; Hersh et al. 1981, Purkey & Smith, 1983).
Other researchers took a different path and have focused on re-examining the variables included in the Coleman Report and have found some nuanced, positive effects on student achievement. Several studies, including the well-known Tennessee STAR experiment, for example, have found statistically significant, positive effects of small classroom size (below 20 students per teacher) on student learning in Kindergarten through third grade (Centra & Potter, 1980; Mosteller, 1995; Walberg, 1982). Others, however, have found little evidence that class size affects learning enough to warrant its financial costs (Hanushek, 1999; Funkhouser, 2009). The debates within the literature on classroom size are representative of the school improvement research as a whole in that there are some positive findings but much debate over meaning and replication. Analysis of
other inputs such as school funding, teacher education and teacher salaries constitutes a large, and


33
somewhat controversial, literature. The wide variety of variables, the broad range of statistical methods used, and the differing levels of empirical quality have resulted in findings that are often seen as contradictory (Hanushek, 1979). Hanushek (2003) calls for more rigorous research on school inputs, stating that "if educational policies are to be improved, much more serious attention must be given to developing solid evidence about what things work and what things do not" (94).
In the last 20 years, scholars have indeed continued to analyze a variety of school-based programs in search of "solid evidence" of positive effects on student learning. The focus of this more recent research has been on topics such as standards and curriculum, bilingual education, literacy programs, and improving the transition from high school into college (Hoffman, 2005; Darling-Hammond, 2010; Kirst & Venezia, 2004, Hoffman, Vargas & Santos, 2008a, 2008b). Initiatives focused on aligning secondary and postsecondary systems are often referred to as K-16 or P-20 initiatives. Research on P-20 systems was epitomized in the work of the Stanford Bridge Project in the late 1990s, which made the case that the underpreparation of high school graduates for higher education was a pervasive and critical problem (Venezia, Kirst & Antonio, 2003). The researchers found the coursework offered in high school and college to be disconnected, and they noted that underrepresented students were "especially likely to be hampered by insufficient access to college preparatory courses" (Venezia, Kirst & Antonio, 2003, 8). The expansion of dual enrollment programs was one of the key recommendations of the Stanford Bridge Project to improve the transition from high school to college. Dual enrollment is seen as an avenue for students to gain stronger academic preparation for college (Kirst & Venezia, 2004; Venezia, Kirst & Antonio, 2003). The next section of the literature review will explore in detail the theory behind how academic preparation and coursework offerings relate to a student's educational achievement.


34
Academic Preparation
Empirical research studies over the past half century have consistently identified that disparities in curriculum offerings, including course options, rigor of curriculum, and quality of courses, contribute to and exacerbate achievement gaps. This stream of literature focuses on the inequity of educational opportunity between groups of students, and the detrimental effect that has on academic success. Dreeben's (1987) study of inner-city elementary students in Chicago, for example, found that black and white students of similar aptitude performed equally well when exposed to the same instructionhigh or low quality (Dreeben, 1987). The problem is that high-quality instruction is not offered in every classroom, and low-income, minority students tend to disproportionately receive inadequate instruction (Darling-Hammond, 2010). This contributes to persistent achievement gaps as evidenced through several studies that have found that students who are exposed to rich, challenging curriculum eventually outperform their peers who are placed into less rigorous classes, even after controlling for socioeconomic background (Alexander & McDill, 1976; Gamaron, 1990; Gamaron & Hannigan, 2000; Oakes, 2005). Peterson (1989), for instance, conducted an experimental study that randomly placed at-risk 7th graders with similar backgrounds into varying levels of math classes. Students placed into the highest math class (containing a prealgebra curriculum) outperformed the other students on assessments given at the end of the school year (Peterson, 1989).
Researchers have found that the disproportionate allocation of high-quality instruction to students occurs primarily through two ways. The first is that schools with minority-majority populations (i.e. serving mostly Hispanic, black or Native American students) offer fewer academically-rigorous courses. Instead of having a selection of honors, Advanced Placement, lab
science and foreign language courses like high schools in wealthy districts do, high schools serving


35
large numbers of minority and low-income students often offer mostly remedial and vocational courses (Pelavin & Kane, 1990; Oakes, 2005; Darling-Hammond, 2010).
The second way that high-quality instruction is allocated away from low-income, minority students is through tracking. In schools with socioeconomically-diverse populations, white and upper-income students tend to be placed into college preparatory classes while minority and low-income students are tracked into lower-level courses (Darling-Hammond, 2010; Oakes, 2005). As Darling-Hammond (2010) phrases it, "curriculum tracks are generally color coded" (52). Oakes (2005) has conducted empirical studies over the past several decades that highlight the pervasiveness of tracking in America's schools. One of her studies, for example, found that after controlling for test scores, white and Asian students were far more likely to be placed into honors courses than their peers (Oakes, 1993). High-achieving Latino students who scored at the 90th percentile on standardized tests had just a 56.3% chance of being assigned to a college preparatory class, as compared to 97.3% of Asian students and 93.3% of white students scoring in the same percentile (Oakes, 1993).
The reasons for the underrepresentation of minority and low-income students in honors, AP and other challenging secondary courses are many. Tracking begins at an early age and, by the time students reach high school, tracked students often do not have the prerequisite skills or test scores to take advanced courses (Darling-Hammond, 2010; Oakes, 2005). More directly, counselors may advise students from low socioeconomic backgrounds away from challenging postsecondary pathways and towards low-status careers (Darling-Hammond, 2010). Further, middle- and upper-income parents tend to be more active in pushing for their children to be placed into advanced courses and programs and to be assigned to the best teachers. As Darling-Hammond (2010) explains, high-quality education is scarce and,
Scarce resources tend to get allocated to the students whose parents, advocates or
representative have the most political leverage. This typically results in the most highly


36
qualified teachers offering the most enriched curricula to the most advantaged students.
(60)
Indeed, other empirical work has found tracking patterns in place for teachers in which the best (most-experienced, most competent) teachers are assigned to the brightest students in upper-level classes, while inexperienced and ineffective teachers are assigned to lower-level classes and students (Finley, 1984; Talbert, 1990).
As this section of the literature review demonstrates, disparities in curriculum offerings and quality of instruction remain a significant problem in today's schools. Theoretically speaking, ensuring that all students have access to rigorous coursework in high school through programs such as concurrent enrollment could improve education outcomes. Concurrent enrollment programs expand the number of accelerated learning options available to schools, especially when programs have clear funding streams (Allen, 2010; Karp, Bailey, Hughes, & Fermin, 2005). Making concurrent enrollment courses more prolific in high schools and targeting them to students of all socioeconomic backgrounds could be a way to ensure traditionally-underserved students have access to enriched, advanced curriculum (Karp, Bailey, Hughes, & Fermin, 2005; Venezia, Kirst, & Antonio, 2003). For students who have been tracked into lower-level courses from an early age and enter high school far behind their peers academically, some concurrent enrollment programs, including Colorado's, offer remedial courses to high school seniors to help students become college ready (Allen, 2010; Rutschow & Schneider, 2011). This a newer strategy; dual enrollment courses have long been targeted to high-achieving students, but schools are now expanding the mission of such programs to serve less academically-prepared students (Allen, 2010; Karp et al., 2007; Rutschow & Schneider, 2011; Venezia, Kirst, & Antonio, 2003).
Metacognitive Learning Skills
When student do not take advanced, college-preparatory coursework, it is not content knowledge alone that students miss (e.g. algebra vs. pre-algebra), but also the development of


37
higher-order thinking and reasoning skills (Darling-Hammond, 2010). Researchers have found that classes in the lower tracks often focus on rote memorization, test taking, and behavioral problems (Eckstrom & Villegas, 1991; Good & Brophy, 1987; Oakes, 2005). In contrast, other research studies have shown that in college-preparatory tracks, teachers engage students in hands-on group activities and projects that encourage students to be creative, to problem solve and to think critically and strategically (Braddock & McPartland, 1993; Garcia, 1993; Wenglinsky, 2002). It is the latter skills that students need to acquire to ultimately be successful in higher education (Conley, 2007, 2010).
In fact, research on college readiness has long made the point that content knowledge alone does not predict success in collegethere is a host of other knowledge, skills, and behaviors students must acquire to be successful in postsecondary education (Attinasi, 1989; Byrd and MacDonald, 2005; Conley, 2005, 2007, 2010; Dickie & Farrell, 1991; Shields, 2002). These other attributes have often been referred to as "nonacademic" or "noncognitive" skills, which is arguably a misnomer because the skills and behaviors are directly related to cognition and thinking processes. Conley (2013) advocates for the terminology "metacognitive learning skills" to be used when referring to "the full range of behaviors, attitudes, and beliefs students demonstrate while engaging in the learning process" (Conley, 2013, 21). Metacognition is comprised of "personality and motivational factors, experiential and contextual intelligence, social skills and interests, and adjustment and student perceptions" (Educational Policy Improvement Center, 2013, para. 4). Students with well-developed metacognitive learning skills will be able to manage their time effectively, think critically, navigate college resources, maintain study routines, have self-awareness of their strengths and weaknesses, analyze and interpret information, and have the confidence to
overcome challenges (Conley, 2010, 2013).


38
Concurrent enrollment is a venue for students to acquire metacognitive learning skills. As Karp (2012) explains,
Dual enrollment can be seen as a social intervention in which potential college students learn about the norms, interpersonal interactions, and behaviors expected for college success. By "trying on" the role of a college student, dual enrollees benefit from early exposure and practice, coming to feel comfortable in a college environment and ultimately becoming successful once they matriculate. (22-23)
Karp (2012) tested facets of sociological theory in a qualitative study that measured the degree to
which students in concurrent enrollment courses gained knowledge of college behaviors, norms and
processes. The students in her study took dual enrollment courses at their high school, taught by a
teacher who was credentialed as a college adjunct. Despite not being physically located on a college
campus, Karp found that dual enrollment courses still gave students the opportunity to assume the
"role" of a college student, which helped them gain a deeper knowledge of what would be expected
of them in a college courses. Karp (2012) concludes that "dual enrollees get ready for college
success by learningbefore they actually matriculateall aspects of the college role" (23). This
finding supports other research demonstrating that students who take concurrent enrollment
courses have higher levels of self-efficacy, or confidence, in their academic abilities (Margolis &
McCabe, 2004).
College Affordability
Even if students have acquired the right knowledge, skills and abilities to be successful in college, college affordability remains a significant barrier to matriculation. Numerous studies have found that low- and middle-income students are more sensitive to tuition and aid changes than wealthier students, meaning that when tuition increases or grant aid decreases, there is a bigger decline in enrollment for low-income students than for upper-income students (see e.g. Heller,
1997; Leslie & Brinkman, 1987; Terenzini, Cabrera, & Bernal, 2001; St. John, 1990). In other research
on college affordability, one study analyzed enrollment rates by income level and standardized tests


39
scores and found that high-income students who scored among the worst on the achievement test were as likely to go to college as low-income students who performed among the best (Kahlenberg, 2004). Put another way, "the least bright rich kids have as much chance of going to college as the smartest poor kids." (Kahlenberg, 2004, 24). The study also found that 22 percent of low-income students with the highest test scores did not go to college, compared to 11 percent of middle-income students and only 3 percent of high-income students with the same scores (Kahlenberg, 2004).
Concurrent enrollment can be an opportunity to make college more affordable by allowing students to earn college credit for free, or at a reduced cost, while still in high school, thus reducing the amount of tuition students will have to pay when they matriculate to college (Hoffman, Vargas, & Santos, 2008a, 2008b; Jobs for the Future, 2006). In some programs, students are encouraged to accumulate enough credits to earn a certificate or associate degree at the same time as they earn their high school diploma.
From this review of the literature, it is theorized here that concurrent enrollment will improve college access and success for its participants for the following central reasons:
1) Concurrent enrollment provides for rigorous academic preparation and enhanced content knowledge;
2) Concurrent enrollment courses are a venue for students to acquire metacognitive learning skills and exposure to higher education; and
3) Courses provide the opportunity for students to earn free, or low cost, college credit, thus reducing the total amount of a college credential.
Following this theoretical framework, a hypothesis regarding college access is stated as follows. Hypothesis 2: High school students who participate in concurrent enrollment programs will
have a greater probability of enrolling in higher education. Similarly, the author expects positive


40
outcomes in terms of college success based on the theoretical framework, leading to a third central hypothesis. Hypothesis 3: First-year college students who had participated in concurrent enrollment programs in high school will have greater academic success and a higher probability of persisting than college students who did not participate.
Summary
Based on a comprehensive review of the literature, several hypotheses have been identified to guide the quantitative analysis of Colorado's concurrent enrollment program. Table 2 provides a summary of the research questions and the associated hypotheses. The following chapter will
discuss the data and methods used to test the hypotheses.
Table 2. Summary of Research Questions and Hypotheses
Research Questions Hypotheses
RQ1: What factors influence whether high schools adopt concurrent enrollment programs?
RQla: What factors influence the extent to which concurrent enrollment programs are utilized by students within high schools?
Hl.l: High schools that have lower academic achievement levels are more likely to adopt concurrent enrollment and have higher student participation rates.
HI.2: High schools that have greater fiscal capacity are more likely to adopt concurrent enrollment and have higher student participation rates.
HI.3: The larger a school, the more likely it is to adopt concurrent enrollment, but the share of students participating may be lower.
H1.4: High schools nearby other schools that have already adopted concurrent enrollment are more likely to offer the program and have higher student participation rates.
HI.5: High schools with greater proximity to a community college will be more likely to adopt concurrent enrollment and have higher student participation rates.
RQ2: How does participation in concurrent enrollment affect the college-going rates of Colorado's high school students?
H2: High school students who participate in concurrent enrollment programs will have a greater probability of enrolling in higher education.
RQ3: How does high school participation in concurrent enrollment affect the college performance and persistence of students?
H3: First-year college students who had participated in concurrent enrollment programs in high school will have greater academic success and a higher probability of persisting than college students who did not participate.


41
CHAPTER III DATA & METHODS
This chapter describes the data and methods used to explore the hypotheses. The study begins with an exploration of factors that influence the adoption of concurrent enrollment programs among and within Colorado high schools using school-level, administrative data in event history analysis and multivariate regression. The author then undertakes propensity score matching and fixed effects regression using student-level data to analyze the effects of participating in concurrent enrollment on college matriculation and success.
Data Sources and Collection
Nearly all the data used in this study were collected through Colorado's two education agencies: the Colorado Department of Higher Education (CDHE) and the Colorado Department of Education (CDE). The author constructed panel data sets by compiling publicly-available data from the agency websites and by procuring a de-identified and secure cross-section of student-level data from CDHE. The datasets are timely, comprehensive, andmost importantlongitudinal between K12 and higher education due to data-sharing agreements in place between the two agencies.
Statewide longitudinal data systems are still a relatively new phenomenon having rapidly expanded in states over the past decade. The wealth of information included in these state data systems has the potential to help transform the public administration of schools and colleges into a truly evidence-based sector. There are obstacles along the way to reaching that goal, however, including the recent backlash from parents and community members around the perceived overreach of government and businesses in collecting data on students. In fact, between 2013 and 2016, 36 states enacted 74 student data privacy laws, some of which establish important procedures for protecting student data, but some of which also constrain the ability of state agencies to share
student data with researchers (Data Quality Campaign, 2017). Thus, it is within this landscape that


42
this study occurred, and the author acknowledges that access to such rich and powerful data is not to be taken for granted in education research. Moreover, there is an important contribution that can be made to the debate around the value of making student data available to researchers if studies such as this prove worthwhile to policymakers and practitioners.
High School Panel
The first panel data set was constructed by collecting high school data from CDE, CDHE and the U.S. Census Bureau for the following academic years: 2009-10, 2010-11, 2011-12, 2012-13, 2013-14, and 2014-15. The full data set includes 388 high schools that served, at a minimum, grades 10, 11 and 12 during the entirety of the study period. High schools that opened recently and were not in existence for all years of the study, closed during the study period, or only served partial grades (e.g. 9 and 10) during the study period were excluded from the dataset.6 The resulting panel is strongly balanced, meaning all schools have data for all years of the study.
Aggregate data on multiple indicators for the 388 high schools were obtained from CDE's website. The indicators collected include school performance ratings, school type (i.e. charter, alternative education campus, or traditional school), student count, district setting, prior dual enrollment program participation rates, and free and reduced price lunch information. The data that were procured from the CDHE provide details about which high schools and higher education institutions offer concurrent enrollment programs and how many high school students were in enrolled in the program during a given academic year. College matriculation rates by high school were also obtained from the CDHE. Lastly, data were collected from the U.S. Census Bureau's American Community Survey for median household income.
6 Given the geographical focus of the policy diffusion research question and hypotheses, online high schools (n=19) were also excluded from the data set.


43
The dataset provides a sufficient scope of variables and a long-enough time span to analyze the first research question concerning the diffusion of concurrent enrollment among and within high schools. The Colorado legislature passed the Concurrent Enrollment Program Act in spring 2009, and the program was fully operational at the start of the 2010-11 school year. The school panel dataset used in this study spans from fall 2009 through spring 2015, allowing for an analysis of the first five years of program implementation.
Student Panel
The second panel data set was created from data collected through CDHE's Student Unit Record Data System (SURDS), which houses comprehensive postsecondary data on students who are enrolled at public colleges and universities in the state, as well as those enrolled at three private institutions: the University of Denver, Regis University, and Colorado Christian University. The CDHE supplements SURDS with data from the National Student Clearinghouse (NSC) to provide information on out-of-state enrollment and enrollment at private institutions. The NSC has a coverage rate of 96 percent of all students enrolled in a U.S. public or private college (NSC, 2013); thus, when this study considers college enrollment patterns, the dataset captures nearly all Colorado high school graduates who attend college, whether in-state or out-of-state, at a public or a private institution. Further, CDHE has established a partnership with CDE that permits the linkage of the postsecondary data with K-12 data using the State Assigned Student Identifier (SASID). The SASID-linked databases provided the means to create a student-level panel dataset that follows cohorts of high school graduates as they move from the K-12 system into higher education. The high school graduating cohorts of 2011, 2012 and 2013 are included in the student-level analysis.
The variables included in the second panel data set provide details, by semester, about what postsecondary institution students are enrolled in, whether they require remedial education, and
how they perform in terms of grade point average, credit accumulation, and persistence. The data


44
that were procured from CDHE around concurrent enrollment include how many credit hours students take and in which high schools and higher education institutions the students are concurrently enrolled.
Research Design
The research design begins with an event history analysis of how concurrent enrollment programs expanded among Colorado high schools. Using the high school as the unit of analysis, the author also uses regression analysis to see if any of the same factors included in the event history analysis affect the magnitude of program participation rates. Next, the author conducts multivariate analyses to evaluate the effects of participating in concurrent enrollment on education achievement using student-level data. Participation in concurrent enrollment is explored both as a dichotomous measure (yes/no) and as an intensity level (i.e. number of credits). The different components of this research design rely on the same dataset but have different guiding questions and units of analysis. A description of the variables, measure, and methods used in the policy diffusion portion of the study are presented first, followed by an explanation of the variables, measures and methods employed for the student-level policy evaluation. The chapter concludes with a summary of the research design.
Policy Diffusion Variables and Measures
The hypotheses relating to the policy diffusion analysis contain several key concepts related to policy adoption, motivation to innovate, resources and obstacles, and external factors. Table 3 summarizes the indicators that are used to operationalize the explanatory variables in the five diffusion hypotheses, and the following sections provide additional details. Table 4 provides a summary of variable descriptions and data sources.
Concurrent enrollment policy adoption. The key dependent variable for all of the hypotheses in the policy innovation and diffusion analysis is adoption of concurrent enrollment


45
programs. First, this study employs a dichotomous measure of adoption, with a value of 1" indicating that the high school has adopted the concurrent enrollment policy and there is at least one student taking a concurrent enrollment course. The study then includes a second measure of adoption to assess how the covariates affect the magnitude of student participation within a high school. That measure is a proportional value calculated by dividing the number of students participating in concurrent enrollment in a given academic year by the total number of students in grades 9 through 12 in that same year.7
Table 3. Concept Measurement Summary: Policy Diffusion
Hypotheses Constructs Indicators
1.1: High schools that have lower academic Academic College enrollment rates
achievement levels are more likely to adopt achievement School performance rating
concurrent enrollment and have higher student participation rates. levels (index of ACT scores, graduation rates, dropout rates and achievement on state standardized tests)
1.2: High schools that have greater fiscal capacity Fiscal capacity Median household income
are more likely to adopt concurrent enrollment and have higher student participation rates. Free and reduced price lunch eligibility
1.3: The larger a school, the more likely it is to adopt concurrent enrollment, but the share of students participating may be lower. High school size Student count
1.4: High schools nearby other schools that have Proximity to Number of High Schools within
already adopted concurrent enrollment are more likely to offer the program and have higher student participation rates. adopters 5 miles offering Concurrent Enrollment
1.5: High schools with greater proximity to a Proximity to Distance in miles from nearest
community college will be more likely to adopt community community college
concurrent enrollment and have higher student participation rates. colleges Number of community colleges within 10 miles
7 Some of the schools included in the study serve more grades than 9-12 (e.g. K-12 schools, or secondary schools serving grades 6 or 7 through 12th grade). In all cases, only the population of grades 9-12 is used as the denominator for calculating the share of students participating in concurrent enrollment.


46
Table 4: Variable Descriptions and Sources
Variable Description Source
High school adoption of concurrent enrollment program Dummy variable (yes = 1; no = 0) indicating whether a high school adopts concurrent enrollment in a given year during the study. Colorado Department of Higher Education
College matriculation rate (%) Annual measure of the percent of high school graduates who enroll in a postsecondary institution in the fall immediately following graduation. Colorado Department of Higher Education
School performance rating (%) Annual index measure of the percentage of points earned on state performance framework that includes ACT scores, graduation rates, dropout rates and achievement on state standardized tests. The higher the percentage of points earned, the better the school performed on the measures. Colorado Department of Education
Median household income (logged) Log of the median household income in the past 12 months (in 2014 Inflation-adjusted dollars) for the five year period running from Jan. 1, 2010 Dec. 31, 2014. The aggregated 5-year survey was used to obtain neighborhood-level estimates. U.S. Census Bureau, American Community Survey 5-year estimates (2014)
Free and reduced-price lunch (FRL) eligible (%) Annual measure of the percentage of students eligible for free or reduced-price lunch. Colorado Department of Education
Student count (logged) Annual measure of the log of the total student enrollment count in October of each school year. Colorado Department of Education
Diffusion of concurrent enrollment Number of high schools within 5 miles offering Concurrent Enrollment. Author's calculations using data from the dependent variable and high school addresses
Community college distance Distance in miles from the high school to the nearest community college. Author's calculations using data from the dependent variable and high school addresses
Concentration of community colleges Number of community colleges within 10 miles of the high school. Author's calculations using data from the dependent variable and high school addresses
Charter school Dummy variable (yes = 1; no = 0) indicating whether the school is a charter school. Colorado Department of Education
PSEO participation Dummy variable (yes = 1; no = 0) indicating whether the high school previously offered Post Secondary Education Options (PSEO). Colorado Department of Education
District Setting Categorical variable: Denver Metro, Outlying City, Outlying Town, Remote, Urban-Suburban. Colorado Department of Education


47
Motivation to innovate. The hypothesized motivation to innovate is lower academic achievement levels, which is operationalized using college matriculation rates and CDE's school performance ratings. College matriculation rates measure the proportion of graduates who enroll in any college in the fall immediately following high school graduation, which relies on CDHE data. CDE's performance rating for each school is also included as a measure of student achievement. CDE provides both numerical ratings and categorical ratings (performance, improvement, priority improvement and turnaround) in its annual performance review of school districts and schools. The high school-level performance rating is an index that includes graduation rates, dropout rates, ACT scores and achievement and growth on the statewide standardized assessment. Both matriculation rates and school performance ratings are available on a yearly basis. The annual data, when included in the event history analysis, are lagged one year to avoid problems with causal inference.
If the hypothesis is that lower performance motivates a school to adopt concurrent enrollment, the data for those indicators need to be from a time prior to the adoption year.
Resources. Fiscal capacity and size are two types of resources important to the analysis of policy diffusion and innovation that are included in hypotheses 1.2 and 1.3, respectively. Per-pupil funding at the school level is not available; only district-level data is available. Including the district-level data masks important funding variance at the high-school level. Instead of including a district-level variable, two school-level measures are included in an attempt to capture the level of wealth and resources of individual high school communities. Previous studies have found a positive correlation between per-pupil spending and the wealth of the local community (e.g. Augenblick, Myers & Anderson, 1997). Thus, as a measure of a local community's fiscal capacity, median household income for the neighborhood immediately surrounding the high school is used in this study. Neighborhood-level estimates were obtained from the U.S. Census Bureau's American Community Survey (ACS) by geocoding each high school's address and matching it with a census


48
tract number and county code. The ACS 5-year survey data contains the median household income in the past 12 months (in 2014 Inflation-adjusted dollars) for the five year period running from Jan.
1, 2010 Dec. 31, 2014. The 5-year survey option was used because it offers neighborhood-level estimates.
An additional measure included to capture the resources of a school is the proportion of students eligible for free or reduced price lunch (FRL). Annual data is available from CDE on the percentage of students qualifying for free or reduced price lunch; the data are lagged one year in the event history analysis. FRL eligibility rates correspond with funding levels as Colorado's school finance formula awards additional per-pupil funds to districts for FRL students. Districts, in turn, use their own formulas to distribute state fundsas well as federal Title I dollarsto schools most in need. While the FRL data and the median income indicator are measuring similar information, they are only moderately negatively correlated. There are instances where one measure may more accurately account for the fiscal situation of a school than the other. There are wealthy communities, for example, that have high proportions of FRL students, possibly due to school choice patterns (e.g. median income in one Denver Metro area school is $108,627 and the percent FRL is 72.4%). In contrast, there are schools that have very low median incomes in the surrounding neighborhood but also have low FRL counts, most likely due to underreporting by families (e.g. median income in one remote southwest Colorado school is $39,476 and the percent FRL is 35.3%). Further, while the FRL indicator captures those schools that may receive additional funding support, not many high schools are Title I served, meaning districts more often direct the funds ear-marked for high poverty schools to elementary and middle schools. Schools that are low-income but not Title I served, or schools that serve middle income families who just fall short of meeting FRL eligibility will operate differently from schools that serve mostly high-income families.


49
One example of how such differences operationalize in terms of fiscal capacity, is through parent fundraising. According to an investigative report of parent fundraising at Colorado schools, fundraising levels vary dramatically by school and are correlated with the wealth of the community (Schimke, 2016). One school, for example, raised $14,400 through the one-day "Colorado Gives Day" fundraising campaign in 2016; within the same school district, a school serving a lower-income population raised only $300 (Schimke, 2016). While that is an extreme example, there are concrete differences in fundraising capabilities by school dependent upon the wealth of the parent population, as well as the wealth of the surrounding neighborhood (who are often asked to contribute to school fundraising campaigns). Even in districts that have school choice, controlling for the income levels of the neighborhood immediately surrounding the school building is still critical. Further, funds raised by the school are increasingly used to support instructional and programmatic needs, as opposed to just supporting extracurricular activities (Schimke, 2016). Consequently, both median household income and FRL eligibility are included in an attempt to capture the different fiscal pressures at play at the school level.
The second type of resource important to this studyschool sizeis measured by counting the number of students enrolled in the school during the annual October count period, which is the official method CDE relies on to assess pupil membership. Annual data is available for student count data, and the data are lagged one year in the event history analysis to avoid causal inference problems.
External factors. The main external factors to be measured in hypotheses 1.4 and 1.5 are proximity to schools that have adopted concurrent enrollment and proximity to community colleges. The first factor is operationalized by calculating for each individual high school how many other high schools were already offering concurrent enrollment within a five mile radius. The
calculation was done for each school year included in the study so that the values could change as


50
the program spread. The data are lagged one year in the event history analysis to ensure that the "diffusion of concurrent enrollment" variable is accurately captured as a predictor variable.
The second proximity factor was measured in two ways: 1) by calculating the distance in miles from the high school to the nearest community college, and 2) by calculating the number of community colleges within 10 miles of the high school. These measures adequately account for those high schools that are located in more populous, urban areas with access to multiple colleges. The author geocoded high school addresses using Texas A&M GeoServices. Once the longitude and latitude was obtained for each high school, STATA's geonear command was used to compute geodetic distances8 between individual high schools and community colleges and among the high schools themselves. The geocoded data was also used to match high schools with unique census tract numbers to then link the Colorado administrative dataset with the ACS dataset.
Control variables. Lastly, two additional variables were included in the empirical models to control for possible confounding factors. These include: 1) an indicator for whether a school was a charter school or a traditional school, and 2) an indicator for whether the high school offered a different dual enrollment program prior to 2011. First, an indicator for whether a school is a charter school is included to control for any potential confounding effects on the variables of interest in the case that charter schools act differently than traditional district-run schools, particularly in terms of deciding programmatic offerings. Second, Berry & Berry (2007) suggest that diffusion models likely need to include variables that capture whether prior policies were in place that could impact the decision to adopt the policy currently at hand. In this case, there was a dual enrollment program available to districts in Colorado prior to the Concurrent Enrollment Programs Act passing in 2009. The program, known as Post Secondary Education Options (PSEO), began phasing out in 2009 and
8 Geodetic distances calculate the length of the shortest curve between two points along the surface of a mathematical model of the earth.


51
was fully phased out by 2011-12. If schools that had PSEO in place wanted to continue to offer similar opportunities, they had to make the transition to Concurrent Enrollment by 2011-12 (CDHE, 2014). It was not required that they make the transition, however, and PSEO schools could choose not to offer any dual, or concurrent, enrollment courses once PSEO was phased out (CDHE, 2014). Although substantially different in nature and mechanism, PSEO participation likely is associated with concurrent enrollment adoption and so an indicator will be included to control for if a high school had participated in the PSEO program.
Policy Diffusion Methodology
This section provides an overview of the two methods used to test hypotheses 1.1 through 1.5: event history analysis and ordinary least squares (OLS) fixed effects regression. Event history analysis was the method used for exploring the diffusion of concurrent enrollment across high schools in the state, while OLS fixed effects regression was used to explore any factors related to the intensity of program participation within high schools once concurrent enrollment was adopted.
Event history analysis. Event history analysis was conducted to explore the possible existence of explanatory relationships in the diffusion of concurrent enrollment programs among Colorado high schools. According to Wong and Shen (2002), event history analysis "has become widely accepted as the most effective way to empirically assess the causes of policy innovation in the states" (168).9 The method allows researchers to identify what factors influence events over time. In this study, the event was the adoption of concurrent enrollment by high schools (/'), and time was measured in discrete units of school years (t). The study period ran from the beginning of the 2010-11 school year until the end of the 2014-15 school year. The legislation that created the concurrent enrollment program passed in the spring 2009. In the 2009-2010 school year a small
9 EHA has also been applied to sub-state entities, see e.g. Hoyman and Weinberg (2006) and Lubell et al. (2002).


52
number of school piloted the program, and concurrent enrollment was officially operational statewide in the 2010-11 school year. Thus, the window of this event history analysis captures the first five years of the implementation of concurrent enrollment in Colorado.
Table 5 displays an overview of the number of high schools in the risk set in each school year, the number of adoptions each year, and the survivor function. In an event history analysis, the survivor function expresses the probability that survival time T is equal to or greater than t, where t represents the actual survival time (Mills, 2011).
S(t) = Pr(T > t)
The output of S(t) in this study is, therefore, simply a proportion of the observations that still had
not adopted concurrent enrollment following school year t.
Table 5: Concurrent Enrollment Adoptions and Survivor Functions, by School Year
School Year Number of Adoptions Cumulative Adoptions Risk Set Survivor Function Std. Error 95% Confidence Interval
2010-11 195 195 388 0.50 0.03 [0.45, 0.55]
2011-12 76 272 193 0.30 0.02 [0.26, 0.35]
2012-13 65 337 117 0.13 0.02 [0.10, 0.17]
2013-14 6 343 52 0.12 0.02 [0.09, 0.15]
2014-15 10 353 46 0.09 0.01 [0.07, 0.12]
As Table 5 depicts, the first year saw the largest number of adoptions with fifty percent of the high schools implementing the program that year. The diffusion of the program continued rapidly after that, with eighty-seven percent of high schools offering concurrent enrollment by the third year. As the survivor function indicates, just 9 percent of high schools (n=36) had not adopted concurrent enrollment by the end of the study period in the 2014-15 school year. The high schools that did not adopt concurrent enrollment by the end of the study period are considered to be right-censored. Event history analysis is preferred over other regression models for policy diffusion
studies because it can account for both censored and non-censored observations when producing


53
estimates of the likelihood than an event will occur at a specified point in time (Mills, 2011; Mokher & McLendon, 2009).
Following Berry and Berry (1990), many public affairs scholars used discrete-time logit or probit models to perform event history analysis (Allison, 1984; Berry & Berry, 1990; Buckley & Westerland, 2004; Wong & Shen, 2001). Buckley and Westerland (2004) note that "this approach to testing diffusion theory with discrete event history analysis is straight-forward, computationally economical, and easy to execute, but it has several shortcomings" (95). One limitation is that the discrete-time models assume that the probability of policy adoption in one year is unrelated to the probability of adoption in previous years, when that may not be the case in actuality (Berry & Berry, 1992; Buckley & Westerland, 2004). The odds, for example, that a high school adopts concurrent enrollment in the first year of the study are likely different from the odds that a school adopts the program in the last year of the study when the policy is more popular and eighty-five percent of other high schools have already adopted it. As a result, scholars have turned to the Cox proportional hazards model, which allows for the probability of policy adoption to change over time while not having to specify the functional form (Buckley & Westerland, 2004; Jones & Branton, 2005, Mills,
2011). The semi-parametric nature of the model lends itself to being robust to different data, even if the author does not know the precise underlying shape of the probability distribution (Mills, 2011). The Cox method also permits the incorporation of time-dependent variables. For these reasons, this research design used Cox proportional hazards models to analyze the diffusion of concurrent enrollment.
The Cox proportional hazards regression model relies on maximum partial likelihood estimation when computing hazard rates. The hazard rate is the likelihood that the event of interest will occur in a specified unit of time given that the observation has survived any prior time periods. The Cox model estimates changes in hazard rates as a function of a set of covariates. While the


54
model does not assume a particular shape of the baseline hazard rate, it does make a strong assumption that the ratio of hazard between any two observations is proportional across time (Box-Steffensmeier and Jones 2004; Mills, 2011). If the proportional hazards assumption is violated, the relative risk may be improperly estimated.
To test the proportional hazards assumption, Schoenfeld residuals were estimated and plotted to see if there was a pattern in any of the covariates' residuals that would indicate time-dependency. The variable for district setting clearly violated the proportional hazards assumption.
As a remedy, the Cox model was stratified on the variable, which essentially sets a separate baseline hazard function for each value of district setting. Once the stratification was conducted, there were no further violations of the proportional hazards assumption in individual covariates or for the model as a whole.
The final Cox model specification can be expressed as:
hi(t) = h0(t) exp(x/0)
where the proportional hazard of high school / adopting concurrent enrollment in school year t is the result of an unspecified baseline hazard function h0(t) and a vector of the exponent of the coefficients of parameters ((B) for the constant and time-varying covariates (xy) in the model (Mills, 2011). The Efron approximation was used in estimating the Cox model, which more appropriately handles "tied" data, or when more than one high school adopts concurrent enrollment in the same time period, than the typically-used Breslow approximation.
Additional model diagnostics that were conducted include the estimation and plotting of Cox-Snell residuals to assess overall model adequacy and the plotting of martingale residuals to assess any nonlinearity in the covariates. As a result of analyzing the martingale residuals, two variables were log-transformed to improve linearity: median income and student count. The results of the Cox-Snell residual analysis indicated that the full model specification was an adequate fit.


55
Ordinary least squares fixed effects regression. A progression of ordinary least squares (OLS) regression models were conducted to investigate if any of the covariates included in the event history analysis model serve as predictors for how deeply a high school implements concurrent enrollment. The dependent variable in the OLS models is a school-level participation rate created by dividing the number of students within a high school taking at least one concurrent enrollment course by the total number of students enrolled at the high school for each academic year within the study. The first OLS model includes the same variables used in the event history analysis as predictors. The subsequent models include a series of time and unit fixed effects with and without a lagged dependent variable.
If the dependent variable and one or more independent variables trend in a direction over time, including time fixed effects in the regression model is often a necessary precaution (Wooldridge, 2006). If the dependent variable and one of the key covariates both are trending upward, for example, the two time series processes may appear to be correlated when they are actually both trending for reasons related to factors unaccounted for in the model. Employing dummy variables for each year of the study (excluding the baseline year) controls for spurious trend relationships. If the time dummy variables end up being statistically significant, and the coefficients of other variables change in a meaningful way, that is evidence of the need to include time fixed effects in the regression model (Wooldridge, 2006).
After including time fixed effects, regression models were run with the addition of unit fixed effects to control for possible omitted variable bias. Fixed effects were included, separately, at the district and school levels to control for district- or school-specific variation, which could have a confounding effect on the high school-level model. High schools within districts are likely influenced by district-level factors such as administrative capacity, the presence of a college preparatory
culture, history of pursing partnerships within the district, or fiscal characteristics. Including district


56
fixed effects results in a within regression analysis where the change in covariates is only analyzed within each district. This could provide a good amount of control for any unobserved confounding factors, while also allowing for some across-school variation (within districts that have more than one high school).
Only including district fixed effects, however, still leaves room for doubt that the model is accounting for all unobserved variables. School-level characteristics such as leadership, culture, academic systems (e.g. curriculum and instructional model), and teacher capacity likely also effect the implementation of concurrent enrollment. One disadvantage of school fixed effects is that they absorb nearly all of the "action" since there is very little change in the time-varying predictors within individual schools over the five years of the study (all time invariant predictors are dropped from a school fixed effects model).
Thus, the author runs both a district- and school-level OLS fixed effects regression model, which can be formerly expressed as:
Yit= 60 + oTn-i + Fn-i + Xjt6 + Ha
where the dependent variable (Y) is the concurrent enrollment participation rate for each high school (/) in time period (t) and is a function of time fixed effects (Tn-i), school or district fixed effects (Fn-i), a vector of the coefficients of parameters (6) for the time-varying covariates (Xu), and the error term (iuit).
Two prominent issues with running OLS regression on time series data are serial correlation and heteroskedasticity. Serial correlation, or autocorrelation, refers to the correlation of error terms among observations and is often present in time series data since the same unit is being measured in repeated time periods (Wooldridge, 2006). If the value of a covariate in one time period is related to its value in the previous time period, for example, then the error terms are likely to be correlated. Serial correlation does not bias the estimates but it does result in an underestimation of the


57
standard errors, which in turn inflates the t-statistic and leads to overestimation of statistical significance. Heteroskedasticity, or a violation of the OLS assumption that error variance is constant, likewise affects the calculation of standard errors and results in misleading claims of statistical significance (Wooldridge, 2006). Diagnostics tests were run on the time series data set and found that both serial correlation and heteroskedasticity were present. To correct for both issues, robust school-clustered standard errors are used in the regression models.
Another issue with running OLS models on this data set in particular stems from the functional form of the dependent variable, which is a percentage bound between 0 and 100. Using OLS simplifies the interpretation of the regression results, and post-estimation analysis found that only 20 of 1820 (1.10%) of predicted values fall outside of the 0 to 100 range (see Papke, 2005 for a similar approach). The regression models were also run using fractional logit to ensure that the results from OLS models are robust and not affected by misspecification error (Papke & Wooldridge, 1996). The fractional logit models substantiated the statistical significance and direction of the OLS results.
Dynamic panel data model. Even though the OLS models control for year and unit fixed effects, there remains a need to investigate the effect of the prior year's participation rate on the current year. Practical reasoning would lead one to suspect that a high school's concurrent enrollment participation rate for one year would be highly predictive of the following year's participation rate. Including a lagged dependent variable with fixed effects in the same OLS model, however, may lead to biased estimates as a result of correlation between the error terms and the covariates (Allison, 2009; Wooldridge, 2010). Economists refer to models that include lagged dependent variables as dynamic panel data models, and there are several approaches that can be used for estimation (Allison, 2009; Williams, Allison & Moral-Benito, 2016; Wooldridge, 2010). This
study employs an approach that uses maximum likelihood estimation and allows for the inclusion of


58
time invariant predictors while still retaining the benefits of fixed effects (Williams, Allison & Moral-Benito, 2016). Other approaches to modeling dynamic panel data, including traditional fixed effects methods and generalized method of moments (GMM), exclude time invariant predictors. As explained above, there is not much within school variation on the time varying predictors, so the inclusion of additional covariates in the model is beneficial to understanding patterns in concurrent enrollment participation rates. The dynamic panel data model using maximum likelihood estimation was run with the dependent variable lagged one year. Full information maximum likelihood (FIML) was used to treat missing data. About 5 percent of schools are missing data on matriculation rates, and using FIML allows for those schools to remain in the estimation by using the data that is available for those schools rather than using list-wise deletion and losing those observations altogether (Arbuckle, 1996).
Policy Evaluation Variables and Measures
The following section provides details on the dependent, explanatory and control variables used in the evaluation of concurrent enrollment. Table 6 summarizes the indicators used to operationalize the variables in the two policy evaluation hypotheses, and Table 7 provides a summary of variable descriptions.
Table 6. Concept Measurement Summary: Policy Evaluation
Hypotheses Constructs Indicators
H2: High school students who College access Immediate enrollment in college
participate in concurrent enrollment Concurrent following high school graduation
programs will have a greater enrollment Concurrent enrollment participation (y/n)
probability of enrolling in higher education. participation Number of concurrent enrollment credit hours
H3: First-year college students who Academic Need for remedial education
had participated in concurrent success & First-year college grade point average
enrollment programs in high school persistence (GPA)
will have greater academic success and Concurrent Fall-to-fall college persistence
a higher probability of persisting than enrollment Concurrent enrollment participation (y/n)
college students who did not participate. participation Number of concurrent enrollment credit hours


59
Table 7. Descriptions of Pre-College Independent Variables and College Outcome Variables
Variable Description
Concurrent enrollment participation Concurrent enrollment credit hours Dummy variable (Took any concurrent enrollment course = 1) Categorical variable for number of concurrent enrollment credit hours attempted (0,1-3, 3-6, 6-12 or 12+)
Student academic characteristics ACT composite score English language learner Special education Continuous variable (min=12; max=36) Dummy variable (Students designated as ELL=1) Dummy variable (Students designated as SPED = 1)
Student and family background White African American Hispanic Asian Other race Gender Free or reduced-price lunch (FRL) Dummy variable (white = 1) Dummy variable (African American = 1) Dummy variable (Hispanic = 1) Dummy variable (Asian = 1) Dummy variable (other race = 1) Dummy variable (male = 1) Dummy variable (FRL-eligible students = 1)
School environment Rural/urban school district Dummy variable (Rural = 1)
College outcomes College enrollment Remedial education need First-year college grade point average (GPA) College Persistence Dummy variable (Enrolled in college anywhere in the fall immediately following high school graduation = 1) Dummy variable (Needed remedial education in at least one math, reading or writing course = 1) Cumulative grade point average in the spring semester of a student's first-year in college Dummy variable (If enrolled in year one and enrolled in year two of college anywhere = 1)
Dependent variables. The dependent variable in Hypothesis 2 is college enrollment, which was measured by considering those students who enrolled in college in the fall immediately following high school graduation. Students who enrolled in college anywhereat an in-state, out-of-state, public or private institutionare captured. This is a dichotomous variable; students who enrolled in college were coded as a 1.
There are several dependent variables that are operationalized from Hypothesis 3. First, students' need for remedial education in college is included as a measure of academic performance. The measure includes both students assessed as needing remediation and those enrolled in
remedial courses who did not have an assessment score on file. This is a dichotomous variable;


60
students who need remedial education in college are coded as a 1. While college-level concurrent enrollment courses require students to be college-ready in the course's subject area (as determined by an appropriate assessment), students could need remedial education in a different content area. For example, a student may take a concurrent enrollment, college-level literature course, but that student may require remediation in math. Further, some students only take career and technical education (CTE) concurrent enrollment courses that do not have the same academic prerequisites as core subject areas. Thus, remedial education is considered to be a worthwhile measure of academic preparation and success to include in the model.
Second, students' postsecondary academic success is assessed by considering the cumulative grade point average after the spring semester of the first year in college. Course-level data is only available through the state's administrative data collection, and thus only students who enrolled in a public college in Colorado are captured in the calculation of remedial rates and grade point averages (CDFIE, 2016). While this is a limitation of the dataset, approximately 75 percent of high school graduates who enrolled in college did so at an in-state, public institution and are captured in the state's data system.
The third dependent variable that is measured in the second hypothesis is college persistence, which is measured using a dummy variable indicating whether a student who enrolled in year one of college returned to enroll in year two of college (returned to any institutionnot just the original institution). Because enrollment data is available through both the state administrative system and the National Student Clearinghouse, the data for this variable includes all students who enrolled in college anywhere, not just those who enrolled in Colorado.
Concurrent enrollment participation. The key explanatory variable for both hypotheses is participation in concurrent enrollment. This study employed two measures of participation: 1) a
dichotomous measure of participation, in which students who graduated high school having taken at


61
least one concurrent enrollment course were coded as a 1; and 2) a categorical measure based on the attempted number of concurrent enrollment credit hours. There are five categories for the credit hours measure: no credit hours, 1-3 credit hours, 3-6 credit hours, 6-12 credit hours and more than 12 credit hours. Zero credit hours is set as the baseline category in the analysis. The remaining four categories were selected after viewing the descriptive statistics and seeing natural breaks between each category that equate roughly to quartiles.
Demographic pre-college independent variables. The empirical analysis included demographic and geographic control variables that, based on prior research, are thought to influence concurrent enrollment participation, college-going behavior and postsecondary outcomes. These measures include gender, high school free and reduced-price lunch (FRL) status, special education (SPED), English Language Learner (ELL) status, race/ethnicity, and ACT scores. The data for FRL, SPED, and ELL students are reported by high schools to the Colorado Department of Education and indicate whether a high school graduate received free or reduced-price lunch, was identified as special education, or was identified as ELL, respectively. Race/ethnicity is self-reported by students to schools and was measured here using dummy variables for African American students, Hispanic students, white students (the baseline group), Asian students and an "other race" category that includes American Indian/Alaskan Native or Hawaiian/Pacific Islander students.10 Gender, FRL, SPED, ELL and race/ethnicity fields are required components of the datasets schools submit to the Colorado Department of Education and there are no missing data. Lastly, a dummy variable for rural schools was included (when school-level effects were not utilized) to capture school-level differences attributable to geographic setting.
10 The categories of race/ethnicities used in this study are representative of the largest groupings of students and were necessary to accurately run the propensity score matching (PSM) analysis. Including separate, smaller sub-groups of students in the analysis did not change the end results but substantially reduced the likelihood of achieving non-biased matches during the PSM analysis.


62
Academic pre-college independent variable. The composite ACT score was used as a proxy control for academic achievement. ACT scores were an important variable to include because performance on the college entrance examination is highly correlated with college attendance. Further, ACT scores are strongly correlated with assessment scores from the statewide accountability tests administered in 9th and 10th grades, indicating the scores are a good control for overall academic aptitude.11 ACT subject scores (reading, writing, math and science) also were collected, but in the effort to achieve a more parsimonious model, composite scores were used.12
During the period of this study, Colorado required all high school juniors to take the ACT since (ACT, Inc., 2009). The test was provided free to students, and one day of the academic year for juniors was devoted to taking the ACT. Nonetheless, some students opted-out of taking the assessment. In addition, when the data from the ACT were matched with postsecondary data from the CDHE, some records were not matched successfully. As a result, across the three high school graduating cohorts, 15.0 percent had missing ACT scores. Multiple imputation was initially used as a treatment for the missing data, and the results did not alter significantly from those that are presented later in the chapter. Thus, list-wise deletion was ultimately used to eliminate the observations with missing ACT data for ease of interpretation. This method does have disadvantages because the population of students with missing data differed from the general population. Those with missing data were less likely to attend college (24.6%) than the students with ACT scores (63.2%), and they were less likely to participate in concurrent enrollment (6.8% compared to 15.4%).
11 The statewide assessments administered in 9th and 10th grades had slightly fewer missing values than the ACT assessment and could be used an alternative measure for academic aptitude. However, students taking concurrent enrollment are most likely to do so in the 11th and 12th grade of high school, and with the ACT being administered in the 11th grade and being designed to assess college readiness, it is a timelier source of performance data and, arguably, a more reliable and valid control variable for the hypotheses being tested here than the statewide grade-level assessments. However, regression models were estimated using the 9th and 10th grade data, which confirmed the statistical significance and direction of the resulted presented here
12 Models that were run with subject scores did not vary from the models run with the composite score.


63
However, nearly one third (32.7%) of students with missing ACT scores were attending an Alternative Education Campus (AEC). On average throughout the study, only 2.3 percent of students were enrolled at an AEC. It is common for AECs to have missing data given that they serve a highly mobile and transient population of at-risk students. AECs typically also have very low college matriculation rates (16.4% compared to 57.7% at traditional high schools in 2014). It is likely that the AEC school status (and what that represents) is a bigger predictor of college outcomes than the ACT scores would be if the data were not missing. As explained later, school fixed effects are used in the final model of the regression analysis to control for such unobserved bias. Further, as stated above, the results from regressions that were estimated following multiple imputation for the missing ACT scores confirmed the findings presented below.
Policy Evaluation Methodology
Multivariate, fixed effects regression and propensity score matching (PSM) were used to determine relationships between concurrent enrollment participation (the dichotomous measure) and the key dependent variables. The average treatment effect was calculated for both methods, which allows for comparisons of the techniques and provides a triangulation of the findings to assess how college outcomes for students who participate in concurrent enrollment are affected as compared to non-participating students. When considering how interactions between race/ethnicity and concurrent enrollment participation affect college outcomes, the author conducted the analysis solely with fixed effects regression given the significant complications posed by including interaction terms in PSM (Garrido et al., 2014; Imbens, 2000). The same methodological complications arise when using categorical predictor variables in PSM; thus, fixed effects regression was used when analyzing the effect of the number of concurrent enrollment credits on college outcomes.
Overview. Randomized controlled trials are the ideal method because both observed and
unobserved factors are accounted for through the process of randomly assigning individuals to the


64
treatment group (Schneider, Carnoy, Kilpatrick, Schmidt & Shavelson, 2007; Singleton & Straits, 2010). In nonrandomized designs, selection biaswhen the participants in the treatment group differ systematically from those who did not participate in the treatmentis a threat to internal validity (Singleton & Straits, 2010). In education research, as in the social sciences generally, it is often difficult to adhere to experimental designs due to practical, situational, or ethical considerations (Titus, 2007; Winship & Morgan, 1999). As a result, researchers have developed and refined analytical techniques that attempt to create quasi-experimental conditions that control for selection bias (Heckman, 1979; Rubin, 1974, 1997; Schneider et al., 2007; Winship & Morgan, 1999). Both multivariate regression and PSM are techniques that fill that role. Multivariate regression allows the author to control for confounding or intervening variables that affect the relationship between the treatment and the outcomes (Singleton & Straits, 2010).
PSM, developed by Rosenbaum and Rubin (1985), has a different focus and controls for those observed variables that affect whether an individual participates in the treatment or not. The theory behind PSM asserts that, if assignment to the treatment is driven solely by observable factors, then after the matching is conducted analysis of the treatment effect can proceed as if assignment was random (Rosenbaum & Rubin, 1985; Winship & Morgan, 1999). PSM begins by generating propensity scores based on observable variables that predict the likelihood that an individual will receive the treatment. Those scores are used to match individuals who received the treatment with similar individuals who did not receive the treatment. Through the matching process a control group is essentially created, which then allows the author to mimic experimental conditions, establish a counterfactual, and estimate the treatment effect (Guo & Fraser, 2010; Rosenbaum & Rubin, 1985; Rubin, 1997; Winship & Morgan, 1999).
PSM has become popular amongst applied researchers dealing with observational data
(Caliendo & Kopeinig, 2008; Dehejia & Wahba, 2002, 1999; Heckman, Ichimura & Todd, 1997). As


65
the number of studies using PSM has increased, the debate around whether PSM is any better of a method than standard regression analysis persists (Brand & Halaby, 2006; Shadish & Steiner, 2010; Shadish, Clark & Steiner, 2008; Shah, Laupacis, Hux & Austin, 2005; Smith & Todd, 2005). Both PSM and multivariate regression analysis condition only on observable variables and several studies have found that regression estimates tend to be similar to PSM estimates (Cook, Shadish, &Wong, 2008; Shadish & Steiner, 2010; Shah, Laupacis, Hux & Austin, 2005). Nonetheless, there are some advantages to PSM. First, PSM estimations avoid bias caused by a misspecification of the functional form, which occurs frequently in regression analysis. As Brand & Halaby (2006) explain, "although matching assumes selection on observables, it does not assume linear selection as does covariate adjustment through regression" (757).
Second, there are several balancing tests that can be performed using PSM that provide information regarding the validity of causal inferences from the data set. A data set with covariates that are balanced between the treatment and control groups is not enough to fully eradicate selection bias concerns if there are unobservable confounders present; but, having a dataset with balanced observable covariates and areas of common support is a minimum requirement for making causal inferences (Shadish & Steiner, 2010). Regression analysis does not typically limit estimates to areas of common support, or to parameters of variables where both treatment and control observations exist (Brand & Halaby, 2006).
Third, there are diagnostic tests allowing the author to assess the sensitivity of the treatment effects to unobservable bias. These tests do not conclusively determine the level of unobserved bias, but provide ways for the research to lend credence to findings in the case of theorized selection bias (Caliendo & Kopeinig, 2008).
Propensity score matching. The pre-college exogenous variables described in Table 2 were
used to generate propensity scores, after which the results were evaluated to ensure there was an


66
even distribution of propensity scores across treatment and comparison groups. PSM requires there to be an individual in the comparison group with a similar propensity score for each individual in the treatment group to make inferences from the results (Garrido et al., 2014). The sample is divided into a number of blocks that is sufficient enough to ensure equal mean propensity scores between treatment and comparison groups. Figure 2 displays the visual output from the results and indicates that there was a satisfactory overlap in propensity scores between the treatment and control groups and an appropriate range of propensity scores.
Figure 2. Distribution of Propensity Score Across Treatment and Comparison Groups
Another important verification step is to check if the propensity scores were accurately specified by ensuring that the covariates are individually balanced in each block of the propensity score for both the treatment and comparison groups (Garrido et al., 2014; Imbens, 2004). Typically, initial specifications are not balanced and variables need to be dropped or transformed. In this study, several iterations of generating propensity scores were performed until balance was achieved.
Initially, dummy variables for each high school (n=423) were included, but balance in the covariates could not be reached given the large number of high schools and the inability to produce


67
matched samples on each covariate within each high school. In the final specification, school fixed effects were not included, but an indicator for rural or urban high schools was included to control for some school-level effects. The race/ethnicity categories were redefined from seven categories to four categories (Hispanic, white, African American, other race). Additionally, the variable for ACT scores was converted from a continuous variable to a categorical variable.13 After making those changes, balance in the covariates was satisfactorily achieved across the blocks using t-tests. Some level of imbalance is expected and acceptable (Austin, 2009), but to further ensure balance had been achieved standardized differences were computed for the covariates across the blocks. The PSM literature has set forth an acceptable amount of imbalance as being maximum standardized differences of 10 to 25 percent (Garrido et al., 2014). The results of this study achieved standardized differences no larger than 2.5 percent. Thus, it was concluded that sufficient balance was achieved.
Treated individuals were matched with comparison individuals who had the most similar propensity score within a certain range of scores, referred to as a caliper. Keeping the matches within a range, or caliper, prevents poor matches from occurring. Here, a caliper of 0.2 of the standard deviation of the logit of the propensity score was used based on prior research that has found that range to produce optimal results (Austin 2011; Rosebaum & Rubin, 1985). One-to-one matching produces the least biased estimates since the first match is always the strongest match, especially given the restriction of the caliper which prevents poor matches. One-to-many matching, on the other hand, increases bias (from poorer matches) but decreases the variance of the estimates by including more counterfactual information for each treated subject (Caliendo & Kopienig, 2008). This study used both one-to-one matching and one-to-four matching, in which each
13 ACT scores were divided into the following categories: Low ACT score = 5 to 16; Medium-Low ACT score = 17 to 20; Medium-High ACT score = 21 to 24; High ACT score = 25-36. The analysis was also run using a continuous variable for ACT scores and the average treatment effect sizes did not vary, but there was more bias present during the initial stages of the PSM process due to the inability to find precise matches with the continuous variable.


68
treatment unit is matched to four control units. Other matching techniques such as kernel, radius and stratification matching were tested, but the one-to-one and one-to-four caliper matching with replacement produced the best balance in covariates across the treatment and control samples, as measured by standardized differences.
Regression analysis. Regression models were estimated, separately, for both concurrent enrollment independent variables (dichotomous and credit hours) on the four college outcome dependent variables. Initially, bivariate models with the dependent variable and the key independent variable were run. Then demographic control variables (gender, FRL status, race/ethnicity) were added, followed by the addition of the controls for academic achievement (ACT composite score, ELL, SPED). The final models included all of the previous controls and added fixed effects for high school and graduation year; this was considered the preferred model specification for both main research questions. Adding fixed effects helps alleviate concerns about omitted variables. In particular, school-specific featuressuch as the availability of college guidance counselors, the presence of a college preparatory culture, school leadership and school location (e.g. rural, urban, suburban)vary and could have a confounding effect on the model. Additionally, adding in time fixed effects for each year of the study (minus the baseline year) controls for any spurious trend relationships occurring over the three-year time period of the study.
When investigating the effects of concurrent enrollment participation on minority and low-income students, interaction terms were added to the fixed effects regression models by crossing each race/ethnicity variable and the FRL variable with concurrent enrollment participation (yes/no). The author used logistic regression for the dichotomous dependent variables (college enrollment, remedial education need, and persistence) and OLS regression for the single continuous dependent variable (first-year GPA). Another way to investigate differential effects by disaggregated groups of students is to divide the sample by student group and run separate regression models. Using a


69
pooled regression analysis with interaction terms was selected as a more parsimonious and efficient way to approach analyzing the effects of concurrent enrollment on different groups of students, but both approaches should, in theory, obtain the same outcomes. The findings chapter reports results from the average treatment effects calculation that was run after the regressions using STATA's teffects and margins suite of commands (Williams, 2012; Wooldridge, 2010).
Data & Methods: Summary
This chapter described the data and methods used to explore the research questions and associated hypotheses. Table 8 provides a summary of the methodological approaches for each research question.
Table 8: Methodological Approaches with Associated Research Questions
Analytic Methods Unit of Analysis Research Questions
Event History High School RQ1: What factors influence whether high schools adopt
Analysis concurrent enrollment programs?
Fixed Effects Regression & Dynamic Panel Data Model High School RQla: What factors influence the extent to which concurrent enrollment programs are utilized by students within high schools?
Student (high RQ2: How does participation in concurrent enrollment affect the
Propensity Score school graduate) college-going rates of Colorado's high school students?
Matching & Fixed Effects Regression Student (college student) RQ3: How does high school participation in concurrent enrollment affect the college performance and persistence of students?
The study begins with an exploration of factors that influence the adoption of concurrent enrollment programs among and within Colorado high schools using school-level, administrative data in event history analysis and multivariate regression. The research then focuses on using student-level data in propensity score matching and fixed effects regression to analyze the effects of
participating in concurrent enrollment on college matriculation and success. All components of the


70
study rely on data from Colorado's statewide longitudinal data system, but the research questions are explored using different units of analysis and methods. This research design allows for a comprehensive investigation of the effects of Colorado's concurrent enrollment policy.


71
CHAPTER IV
POLICY DIFFUSION FINDINGS & DISCUSSION
This section presents findings from the descriptive statistics of the dataset and from the inferential statistics used to analyze the first research question regarding what factors lead some schools to adopt concurrent enrollment more quickly and implement the program more intensely as compared to other schools. The results from the event history analysis are presented following the descriptive statistics; the OLS fixed effects regression results are discussed thereafter.
Descriptive Statistics
The diffusion of concurrent enrollment throughout Colorado high schools was rapid and nearly complete by the end of the five year study, in 2015. Figure 3 provides a visual of program adoption by school districts and high schools during the first 5 years after the state legislation passed.
Tables 9 and 10 contain descriptive statistics for the data. As Table 9 depicts, the means of the time-varying covariates did not alter dramatically from 2011 to 2015, with the exception being the diffusion of concurrent enrollment indicator, which captured the rapid increase in the number of high schools with concurrent enrollment (and the corresponding increase in the number of observations that had nearby high schools offering the program). All time-varying covariates were lagged one year to ensure that they were being accurately captured as predictor variables. As an example, for the 2010-11 school year, the time-dependent covariates reflect values from the 2009-10 school year.
Table 10 displays the mean values of the covariates by year of concurrent enrollment adoption. In 2011, for example, 195 high schools adopted concurrent enrollment. The average matriculation rate for those high schools for the year prior to adoption was 59.1 percent, which was higher than the average rates for high schools adopting later and nearly twice the mean for high


72
schools that did not adopt concurrent enrollment during the study period (31.1%). The mean distance to the nearest community college was higher for schools adopting earlier (2011-2013) than for schools adopting later, or not adopting at all.
2010-11 (Year 1 of study) High Schools 2014-15 (Year 5 of study)
Figure 3. Adoption of Concurrent Enrollment Programs from the 2010-11 School Year to the 2014-15 School Year, by School Districts and High Schools. The upper two maps indicate that school districts have at least one high school offering concurrent enrollment if the district boundary is shaded. In the bottom two maps, each dot represents an individual high school that offers concurrent enrollment.
The variable for charter schools also reveals differences between cohorts. Of the 195 schools adopting concurrent enrollment in 2011, only 4 percent were charter schools even though charters comprise about 13 percent of all high schools in the study. Charter schools are overrepresented, though, in the later adoption years of 2014 and 2015. Additionally, of those high schools that did not adopt concurrent enrollment, 43 percent were charter schools. Another point
of interest is the variable for prior participation in dual enrollment (PSEO), which indicates that 63


73
percent of the first cohort of adopters had previously had the PSEO program in place; the mean of that variable declines as the study progresses. Only 8 percent of the 37 schools that did not adopt concurrent enrollment had PSEO. Lastly, Denver metro area and rural schools (outlying city, outlying town and remote schools) tended to adopt concurrent enrollment on the earlier side of the study period, whereas urban-suburban schools adopted in the later years. Of the non-adopters, about half
are located in the Denver metro area.
Table 9: Descriptive Statistics for All High Schools, Beginning and End of Study
Time-Varying Covariates 2011 M (SD) Min Max 2015 M (SD) Min Max
College matriculation rate (%) 53.33 0 100 (21.93) 53.07 3.8 92.3 (20.01)
School performance rating 63.50 25 100 (17.32) 68.74 27.6 100 (13.71)
Student count (logged) 5.76 1.39 8.16 (1.30) 5.76 1.61 8.16 (1.32)
Diffusion of concurrent enrollment (# of CE High Schools within 5 miles) 0.15 0 5 (.611) 4.56 0 26 (5.98)
Free/reduced-price lunch students (%) 37.68 0 100 (23.08) 42.64 0 100 (23.12)
Fixed Covariates 2011 2015 M (SD) Min Max
Median household income (logged) 10.89 9.68 11.81 (.37)
Community college distance 11.72 0 67.11 (13.00)
Concentration of community colleges 1.91 1 6 (1.35)
Charter school 0.13 0 1 (.33)
PSEO participation 0.52 0 1 (.50)
District Setting
Denver Metro Area 0.34 0 1 (.47)
Outlying City or Town 0.22 0 1 (.42)
Remote 0.23 0 1 (.42)
Urban-Suburban 0.21 0 1 (.40)
N 388
Standard Deviations (SD) in Parentheses


74
Table 10: Comparison of Variable Means, by High School Adoption Year
If adoption year was: 2011 2012 2013 2014 2015 Did Not Adopt
Time-Varying Covariates Lagged M* /W**
College matriculation rate (%) 59.06 53.15 52.89 54.50 33.53 31.14
School performance rating 65.36 65.95 70.76 77.68 56.43 62.01
Student count (logged) 5.96 5.95 5.72 5.60 5.13 4.67
Diffusion of concurrent enrollment (CE) (# of CE High Schools within 5 miles) 0.22 1.66 1.80 5.83 4.3 2.33
Free/reduced-price lunch students (%) 38.06 36.55 38.71 37.47 45.15 45.06
Fixed Covariates Fixed M (2011-2015) /W**
Median household income (logged) 10.90 10.88 10.88 10.84 10.94 10.86
Community college distance 11.65 13.06 12.18 9.24 7.56 9.95
Concentration of community colleges 1.97 1.68 1.77 1.83 1.90 2.33
Charter school 0.04 0.08 0.22 0.33 0.40 0.43
PSEO participation 0.63 0.57 0.43 0.17 0.20 0.08
District Setting
Denver Metro Area 0.40 0.35 0.12 0.00 0.20 0.51
Outlying City 0.26 0.24 0.09 0.17 0.40 0.19
Remote 0.24 0.21 0.31 0.17 0.00 0.16
Urban-Suburban 0.10 0.21 0.48 0.67 0.40 0.14
N 195 76 65 6 10 36
*Meansfor time-varying covariates for high schools adopting concurrent enrollment in 2011 through 2015 are lagged one year to reflect values for the year prior to adoption
**Meansfor time-varying covariates for high schools that did not adopt are an average of values from 2011-2015
Figure 4 depicts the mean percentage of high school students participating in concurrent enrollment over time, by adoption year cohorts. In the first year the concurrent enrollment program was fully operational (2010-2011), 195 high schools adopted concurrent enrollment, and, on average, about 12 percent of the students in those high schools participated in the program. By 2015, for those same 195 high schools, the mean participation rate increased to 18.2 percent. In 2012, 76 high schools adopted concurrent enrollment, and in those schools about 8.1 percent of students, on average, took at least one concurrent enrollment course. Participation rates were
similar at the 65 high schools that adopted the program in 2013. Both the 2012 and 2013 cohorts of


75
adopters saw increases in average participation rates overtime. Very few high schools adopted concurrent enrollment in 2014 and 2015, and the mean participation rates at those 16 high schools was quite smallranging from 1.9% to 2.4%.
2010-11 2011-12 2012-13 2013-14 2014-15
School Year
Year of Concurrent Enrollment Adoption 2011 12012 2013 2014 2015 State Mean % CE Participation n=195 n=76 n=65 n=6 n=10
Figure 4: Average Percentage of High School Students Participating in Concurrent Enrollment (CE) within High Schools, by Adoption Year Cohort from 2010-11 to 2014-15. If a high school adopted CE in 2011 (n=195) it represented in the dark shaded bar on the far left of the series. The dotted line represents the statewide average of the percent of students participating in CE during a school year.
While all adoption year cohorts increased average participation rates over time, regardless of the year of adoption, there are apparent differences in the level of participation by year of adoption, with early adopters having higher starting levels of participation rates as compared to those adopting later. This indicates that the degree to which a school uses the program is dependent upon not only how many years the program has been in place, but also whether or not the school was an early adopter. It also likely suggests that early adopters had prior dual enrollment programs in place. This pattern is also depicted in Figure 5, which presents descriptive statistics of participation rates in map form for key variables of interest: percent FRL, matriculation rates, PSEO
participation and year of program adoption.


76
2.22
Had PSEO
1
100.00
O 20.00
40.00
60.00
50.00
93.00
0.0
100.0
Adoption Year 2011 2012
2013
2014 2015
Figure 5. Maps of Colorado high schools and Concurrent Enrollment (CE) participation rates by covariates of interest. Each circle represent a high school; the size of the circle reflects the percent of students participating in CE in 2014-15 school year. Shading in the top left map indicates % of students who were FRL eligible in the 2013-14 school year, with darker shading indicating higher FRL rates. The top right map reflects matriculation rates with high rates designated by darker shading; bottom left map displays schools that did not offer PSEO (dark blue) in contrast to those that did offer the prior dual enrollment program (light blue); bottom right map depicts the year of program adoption with darker circles representing high schools who adopted earlier in study.


77
In the Figure 5 maps, each dot represents a high school, with larger circles representing higher participation rates. The maps generally reveal there is variation across the state on these indicators. Some schools, for example, have high concurrent enrollment participation rates and high percentages of FRL students, while other schools with high participation rates have low FRL percentages. Interestingly, the PSEO map reveals several instances of high schools that have high participation rates that did not previously have PSEO in place (i.e., large, dark shaded circles). The year of adoption map appears to be dominated by darker shaded circles representative of early adopters that have higher participation rates, which was the pattern seen in Figure 4.
Event History Analysis
The findings of the multivariate event history analysis are presented in Table 11. The Cox proportional hazards model results are displayed as exponentiated coefficients, referred to as hazard ratios. A ratio that is greater than 1 indicates the high school is more likely to adopt concurrent enrollment as the values of the covariate increase. Hazard ratios that are less than 1 indicate that a high school is less likely to adopt concurrent enrollment at higher values of the covariate. A ratio of 1 is interpreted as there being no association between the covariate and the hazard of adopting concurrent enrollment. Each hypothesis is tested separately with the control variables included, and then the full model is specified. The results are largely robust across the models. Three of the five hypotheses contain statistically significant findings and both of the control variables are statistically significant.
Looking at the Hypothesis 1.1, the coefficient for matriculation rates is statistically significant at p<0.01, but the direction is the opposite of what was hypothesized. For every 1 percentage increase in matriculation rates, the likelihood of adopting concurrent enrollment increases by an estimated 1.5 percent. It was hypothesized, based on the literature, that schools
with lower college-going rates might be quicker to adopt the program given its link to improving


78
matriculation and the school's need to improve outcomes. While the direction is not what was hypothesized, it is also not surprising; it would be easy to justify having written the hypothesis in the opposite direction given that high schools with already-established cultures of college readiness would also be more likely to take advantage of the concurrent enrollment program.
Table 11: Cox Proportional Hazards Model Results
Relevant Hypotheses Variable Model 1 Model 2 Model 3 Model 4 -Full Model-Model 5
Hypothesis 1.1 Matriculation Rates (%) 1.015*** 1.015***
(0.00308) (0.003)
School Performance 0.992* 0.998
Rating (%) (0.00424) (0.005)
Hypothesis 1.2 Median Household 1.125 1.035
Income (logged) (0.168) (0.159)
FRL eligibility (%) 1.003 1.012***
(0.00264) (0.003)
Hypothesis 1.3 Student Count 1.316*** 1.224***
(logged) (0.0679) (0.082)
Hypothesis 1.4 Diffusion of 0.967 0.976
Concurrent Enrollment (0.026) (0.027)
Hypothesis 1.5 Community college 0.994 0.998
distance (0.005) (0.006)
Community college 1.064 1.015
concentration (0.059) (0.056)
Controls Charter school 0.437*** 0.393*** 0.447*** 0.435*** 0.469***
(0.0905) (0.0820) (0.0916) (0.088) (0.111)
PSEO participation 1.571*** 1.860*** 1.667*** 1.833*** 1.494***
(0.171) (0.206) (0.183) (0.200) (0.164)
Observations 727 782 789 790 722
Likelihood ratio 79.15 75.48 111.6 79.51 108.1
df 4 4 3 5 10
Prob>chi2 0.0319 0.0276 0.0368 0.0265 0.0395
*** p<0.01, ** p<0.05, p<0.1 Robust standard errors in parentheses Coefficients are expressed as hazard ratios
All models are stratified by District Setting (Remote, Outlying City, Urban-Suburban, Denver Metro Area)
The second hypothesis concerning fiscal capacity has mixed findings. High schools with higher percentages of free and reduced price lunch-eligible students have higher hazard ratios. This statistically significant coefficient indicates that for every 1 percentage point increase in the


79
percentage of FRL-eligible students, the likelihood of adopting concurrent enrollment increases by an estimated 1.2 percent.14 The coefficient on median household income is not significantly different from 1.
Hypothesis 1.3 is statistically significant and in the predicted direction. Larger schools have a higher chance of adopting concurrent enrollment more quickly than smaller schools. Hypotheses 1.4 and 1.5 were both rejected with nonsignificant results on the diffusion of concurrent enrollment measure and the proximity to community colleges measures. The model indicated no association between having more neighbors offering concurrent enrollment and time to adoption. One reason for this finding may be how quickly the program diffused throughout the state that the impact of "peer pressure" was not captured in this model. The findings do also reinforce other studies that did not find statistically significant effects of regional diffusion (Mokher & McLendon, 2009). The coefficients for distance to the nearest community college and number of nearby community colleges were also nonsignificant. This is not a surprising finding given the descriptive statistics, which reveal that high schools that were first to adopt concurrent enrollment were as likely to be located in rural areas where community colleges are few and far between as in urban areas where community colleges are more prolific.
Lastly, as expected, both control variables were statistically significant. The chance of adopting concurrent enrollment was about 1.5 times greater for those high schools that previously offered the PSEO program compared to those schools that did not have PSEO in place. Charter schools were half as likely to take up the concurrent enrollment program as compared to traditional schools as indicated in the results and in Figure 5. The hazard function in Figure 5 provides a visual display of the large magnitude of the effect of charter schools on the likelihood of adopting concurrent enrollment. As a means of comparison, the hazard function for matriculation rates is also
14 Percent changes were calculated using the formula (eb- 1)*100.


80
displayed. While the coefficient was highly significant for the matriculation rates indicator, the effect size was not as large as that of the charter school variable, which is seen in Figure 5. Both graphs also indicate a relative proportionality in the hazard functions, supporting the Cox proportional hazards assumption.
Figure 6. Cox Proportional Hazards Regression Smoothed Hazard Functions for Charter Schools and College Matriculation Rates. Non-charter schools (dotted line, left) and high schools with high college matriculation rates (75th percentile dotted line, right) had a higher likelihood of adopting concurrent enrollment as compared to charter schools (solid line, left) and high schools with low matriculation rates (25th percentile solid line, right), respectively.
OLS Fixed Effects Regression Analysis
The findings of the OLS regression analysis are presented in Table 12 and Figure 6. The first model consists of the same variables that were included in the event history analysis (EHA) Cox proportional hazards regression model. While the Cox regression was stratified by district setting, the OLS regression version includes district setting categories as dummy variables, with Denver Metro Area serving as the baseline, or reference, group. The second model builds off of the first and also includes time fixed effects by adding dummy variables for school year, with the 2010-11 school year serving as the baseline group. The third model expands on Model 2 by including district fixed effects. The district fixed effects are absorbed due to the large number of districts (n=178);
consequently, coefficients are not displayed.


81
Table 12. Predictors of Student Participation Rates in Concurrent Enrollment (CE)
Variable (Model 1) EHA Model (Model 2) Year Fixed Effects (FE) (Model 3) Year FE & District FEb (Model 4) Year FE & School FEC
College matriculation rates (%) 0.187*** 0 197* * 0.095** -0.017
(0.032) (0.032) (0.038) (0.037)
School performance rating (%) 0.064 0.029 0.037 -0.067**
(0.046) (0.048) (0.053) (0.030)
Median Household Income (logged) -0.474 -0.834 -1.149
(1.341) (1.342) (1.782)
FRL eligibility (%) 0.141*** 0.131*** 0.096** -0.019
(0.036) (0.037) (0.043) (0.050)
Student Count (logged) -2.840*** -2.950*** -1.173 -5.473*
(0.577) (0.576) (0.754) (3.263)
Diffusion of CE 0.155* -0.114 -0.224** -0.173**
(0.092) (0.117) (0.101) (0.081)
Community college distance -0.060 -0.058 -0.006
(0.062) (0.062) (0.159)
Community college concentration -0.624* -0.080 -0.625
(0.359) (0.376) (0.578)
Charter school -0.589 -0.444 0.716
(2.553) (2.550) (3.239)
PSEO participation 4.695*** 4 795* * 2.193
(1.024) (1.020) (1.355)
Outlying3 2.134 1.268
(1.519) (1.527)
Remote 0.999 0.305
(2.188) (2.191)
Urban-Suburban -3.025** -3.428***
(1.260) (1.258)
2011-12 2.268*** 2.417*** 2.700***
(0.589) (0.581) (0.547)
2012-13 5.013*** 5.284*** 5.654***
(0.774) (0.804) (0.810)
2013-14 5.384*** 5 799* * 6.317***
(0.920) (0.919) (0.885)
2014-15 6.078*** 6.452*** 7.257***
(0.953) (0.968) (0.893)
Constant 10.238 13.397 14.414 43.278**
(14.862) (14.945) (18.662) (18.330)
Observations 1,828 1,828 1,828 1,828
Adj. R-squared 0.219 0.239 0.538 0.753
*** p<0.01, ** p<0.05, p<0.10 a District Setting baseline group is Denver Metro Area b District fixed effects included but not reported; 178 categories absorbed c School fixed effects included but not reported; 383 categories absorbed Robust school-clustered standard errors in parentheses for all models


82
The fourth, and final, model includes school, rather than district, fixed effects, which are also absorbed (n=383). Including school fixed effects omits several indicators that do not vary over time (i.e., during the study) within high schools, including median household income, community college distance, community college concentration, charter school and PSEO participation. School fixed effects are important to include, however, since the treatment is at the school-level and characteristics specific to schools such as leadership, culture, academic system, and teacher capacity likely affect the implementation of concurrent enrollment.
None of the covariates remain statistically significant throughout all models. College matriculation rates and FRL eligibility are statistically significant in Models 1 through 3 but are not statistically significant at p<0.1 when school-level fixed effects are added (Model 4), as a result of the fixed effects controlling for unobserved variables and absorbing across-school and across-time variation. Regarding the college matriculation variable, while it is lagged one year, there is still a threat of endogeneity, if an increase in concurrent enrollment participation rates leads to an increase in matriculation rates.
The variable for the size of the high school is statistically significant at p<0.10 in Model 4.
The coefficient indicates that a 10 percent increase in the count of students in a high school leads to a 0.52 percent decrease in the participation rate, meaning that smaller high schools are more likely to have a larger share of their students participating in concurrent enrollment, although the effect size in this model is substantively small. To put it in more relative terms, an increase of 60 students at the average high school (going from the mean of 592 students to 652 students) is associated with approximately 3 fewer students participating in concurrent enrollment (decreasing from about 72 students participating to 69).
The charter school and PSEO covariates are not statistically significant in Model 3, which
includes district fixed effects; however, Model 4 is the preferred specification and both variables are


83
omitted in that model because there is no variance within a school on those indicators. Therefore, it is unclear from these fixed effects models what the effect of being a charter school or having had PSEO in place prior to 2009 is on participation rates. The diffusion of concurrent enrollment measure, which captures the number of neighboring high schools offering the program, is statistically significant in both the district and school fixed effects models (Model 3 and Model 4). The coefficient from Model 4 can be interpreted as an increase of 1 additional neighboring high school offering concurrent enrollment reduces the share of students participating by 0.17 percent. While the result is not in the predicted direction, the effect size is very small in substantive terms.
In the first two models, before district and school fixed effects were added, a variable for district setting was included. The coefficient for urban-suburban high schools was statistically significant at p<0.1, indicating those schools were associated with a 3.4 percentage point lower participation rate when compared to Denver metro area high schools. The descriptive statistics showed urban-suburban high schools were slower to adopt concurrent enrollment, and it appears that those schools also experience a lower overall participation rate once they do adopt concurrent enrollment. Lastly, the year fixed effects also reflect what was described in the descriptive statistics; average school-wide participation rates in concurrent enrollment have increased over time. The 2014-15 school year is associated with a 7.3 percentage point increase in the mean participation rate as compared to the 2010-11 school year.
Dynamic Panel Data Model
Table 13 displays the results from the dynamic panel data model, which includes a lagged value of the dependent variable and school and year fixed effects, and uses the maximum likelihood estimator to produce estimates. The model was estimated with the dependent variable lagged one year, which removes one year (2011) from the dataset. The coefficient for the lagged effect


84
indicates that a high school will see a 0.75 percent increase in this year's participation rate for every 1 percent increase in last year's participation rate.
Table 13. Dynamic Panel Data Model using Maximum Likelihood for Concurrent Enrollment (CE) Participation Rates in High Schools____________________________________
Variable Model 1
CE Participation Rate t-1 0.750*** (0.054)
Matriculation Rates (%) -0.026 (0.037)
School Performance Rating (%) 0.047 (1.171)
Median Household Income (logged) 4.633** (2.334)
FRL eligibility (%) -0.018 (0.035)
Student Count (logged) -7.060*** (1.528)
Diffusion of CE 0.052 (0.153)
Community college distance -0.324*** (0.099)
Community college concentration 0.346 (0.674)
Charter school -6.224*** (1.895)
PSEO participation 5.162*** (1.352)
Observations 388
Likelihood ratio 105.11
df 65
Prob>chi2 0.0012
*** p<0.01, ** p<0.05, p<0.10
Robust school-clustered standard errors in parentheses
As the results for the other variables in Table 13 show, even with the large effect size of the lagged dependent variable, other predictor variables remained statistically significant, including student count, which was statistically significant in the school fixed effects model. In the dynamic panel data model, however, the student count coefficient is much higher than in the fixed effects model; here, a one percent increase in student count is associated with a 7.1 percent decrease in
concurrent enrollment participation rates, all else held constant. Other differences between the


85
dynamic panel data model and the district fixed effects model are the non-significance of the school performance rating and the diffusion of concurrent enrollment variables.
An important advantage of the dynamic panel data model is that it allows for the effects of time invariant indicators to be observed. The following variables were omitted from the school fixed effects model and are statistically significant (p<0.01) in the dynamic panel data model: median household income, distance to the nearest community college, charter school, and PSEO. A one percent increase in median household income is associated with approximately a 4.6 percent increase in participation rates within a high school, accounting for the prior year's participation rate and holding all else constant. Adding an additional mile to the distance from the nearest community college results in a small decrease of 0.32 percent in the participation rate, while being a charter school is associated with a large decrease of 6.2 percent in the share of students in concurrent enrollment as compared to non-charter schools. High schools that participated in PSEO before concurrent enrollment see a 5.2 percent increase in participation rates as compared to those schools that did not have PSEO, all else equal.
Taken altogether, the results provide moderate support for Hypotheses 1.1 through 1.3, and less support for Hypotheses 1.4 and 1.5 (see Table 14). The following section provides further
discussion of the results and draws conclusions.


86
Table 14: Summary of Statistically Significant Results across Methods and Hypotheses
Event History Fixed Effects Dynamic Data Panel
Method Analysis Analysis Model
(Cox PH (OLS with School & (Lagged DV &
Model) Year FE) School/Year FE)
Dependent HS Adopted CE Participation CE Participation
Variable CE (Y/N) Rate Rate
Hypotheses High schools that have... Variable Results B(SE)
1.1: Lower academic College 1.015*** Non-significant Non-significant
achievement levels are matriculation (0.003)
more likely to adopt CE and have higher student participation rates. rates
School Performance Rating Non- significant -0.067** (0.030) Non-significant

1.2: Greater fiscal capacity Median Non- Omitted 4.633**
are more likely to adopt CE household significant (2.334)
and have higher student income
participation rates. % FRL 1.012*** Non-significant Non-significant
Eligibility (0.003)
1.3: More students have a Student 1.224*** -5.473* -7.060***
greater likelihood of adopting CE, but the share of students participating may be lower. count (0.082) (3.263) (1.528)
1.4: Other schools nearby Number of Non- -0.173** Non-significant
that have already adopted high schools significant (0.754)
CE are more likely to offer w/in 5 miles
the program and have offering CE
higher student participation rates.
5: Greater proximity to a community college will be more likely to adopt CE and have higher student participation rates. Distance from CC Non- significant Omitted -0.324*** (0.099)
# of community colleges w/in 10 miles Non- significant Omitted Non-significant
Charter 0.469*** Omitted -6.224***
Control variables (0.111) (1.895)
PSEO 1.494*** Omitted 5.162***
(0.164) (1.352)
*** p<0.01, ** p<0.05, p<0.10


87
Conclusion & Discussion
This study sought to understand influences on the adoption and utilization of concurrent enrollment among Colorado high schools using a panel data set spanning the first five years following the enactment of the Concurrent Enrollment Programs Act of 2009. This pursuit was particularly compelling given that recent research has shown a positive association between concurrent enrollment participation and college outcomes, and given that Colorado's program was nearly fully diffused within five years. The goal was to use event history and regression analysis to see if there were any best practices that could be gleaned from the Colorado case study and applied to other states trying to scale up similar programs. There were several important findings that resulted from this research, although questions remain.
Academic achievement levels
The first hypothesis posited that schools with lower academic achievement levels would be more motivated to adopt concurrent enrollment, which is seen as a strategy for boosting achievement and postsecondary readiness. The variable for school performance rating, which encompasses information on student academic achievement and growth, is nonsignificant across models with the exception of the school fixed effects model, in which the effect size is very small. The college matriculation rate variable is nonsignificant in the participation rate models but is statistically significant in the event history analysis model. The direction of the coefficient is not what was hypothesized, but, as noted previously, the finding is not surprisinghigh schools with already-established cultures of college readiness are likely to take advantage of an additional college access program. Given the eventual, widespread diffusion of the program, it is evident that even high schools with historically low college-going rates have implemented concurrent enrollment as
well.


88
Further, as Table 14 summarizes, schools with higher percentages of FRL students were also more likely to adopt concurrent enrollment. While that variable is technically included as part of the school resources hypothesis, there is a well-known inverse relationship between income status and college-going rates. Thus, the two findings are seemingly contradictory at first glance. Upon reflecting on the results further, though, it is understandable that both findings occurred simultaneously. It seems natural that there would be a higher propensity for the adoption of college preparatory programs in schools where a majority of students matriculate to college. Such schools already have a college-going culture established, and likely have parents who push for the inclusion of opportunities that would advance their child's education. On the other hand, Colorado's Concurrent Enrollment Programs Act was specifically passed to expand access to those students who typically had not been includedthat is, low-income and minority students. Given the rapid diffusion of the program, and the positive, statistically significant coefficient on the FRL variable, one could conclude that high schools have taken up the opportunity to expand access to new groups of students as the law intended. Future research could seek to further untangle these effects, and also explore to what degree those schools that had initially low college-going rates and high FRL rates are seeing positive gains in their postsecondary outcomes as a result of offering concurrent enrollment. Fiscal Capacity
From the results for Hypothesis 1.2, there is some evidence that fiscal capacity relates to concurrent enrollment participation. The EHA results show that schools with higher proportions of low-income students were quicker to adopt concurrent enrollment, but the median household income was not statistically significant. When considering effects on participation rates, the opposite case is seen. The FRL variable is not statistically significant, but the median household income variable is statistically significant and substantively large in the lagged dependent variable
model. The results of that model suggest that high schools in wealthier communities have higher


89
levels of participation in concurrent enrollment. These somewhat conflicting results mirror the previous discussion. While schools that have lower-income populations were quick to adopt the program, perhaps motivated by the intent of the program to expand access, it appears that those schools with more advantaged populations (likely with strong college going cultures and higher numbers of eligible students) have higher shares of students actually participating in courses.
Size of High School
School size was a statistically significant variable in both the event history and the regression analyses. Larger high schools were quicker to adopt concurrent enrollment, leading one to conclude that organizational resources matter in the initial adoption of a program, as hypothesized. The hypothesis regarding school size also posited that the number of a students in a school may be inversely related to the share of students participating in concurrent enrollment. The results confirmed that supposition, with the regression models indicating that smaller high schools were more likely to have high proportions of students participating in concurrent enrollment than large high schools. That is not to say there are not examples of large high schools that have high percentages of students in concurrent enrollment, but rather, on average and holding all else constant, it is more likely for smaller schools to more intensely use the program. As discussed earlier, this can be due to the fact that when a small high school offers a concurrent enrollment course, for example a math class for seniors, that class comprises a large percentage of its overall enrollment, whereas offering one class at a large high school will only consist of a small share of the overall student body.
There is also the consideration that for smaller schools, many located in remote areas, concurrent enrollment provides access to content the school is not able to deliver on its own. A high school, for instance, can go through a community college and use its instructors to offer college-
level chemistry, whereas without concurrent enrollment that school would not be able to offer the


Full Text

PAGE 1

UNDERSTANDING THE EXPANSION AND EFFECTS OF b y BRENDA BAUTSCH DICKHONER B.A., Duke University, 2006 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillm ent of the requirements for the degree of Doctor of Philosophy Public Affairs Program 2017

PAGE 2

ii This thesis for the Doctor of Philosophy degree by Brenda Bautsch Dickhoner has been approved for the Public Affairs Program by Todd Ely, Chair Paul Teske K elly Hupf eld Matt Gianneschi Date: May 13, 2017

PAGE 3

iii Dickhoner, Brenda Bautsch (PH.D., Public Affairs Program ) Understanding the Expansion and Effects of rogram Thesis directed by Assistant Professor Todd Ely ABSTRACT One o f the prominent approaches among states to improve college access and success is concurrent enrollment, which provides high school students the opportunity to enroll in a college course for which they may receive both high school and college credit. This s tudy set out to understand, first, what factors le a d some schools to adopt concurrent enrollment more quickly and implement the program more intensely as compared to other schools The study also sought to evaluate how effective concurrent enrollment is at improving college access and success for all students, including low income and minority students. The dissertation finds that fiscal capacity, organizational capacity, school type and prior program offerings are key predictors of the adoption and impleme nta tion of c oncurrent enrollment programs. Additionally, p articipation in concurrent enrollment in high school results in positive gains in college enrollment rates, first year grade point averages, and college persistence rates, and results in a decrease in the need for remedial education. While concurrent enrollment, on average, improves college outcomes for all students, low income students experience a greater positive impact on their outcomes than higher income students. Moreover, Hispanic students who take concurrent enrollment courses see a greater impact on their likelihood of going to college than white students who participate in the program. The form and content of this abstract are approved. I recommend its publication. Approved: Todd Ely

PAGE 4

iv DEDICA TION For Blair, whose love and support means e ver ything And for Grayson I hope you a lways pursue your dreams no matter how l ong the road ahead seems.

PAGE 5

v ACKNOWLEDGEMENTS I am extraordinarily indebted to Dr. Todd Ely, who provided a dvice guidance and mentorsh ip over the past six years. Dr. Ely has a n enviable aptitude for statistics and challenged me to explore various quantitative methods in an effort to carry out a rigorous and respectable research design I learned more than I could have imagined t hanks to the patient facilitation of Dr. Ely. He even made the process enjoyable as much as suc h a process can be enjoyed D rs. Paul Teske and Kelly Hupfeld lent their public affairs and education policy expertise to provide valuable feedback, particularly in the beginning stages as I was preparing what would be the roadmap for my research. I a m grateful that Dr. Teske Tanya Heikkla, Chris Weible, and Peter deLeon along with many others have created such a wonderful and welcoming PhD program for prac titioner s. T he School of Public Affairs facult y encourages the blending of theory with practical appli cation and warmly a ccepts practitioner students such as myself into their scholarly sphere. I am incredibly grateful that Dr. Matt Gianneschi served on my committee as my outside reader. Dr. Gianneschi program and has a wealth of knowledge about education policy through his roles i n state government in the policy sector and as a college leader. Dr. Gianneschi was also one of the i ndividuals who helped me land on the topic of concurrent enrollment; without him and Dr. Beth Bean I might still be wandering the doctoral wilderness in search of worthy topic. I a m apprec iat ive of Dr. Bean for not only helping me find a topic and a rich d ata set, but also for providing moral support as I worked for her at the Colorado Department of Higher Education (CDHE) Maggie Yang, Michael Vente and all of the CDHE staff were tremendously helpful and patient with my multiple data req uests. while I was at C DHE, s hared an abundance of knowledge with me to help inform the background,

PAGE 6

vi context and discussion portions of this dissertation. Michelle also happens to be a dear friend, and I am so grateful for her support and friendship in addition to the concurrent enrollment insight. I o we a great deal of gratitude to Dr. Julie Bell at the National C onference of State Legislatures, who was my boss when I had the crazy idea to enter a Ph D program. Dr. Bell encour aged me to apply and supported m y acceptance into the program with her letter of recommendation. She also permitted me to work a flexible schedule as I completed my coursework. Thank you, Dr. Bell, for your belief in me I truly woul d not be at t his milestone without you. Alyssa Pearson at the Colorado Department of Education has been an amazing friend and supervisor t his past year as I have completed and defended my dissertation. Thank you, Alyssa, for your unwavering support, overflowing optimism, delicious baked goods, generosity of spirit and for being an inspiration to all You are an excellent role model for public administrators everywhere. Last although certainly not least, I want to acknowledge my family and friends who have supported me on this long road. My parents instilled in me a love for education at a young age. My dad made it possible for me to attend the college of my dreams and my mom made it possible for me to persist through gradua te school. She helps tak e care of Grayson and moe, brings me food any My mom once said she would never be satisfied until I earned my doctorate so I am pleased to finally meet her lofty expectations. Thank you, Mom and Dad for all you do and for valuing education so much! My husband, Blair has been my rock and is the reason I ve made it to the finish line While this process has be en long and grueling at times, one positive outcome is that Blair was able to pursue multiple hobbies while I worked, including guitar playing, marathon running and beekeeping. I will be the proud recipient of a doctoral diploma and homemade honey! Finally I am lucky enough to have many friends too many to name who grabbed a drink with me when I needed one and understood when I had too much work to g o get a drink. Thank you, all

PAGE 7

vii TABLE OF CONTENTS I : IN TRODUCTION ................................ ................................ ................................ ................................ ... 1 Problem Significance ................................ ................................ ................................ .......................... 3 Concurrent Enrollment Policy Landscape ................................ ................................ .......................... 6 ................................ ................................ ............ 14 Contributions to the Field ................................ ................................ ................................ ................ 17 Summ ary & Research Questions ................................ ................................ ................................ ...... 22 II : LITERATURE REVIEW AND H YPOTHESES ................................ ................................ .......................... 24 Policy Diffusion & Innovation Theory ................................ ................................ ............................... 24 Education Theory ................................ ................................ ................................ .............................. 30 Summary ................................ ................................ ................................ ................................ ........... 40 III : DATA & METHODS ................................ ................................ ................................ ........................... 41 Data Sources and Collection ................................ ................................ ................................ ............. 41 Research Design ................................ ................................ ................................ ............................... 44 Data & Methods: Summary ................................ ................................ ................................ .............. 69 IV : POLICY DIFFUSION FINDINGS & DISCUSSION ................................ ................................ .................. 71 Descriptiv e Statistics ................................ ................................ ................................ ......................... 71 Event History Analysis ................................ ................................ ................................ ...................... 77 OLS Fixed Effects Regression Analysis ................................ ................................ .............................. 80 Dynamic Panel Data Model ................................ ................................ ................................ .............. 83 Conclusion & Discussion ................................ ................................ ................................ ................... 87 V : POLICY EVALUATION FINDINGS ................................ ................................ ................................ ....... 93 Descriptive Statistics ................................ ................................ ................................ ......................... 93

PAGE 8

viii Effects of Concurrent Enrollment Participation on College Outcomes ................................ ............ 95 Concurrent Enrollment Effects for Low Income Students and Minority Students ........................ 104 Effects of Concurrent Enrollment Credit Hour Levels on College Outcomes ................................ 112 Conclusion & Discussion ................................ ................................ ................................ ................. 117 VI : CONCLUSION ................................ ................................ ................................ ................................ 120 Key Findings ................................ ................................ ................................ ................................ .... 122 Implications for Research and Practice ................................ ................................ .......................... 123 Limitations and Future Research ................................ ................................ ................................ .... 135 REFERENCES ................................ ................................ ................................ ................................ ....... 141

PAGE 9

ix L IST OF TABLES Table 1. Thematic Analysis of Model State Policy Elements and Standards ................................ ........ 10 Table 2. Summary of Research Questions and Hypotheses ................................ ................................ 40 Table 3. Concept Measurement Summary: Policy Diffusion ................................ ................................ 45 Table 4: Variable Descriptions and Sources ................................ ................................ ......................... 46 Table 5: Concurrent Enrollment Adoptions and Survivor Functions, by School Year .......................... 52 Table 6. Concept Measurement Summary: Policy Evaluation ................................ ............................. 58 Table 7. Descriptions of Pre College Independent Variables and College Outcome Variables ........... 59 Table 8: Methodological Approaches with Associated Research Questions ................................ ....... 69 Table 9: Descriptive Statistics for All High Schools, Beginning and End of Study ................................ 73 Table 10: Comparison of Variable Means, by High School Adoption Year ................................ ........... 74 Table 11: Cox Proportional Hazards Model Results ................................ ................................ ............. 78 Table 12. Predictors of Student Participation Rates in Concurrent Enrollme nt (CE) ........................... 81 Table 13. Dynamic Panel Data Model using Maximum Likelihood for Concurrent Enrollment (CE) Participation Rates in High Schools ................................ ................................ ................................ ...... 84 Table 14: Summary of Statistically Significant Results across Methods and Hypotheses .................... 86 Table 15: Descriptive Statistics for Overall Sample and by Concurrent Enrollment (CE) Particip ation 94 Table 16. Propensity Score Matching Average Treatment Effects ................................ ....................... 98 Table 17. Progression of Logistic Regression Models Esti mating the Effect of ................................ .. 101 Table 18 Average Treatment Effects ................................ ................................ ................................ 102 Table 19. Comparison of Average Treatment Effects ................................ ................................ ........ 104 Table 20. Progression of Logistic Regression Models Estimating the Effect of Concurrent Enrollment Participation on College Matriculation ................................ ................................ .............................. 106

PAGE 10

x Table 21. Regression Models Estimating the Interaction Effects of Concurrent Enrollment Participation on College Outcomes ................................ ................................ ................................ .... 108 Table 22. Credit Hours Descriptive Statistics for Concurrent Enrol lment Students ........................... 113 Table 23. Sample Means of Key College Outcomes by Concurrent Enrollment Credit Hours ........... 113 Table 24. Progression of Regression Models Estimating the Effect of Concurrent Enrollment Participation on College Matriculation ................................ ................................ .............................. 114 Table 25. Average Treatment Effects of Credit Hours Levels on College Outcomes ......................... 115 Table 26. Comparison of Statewide Evaluations Assessing Effect of Dual Enrollment Programs on College Matriculation ................................ ................................ ................................ ......................... 133

PAGE 11

xi LIST OF FIGURES Figure 1. Number of Adopted Bills Pertaining to Dual Enrollment Programs across the U.S., by Year 9 Figure 2. Distribution of Propensity Score Across Treatm ent and Comparison Groups ...................... 66 Figure 3. Adoption of Concurrent Enrollment Programs from the 2010 11 School Year to the 2014 15 School Year, by School Districts and High Schools. ................................ ................................ .............. 72 F igure 4: Average Percentage of High School Students Participating in Concurrent Enrollment (CE) within High Schools, by Adoption Year Cohort from 2010 11 to 2014 15. ................................ ......... 75 Figure 5. Maps of Colorado high schools and Concurrent Enrollment (CE) participation rates by covariates of interest.. ................................ ................................ ................................ .......................... 76 Figure 6. Cox Proportional Hazards Regressi on Smoothed Hazard Functions for Charter Schools a nd College Matriculation Rates. ................................ ................................ ................................ ............... 80 Figure 7. Participation in Concurrent Enrollment, by Graduation Year, Gender and Race/Ethnicity .. 95 Figure 8. Standardized bias differences (%) across all covariates in original and matched samples ... 97 Figure 9. Probability of College Matr iculation, by Concurrent Enrollment Participation and Free or Reduced Price Lunch (FRL) Status and Race/Ethnicity (Hispanic or white) ................................ ....... 109 Figure 10. Probability of College Remediation, by Concu rrent Enrollment Participation and Free or Reduced Price Lunch (FRL) Status ................................ ................................ ................................ ...... 110 Figure 11. Probability of College Persistence, by Concurrent Enrollment Participation and Free or Reduced Price L unch (FRL) Status ................................ ................................ ................................ ...... 112

PAGE 12

1 CHAPTER I INTRODUCTION and earn family sustaining wages (Carnevale, Smith & Strohl, 2013). Access to a high quality K 12 education that prepares students for postsecondary education, however, is not a guarantee in income and minority students consistently have lower levels of academic achievement than their peers at all p oints along the education pipeline, inclu ding high school graduation, college enrollment, and college degree attainment (Bettinger & Long, 2005; Darling Hammond, 2010; Kahlenberg, 2004; Terenzini, Cabrera, & Bernal, 2001; U.S. Department of Education [USDO E] 2006). States across the country have implemented countless policies to better prepare students for life after high school, but achievement gaps persist. Colorado, which has the second largest gap in the country in the college degree attainment betwee n majority and minority students ( NCHEMS, 2013 ), is no exception. Several laws passed by the Colorado legislature in the last decade have targeted improving the transition from high school to college. 1 The question remains, though, as to how effective are those laws at improving educational outcomes, particularly when policies create voluntary programs for schools and students law for example, was specifically designed to improve college readiness for traditionally underse rved students by bolstering access to rigorous, college level coursework (C.R.S. §22 35 101). Under the policy, qualified students in grades 9 through 12 can take tuition free college courses at their high school, a postsecondary institution, online, or in a hybrid format and simultaneously earn high school and college credits (CDE, 2010). Th is law creates 1 See, for example SB08 212: Preschool to Postsecondary Education Alignment Act ( Colorado Achievement Plan for Kids ) ; SB 09 256: Individual Career and Academic Plans; HB09 1319: Concurrent Enrollment Programs Act; HB07 1118: High School Graduation Requirem ents; SB09 163: The Education Accountability Act; HB 12 1155: Supplemental Academic Instruction.

PAGE 13

2 the operational framework the funding mechanism, participation requirements, and oversight for the concurrent enrollment program I t i s a voluntary initi ative however schools can choose whether or not to adopt the program. Proponents of concurrent enrollment argue that it increases academic preparation for college and provides momentum toward degree attainment by giving students the opportunity to enter college with credits already accumulated (An, 2013; Hoffman 2005 ). Prior research has found positive associations between concurrent enrollment participation and college access and success outcomes (Allen & Dadgar, 2012; An, 2013; Giani et al., 2014; Tayl or, 2015 ). Often the previous research has focused on small scale, institution specific programs and used imperfect methods Consequently, rigorous, empirical analys e s of state wide programs are still needed (Allen & Dadgar, 2012; Bailey & Karp, 2003; Blan co, 2006; Giani, Alexander, & Reyes, 2014; Hoffman, 2012; Rutschow research T he purpose of this study is to examine nt program on college access and success, as well as to analyze decisions by high schools to offer concurrent enrollment programs and by students to enroll in them T o address these questions, t he dissertation begins with an introduction to the problems under investigation and background on how concurrent enrollment state policies purport to solve those problems, both nationwide and in Colorado This introductory chapter concludes with a summary of contributions the study will make to research and practic e and sets forth formal research questions. Chapter Two provides a review of relevant literature from the public affairs and education domains and presents testable hypotheses. Chapter Three includes a description of the data collection, an explanation of variables and measures, and detailed review of the various methods employed to answer the research questions. Chapters Four and Five present findings from the empirical research, with C hapter Four focusing on a n analysis of factors that influence the

PAGE 14

3 adopt ion of concurrent enrollment programs at the school level; C hapter Five focus es on a n analysis of the effects of participating in concurrent enrollment on college matriculation and success at the student level Chapter 6 summarizes the study and its implic ations for research and practice Problem Significance Achievement G aps Low income and minority students, on average, lag behind their peers on nearly every important education milestone (An, 2012; Bettinger & Long, 2005; Darling Hammond, 2010; Kahlenberg 2004; Oakes, 2005). Children from low income families, for example, are more likely to have low er reading abilities by the third grade than high income students (Hernandez, 2011). Achievement data from the National Assessment of Educational Progress (NAE P) shows that black and Hispanic students, on average, score two grade levels below white students when taking the NAEP exam in 4 th and 8 th grades ( USDOE 2009, 2011). Low literacy levels in early grades ha ve been linked to diminished achievement in later years, including decreased high school graduation rates (Hernandez, 2011). Early indicators are important to measure because low income students are about five times as likely to drop out of high school as high income students (Kahlenberg, 2004). In high school, disparities in curriculum offerings and quality of instruction remain a significant problem with low income and minority students disproportionally receiving lower quality instruction and fewer advanced course options. Oakes (1993, 2005) found tha t even after controlling for test scores, white and Asian students are far more likely to be placed into h onors courses than their peers. H igh achieving Latino students who scored at the 90th percentile on standardized tests had just a 56 percent chance of being assigned to a college prepara tory class, as compared to a 97 percent chance for Asian students and 93 percent chance for white students scoring in the same percentile (Oakes, 1993 2005 ).

PAGE 15

4 Due to a variety of factors, including lack of access to con sistently high quality instruction and rigorous curriculum, achievement gaps that can be observed as early as pre Kindergarten persist for many children throughout t heir entire education al careers The transition from high school to college is no exception l ow income and minority students are less likely to enroll in or graduate from college than their white, affluent peers (Adelman, 2006; An, 2012; Kahlenberg, 2004). For low income, minority students who do attend college, they tend to be less academic ally prepared than their peers ; s t udies on the relationship between income and race/ethnicity and college remediation rates indicate persistent achievement gaps ( see, e.g., Bettinger & Long, 2005). In Colorado the focus of this study 8 2 percent of Af rican Amer ican students and 70 percent of Hispanic students need remediation at communi ty colleges, as compared to 50 percent of w hite students (Colorado Department of Higher Education, 201 6). Also in Colorado, 53.4 percent of low income students are not ready for college level courses in at least one content area as compared to 31.4 percent of wealthier students ( Colorado Department of Higher Education, 201 6) The fact that half of all white high school graduates who immediately attend a community college are not academically prepared is indicative of systemic challenges in readying our young adults for postsecondary education. That statistic already discounts the numerous students who dropped out of high school or those who graduated high school but chose not to matriculate to college. Further, while the remedial education rate for white students is concerning in and of itself, having remedial education rates that are 20 to 30 percentage points higher for minority students is an alarming trend Returns to Educati on C losing achievement gaps, particularly around college access and success, remains a significant imperative for society from an equity perspective, as well as from an economic perspective. I f achievement gaps persist, then the U.S. society and economy wi ll continue to

PAGE 16

5 experience negative externalities stemming from lower individual quality of life. Research has time and again found that individuals without a college credential are far more likely to face severe challenges throughout life including jobless ness, welfare, incarceration, family instability and health problems (Hout, 2012; Kingston et al., 2003). Th e se challenges are costly and burdensome to the taxpayers who subsidize prisons, social support systems and healthcare. Researchers, however, have a lso long questioned the notion of whether education causes better outcomes or simply reflects advantages bestowed upon certain individuals as a matter of chance. Nonetheless, there is substantial empirical evidence that education provides positive returns on investment for individuals. The literature, for example, on wage premium s for attending college has consistently found that individuals accrue increased earnings for additional years of education using a variety of statistical approaches to control for selection bias, including instrumental variables and natural experiments ( Angrist & Krueger, 1992; Hausman & Taylor, 1981; Hout, 2012, Kane & Rouse, 1995). More recent research has also found that the benefits of higher education are greater for those who are less likely to attend and graduate that is, students who typically perform somewhere in the middle of the spectrum of academic ability (Attewell & Lavin, 2007; Brand & Xie, 2010; Hout, 2012; Maurin & McNally, 2008) While students of higher ability ma y graduate from college at higher rates and earn higher wages, their education has a lesser effect on their success than students of lower academic ability who gain greater wage premiums from higher education (Brand & Xie, 2010; Hout, 2012). This strand of literature has important implications for policymakers in that it supports continued efforts by states to expand higher education access to students who are at risk of not attending. There is also empirical evidence that societal and economic benefits acc rue when higher education completion rates increase. Some studies have found that increasing the number of college graduates in a labor market raises the productivity levels of less educated workers and may

PAGE 17

6 also increase their wages (Moretti, 2012; Mas & M oretti, 2009). Researchers have also linked college graduates with higher rates of volunteerism and positive views of civil liberties and minorities (e.g. Brand, 2010; Kingston et al., 2003). Putnam (1995 672 education is by far the strongest correlate that I have discovered of civic engagement in all its forms." From the economic perspective, labor economists project that jobs in particular, those that provide family sustaining wages will increasingly require postsecondary credentials The labor demand for college educated workers is projected to surpass supply by 2020, which could stymie economic growth ( Carnevale, Smith & Strohl, 2013) Center on Education and the Workforce: Esse ntially, postsecondary education or training has become the threshold requirement for access to middle class status and earnings in good times and in bad. It is no longer the preferred pathway to middle class jobs it is, increasingly, the only pathway. ( Ca rnevale, Smith & Strohl, 2013 13) As this short review indicates, there is a compelling case for expanding higher education opportunit ies to more students. Policymakers often understand this and thus have turned their attention in recent years to expandi ng college access through c oncurrent enrollment. The following section provides an overview of the national policy landscape surrounding concurrent enrollment. Concurrent Enrollment Polic y Landscape Concurrent enrollment is a term used in 15 states, includ ing Colorado, to refer to opportunities for high school students to enroll in a college course for which they may receive both high school and college credit. Unlike other accelerated learning options such as Advanced Placement (AP), students earn college credit if they receive a passing grade in the course just as a college student would rather than by earning a certain score on an end of course exam (Allen, 2010).

PAGE 18

7 co llege degree. Forty states use the terms dual enrollment or dual credit to refer to the same arrangement ; 2 the terms are used interchangeably in the following section Concurrent enrollment programs have been available in public high schools for at lea st the last half century, mostly as an enrichment opportunity for academically advanced students. Programs have grown exponentially since the early 2000s when certain policymakers began expanding concurrent enrollment opportunities to students who are trad itionally underserved, including students o f color and low income students, as well as to students who a re not high academic performers ( Hoffman, Vargas, & Santos, 2008a ). In the 2001 02 school year, public high schools across the country reported approxi mately 1.2 million enrollments in dual credit courses (Kleiner & Lewis, 2005). That number is a duplicated student count it is inclusive of each course enrollment during the school year. A decade later, during 2010 11 school year, dual enrollment participa tion at public high schools increased to just over 2 million (Thomas, Marken, Gray & Lewis, 2013). In 2001 02, 71 percent of high schools had dual enrollment programs; by 2010 11 that figure increased to 82 percent. C oncurrent Enrollment Promises and Ch allenges Concurrent enrollment is promising to policymakers and practitioners because it is seen as a way to expose more students to rigorous curriculum that high schools may be lacking. Providing students exposure to college is thought to be a strategy f or developing metacognitive skills 3 readying students for the demands of college life and increasing college aspirations Policymakers are also drawn to concurrent enrollment as a way to increase college affordability by offering college courses at low o r no cost to families. 2 In some instances, multiple terms are used within states. 3 Students with well developed metacognitive learning skills will be able to manage their time effe ctively, think critically, navigate college resources, maintain study routines, have self awareness of their strengths and weaknesses, analyze and interpret information, and have the confidence to overcome challenges (Conley, 2010, 2013).

PAGE 19

8 The challenges to policy implementation, however, are also multifold. While state policies around concurrent enrollment have proliferated, expanding access to low income students and students of color still remains a challenge F urther, while many states are attempting to increase access by ensuring there are no costs to students state budgets are continually under constrain t, leaving little dedicated funding available for concurrent enrollment. Even in state s where students do n ot shoulder tuition costs, school districts and colleges still need to establish a financially viable model for operating the program. Cash strapped states, districts and colleges increasingly have to find creative ways to fund concurrent enrollment progra ms or risk scaling back access (Borden et al, 2013; Zinth, 2014b, 2015b) Another challenge is ensuring course rigor and quality when courses are taught at a high school or online, as opposed to on the college campus. Offering courses in a high school set ting greatly expands access and eases the logistical hurdles of transportation and scheduling for off campus courses, but it requires more oversight to ensure consistency of rigor (Borden et al, 2013; Lowe, 2010; Zinth, 2015a). It is also challenging to fi nd high school teachers with the necessary qualifications to teach concurrent enrollment courses, especially in rural areas (Zinth, 2014 a ). These challenges and promises have spurred a great deal of legislative activity in recent years. State Policy Accord ing to the Education Commission of the States (ECS), as of 2016, 47 states have statutes and/or regulations in place governing dua l enrollment programs (ECS, 2017 ). However, a great deal of variation among the 47 state policies exists regarding funding, el igibility, course type, instructor qualifications, general oversight and monitoring, and credit transferability (Borden et al., 2013). Further, state policies continue to evolve as states make modifications to their programs in th e se areas. According to da ta collected by ECS over the past five years alone, 143 bills were adopted by state legislatures concerning dual enrollment programs; in the last ten years states

PAGE 20

9 passed a total of 243 bills (ECS, 2017) Figure 1 displays the number of bills signed into law by states a cross the count r y per calendar year. The chart shows low points (2008 and 2010) and high points (2013), but depicts that the number of adopted bills has remained near the average of 24 bills in most years over the past decade. Legislative changes have focu sed on clarifying or expanding funding streams, integrating career and technical education opportunities, promoting options that increase the number of qualified instructors, modifying student eligibility requirements, and implementing provisions to help e nsure dual credit courses are as rigorous as traditional college courses. Figure 1 Number of Adopted Bills Pertaining to Dual Enrollment Programs in the U.S., 2007 2016 Data collected from the Edu cation Commission of the States (ECS) State Po licy Database retrieved February, 11, 201 7 Model P olicy E lements With the high level of legislative a ctivity around dual enrollment, p olicy researchers have delved into the numerous state policies and based on other research and best practices, have identified key components that states should include in their dual enrollment policies T his section revie ws three prominent sets of model policy elements and program stan dards, which are synthesized in Table 1 ECS and Jobs for the Future (JFF) have issued specific guidance for policymakers T he National Alliance of Concurrent Enrollment Partnerships (NACEP) issued guidance focused o n program oversight and ensuring academic rigor. 22 13 30 10 25 26 38 25 26 28 0 10 20 30 40 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 Number of Bills

PAGE 21

10 Table 1 Thematic Analysis of Model State Policy Elements an d S tandards Themes Jobs for the Future NACEP Education Commission of the States Level Program Quality: Course rigor, instructor qualifications and course credi t Quality Assurance States should ensure that college courses offered to high school students use the same syllabi and exams as comparable cou rses taught on a college campus The postsecondary institution conferring credit should set the qualifications for faculty teaching dual credit courses D ual enrollees earn both high school and college credits upon s uccessfully completing courses Curriculum College courses offered in the high school are of the same quality and rigor as the courses offered on campus at the college/ university Faculty Concurrent enrollment i nstructors meet the academic requirements for faculty and instructors teaching in the sponsoring postsecondary institution and are provided discipline specific professional development Assessment Stude nts enrolled in concurrent enrollment courses are held to the same standards of achievement as students in on campus courses including grading standards and assessment methods Ensuring Course Quality Courses meet same level of rigor as traditional college courses Instructors meets same expectations as college faculty and receive support Transferability of Credit Postsecondary Institutions accept dual enrollment credit as transfer credit, provided measures for quality are ensured Financial Provisions $$$ Sustainable Funding and Finance States should develop funding policies that: A llow high school students to take college courses free of tuition and non course related charges Permit b oth districts a nd postsecondary institutions to claim per pupil funding allocations to support the cost of offering dual credit courses Finance Responsibility for tuition does not fal l to parents Districts and postsecondary institutions are fully funded or reimbursed for participating students

PAGE 22

11 Table 1 (cont.) Themes Jobs f or the Future NACEP Education Commission of the States Level Student Access and Support Eligibility and Access A state's eligibility requirements are determined b y the secondary and postsecondary sectors together S tudents have multiple ways to demonstrate readiness, including a combination of tests, end of course grades, teacher recommendations, and work portfolios. Academic and Social Supports States should requir e that districts colleges specify/ document key roles and responsibilities in memoranda of understanding or cooperative agreements, including the provision of a college liaison for student advisement and support States should provide support and funding for programs serving students who are overage and under credited and youth who have dropped out of high school Students Students officially register with a college Students meet pre requis ites The concurrent enrollment program provides st udents with a handbook of rights/responsibility of college students Access All eligible students may participate, based on d emonstration of ability to access college level content Caps on the maximum number of courses allowed should not be overly restricti ve Students earn high school and college credit for successful completion of approved postsecondary courses All students and parents are annually provided wish program information Counseling is made available to students and parents before and during progr am participation Reporting and Program Evaluation System for Accountability States should report annually on dual enrollment participation and impact and develop administrative structures to support program leaders and dual enrollment partnerships. State s should also designate a state board or governing body as having the authority and responsibility to guide dual enrollment policy. Aligned Data Systems States should develop unit record statewide data systems that identify dual enrollees by demographic ch aracteristics and monitor student progress longitudinally across the K 12 and higher education systems Program evaluation Concurrent enrollment programs display greater accountability through required impact studies, student surveys, and course and program evaluations Ensuring Course Quality (cont.) Districts and institutions publicly report on student participation and outcomes Programs undergo evaluation based on available data Sources: NACEP, 2011 ; Ward & Vargas, 2012 ; Zinth, 2014b

PAGE 23

12 NACEP quality concurrent enrollment partnerships and hold them accountable to high stand 2017). The organization administers the only national set of quality standards for concurrent enrollment programs, which it uses to accredit individual postsecondary institutions that offer concurrent enrollment programs across the country. N ACEP advocates for states to use the standards as a quality measure in statewide concurrent enrollment programs. There are currently 17 states that have modeled their quality standards (as set in statue or regulation) on the NACEP standards including Colo rado The standards are categorized around curriculum, faculty, students, assessment and program evaluation and are geared toward s ensuring that courses taught by high school teachers, in particular, are as rigorous and high quality as courses taught by po stsecondary faculty on college campuses (NACEP, 2011) ECS identifies 13 policy components organized under the categories of access, finance, ensuring course quality and transferability of credit (Zinth, 2014 b ) The guidance to policymakers notes that the policy components were selected because they more diverse gro u p of students successfully participates in high quality dual enrollment courses and receives credit that will be transferable to other public postsecondary i n (Zinth, 2014 b 4). Jobs for the Future undertook their policy scan with a lens similar to that used by ECS, but focused more on the key policy components needed to close achievement gap s The organization posits that state po licies have the potential to facilitate meaningful partnerships between high schools and colleges that result in a seamless transition into higher education for students who might not otherwise attend. Of the 47 statewide policies JFF reviewed, however, th ey found that have established sufficient mechanisms to ensure that all students, including those

PAGE 24

13 Vargas, 2012, 4). The six mechanisms JFF identified as important are categorized under quality assurance, eligibility and access, academic and social supports, systems for accountability, aligned data systems and sustainable f unding a nd f inance After analyzing the different policy elements and standards a mong ECS, JFF and NACEP four themes were identified that pro vide a coherent grouping of the elements : 1) program quality, 2) student access and support, 3 ) reporting and program evaluation, and 4 ) financial provisions. Given ty, its standards are concentrated under that first theme, but they do also address student access and program evaluation. The ECS and JFF model policy elements include more guidelines around program evaluation and financial provisions. In terms of progra m quality, some states require or encourage their postsecondary institutions to seek NACEP accreditation as a way to ensure concurrent enrollment courses are rigorous Other states defer to local control and leave it to individual colleges to monitor concu rrent enrollment course quality. There are two model components regarding financial provisions that are recommended by JFF and ECS. The first concerns keeping costs away from students and families so that the program is open and affordable to all. The sec ond model policy element focuses on keep ing costs low for districts and colleges. There are a variety of f unding approaches across states; JFF and ECS recommend that states cover the full costs of concurrent enrollment, or, at a minimum, allow both K 12 an d higher education systems to collect per pupil funding for student enrollments to offset costs The latter method is positively as it ensures both systems have the means and incentive to participate ( Hoffman, 2005; Lerner & Brand, 2006; Ward & Vargas, 2012; Zinth, 2014b ). An additional theme identified concern s ensuring students have adequate support throughout the process i ncluding before, during and after the concurrent enrollment c ourse takes

PAGE 25

14 place and equitable access to the program JFF advocates for more open access wher ein students may demonstrate readiness for college level coursework through portfolios, end of course grades, and teacher recommendations. In many cases, readines s is demonstrated through a course placement assessment. Lastly, a theme across all three organizations was the importance of tracking student data, reporting outcomes, and evaluating the effectiveness of the program in meeting its intended goals. With the legislation is explored in more detail in the next section. ograms Act, there were dual enrollment opportunities available to Colorado high school students, but there was no s tate level coordination of th e p rogram s which resulted in little accountability or attention to quality and low participation rates, particularly for low income an d minority students (CDE, 2010; CR 102(d)). In 2007, Governor Ritter convened a P 20 Education Coordinating Council to develop policies that would foster a seamless education system in which all students receive a high quality education from pre school through graduate school and e nter the workforce prepared to meet the 20 educated states due to imported talent 12 students are not persisting to and through college at high rates (Lopez, 2011; NCHEMS, 2013). Postsecondary access and success was a foca n 2009, at the recommendation of the coun cil, legislative leaders introduced the bipartisan Concurrent Enrollment Programs Act (H ouse B ill S enate B ill chambers of the legislature a rare feat.

PAGE 26

15 Policy Goal s The concurrent enrollment p rogram was specifically created to reach traditionally underserved populations. As the legislative declaration of the Concurrent Enrollment Programs Act states: Historically, the beneficiaries of concurrent enrollment programs have often been high achievin g students. The expanded mission of concurrent enrollment programs is to serve a wider range of students, particular ly those who represent communities with historically low The program is also seen as a way to fulfill state goals of halving the high school dropout rate and doubling the number of postsecondary credentials earned by Coloradans (Lopez, 2011; CRS enrollment courses and to improve the quality of the programs. L egislation also specifically permits students to take co ncurrent career and technical education (CTE) courses, which fits with the program intent of accelerating students to a credential through multiple pathways. 4 Key Policy Features oncurrent enrollment. JFF closely evaluated every statewide policy against their six model policy elements, and Utah ( Jobs for the Future [JFF] 2012). One key feature of the legislation is that it establishes a transparent funding process that shares costs between high schools and colleges, while keeping costs low for families The funding mechanism permits both districts and colleges to collect state fundi ng for students in concurrent enrollment to help defray As mentioned in the previous section, this 4 The Concurrent Enrol lment Programs Act also creates the th year by school districts to receive instruction beyond the senior year. The focus of this dissertation will be on the 9 th 12 th grade Concurrent Enrollment program; ASCENT s tudents will be excluded from the analysis.

PAGE 27

16 funding mechanism is a model policy element according to both JFF and ECS (Ward & Vargas, 2012; Zinth, 2014 b ). School districts use per pupil revenue (PPR) to cover tuition costs for concurrent enrollment students. Districts pay tuition to the postsecondary institution s directly on behalf of student s Previously in Colorado, families would pay for tuition costs and would possibly be reimbursed by the district later. That process, howe ver, can be prohibitive to low and middle income students and reduce access. P artnering postsecondary institution s are allowed to include concurrent enrollment student s in its determination of enrollment numbers for funding purposes. Lastly, students apply for and authorize the institution to collect the Colorado Opportunity Fund stipend to pay that portion of the tuition ( C.R.S. 22 35 105 (2)). While students and families do not pay any tuition costs, they may be responsible for books, transportation tech nology or fees, depending on local financial arrangements. Further, if students do not complete the course and do not have the permission of their principal for a non completion, they may be required to reimburse the school for the tuition costs ( C.R.S. 22 35 105 (4) ). Students and parents fill out formal paperwork to apply for concurrent enrollment, and the terms for repayment, if any, should be specified in the application (CDE, 2016). Districts are required by statute to notify families of concurrent en rollment opportunities and if any schools within the district want to concurrently enroll students, the district must enter cooperative agreements m ust, at a minimum include the following elements : The amount of academic credit to be granted for successfully completed course work by concurrently enrolled students; A requirement that concurrent enrollment course work quali fies as academic credit towards a certificate or degree, or basic skills credit ; A requirement that the local education provider ( i.e. school district, charter school or Board of Cooperative Services) pay tuition for courses completed by a student, according to the negotiated amount ; A requirement tha t the local education provider and the postsecondary institution establish an academic plan of study for concurrently enrollment students and a plan for the district to provide ongoing counseling and career planning ;

PAGE 28

17 Confirmation by the district of the st Identifier (SASID) for funding and enrollment purposes; Authorization for payment of the C ollege Opportunity Fund on behalf of the student ; Consideration and i dentification of ways for concurrent enrollment student s to remain eligible for interscholastic high school activities ; and Additional fi nancial provisions (( C.R.S. 22 35 10 4(6)) The cooperative agreements set forth the basic ground rules for the partnership between high schools, districts and postsecondary insti tutions. Often included in the agreements in addition to the components listed above, are the specific fiscal and operational arrangements regarding course location and instructors. The concurrent enrollment classes must be offered by an eligible institut ion of higher education, but can be delivered on the high school campus, college campus, online, or in a hybrid format. If the courses are taught by high school teachers they must be credentialed as college adjunct faculty. The concurrent enrollment prog ram rules specify that all qualified students in the ninth grade or higher in a public school may take courses for both high school and college credit. T o determine if a student is qualified, institutions of higher education use the same course prerequisit es they use with all other postsecondary students seeking to enroll in the same class on their campus (CRS § 22 35 104 (4)(a)). High schools and colleges have to collaborate to ensure that students are properly assessed and meeting p r ere quisite requirement s for course placement. Colleges are ultimately responsible for the course content and the quality of instruction, even if the course takes place on a high school campus taught by a high school instructor (who has been approved as an adjunct faculty member ). Contributions to the F ield Since Colorado is seen as having a model state concurrent enrollment policy ( JFF, 2012; Lopez, 2011), this study uses Colorado as a case study and begins by exploring the factors that led some schools in the state to adopt co ncurrent enrollment more quickly a nd implement it more

PAGE 29

18 widely than other schools. After understanding the key conditions at the school level the author analyze s student level behaviors by exploring what types of students are choosing to participate in the program and what the effects are of taking concurrent enrollment courses on college access and success. The study considers, in particular, if concurrent enrollment improves postsecondary outcomes among tradi tionally underserved students The findings of this study will be valuable to practitioners, policymakers, and other researchers because state policy continues to be heavily relied upon as a lever for changing educational outcomes, y et, there is not a clear understanding of whether, or how, state poli cy affects behaviors at the institutional and student levels P olicy diffusion behavior is especially informative at the sub state level in Colorado because the state has a strong local control culture, and policies and behaviors can vary by locality. Whil e Colorado provides an appropriate case study for the questions at hand, other states are also experimenting with education reforms under si milar conditions. Therefore, the findings of this study can be generalized to other states and other related educati on policy areas. Policy Diffusion and Innovation Research This study also seeks to contribute to p olicy diffusion and innovation theory. The theory is most often applied to state governments (Berry & Berry, 2007). There have been studies conducted of local governments, but the body of research is much smaller and focuses on municipalities (Shipan & Volden, 2008). Thus, this research w ill contribute to the continual exploration of the theory by applying it to a unique unit of analysis high schools. There is no apparent study on the diffusion of concurrent enrollment across high schools. Additionally, the vast majority of the studies conducted using policy innovation and diffusion have focused on the adoption of a policy without considering what occurs after a doption in the implementation stage. Scholars have identified this gap in the literature and have called for studies

PAGE 30

19 to apply policy diffusion analysis beyond a simple dichotomous measure of adoption to measures of policy implementation (Shipan & Volden, 2 012). T his study will seek to fill this gap in the literature by conducting an analysis of the factors that influence policy implementation, as measured by the share of students taking concurrent enro ll ment courses within a high school. Lastly, more recent diffusion research has focused on the importance of the characteristics of the policy itself in terms of salience, complexity and compatibility to the diffusion process (Boushey, 2010, Makse & Volden, 2011, Nicholson Crotty, 2009). These policy attributes are theorized to affect how quickly policies are adopted among states and municipalities. Makse and Volden (2011), for example, analyzed the diffusion of criminal justice laws across states and found that compatible policies those that fit seamlessly into current practices are quicker to diffuse than complex policies that require a great shift in the status quo. Given that this is a relatively newer stream of diffusion research, this study will provide a modest contribution to the literature on policy char could be considered a compatible policy according to the typology of policy characteristics (Makse & Volden, 2011, Shipan & Volden, 2012). Education Research This study also and to what extent students participating in concurrent enrollment see improvement in educational outcomes in terms of college access and college readiness (whether students are pr epared to academic ally succeed once in college). Education researchers often struggle with controlling for selection bias due to limitations on available data and analytical methods, and this is true for prior research on concurrent enrollment (Allen & Dad gar, 2012; An, 2012; Le, Casillas, Robbins, & Langley, 2005). This study will contribute to the education research field by attempting to better control for selection bias to more precisely isolate the effects of this particular intervention.

PAGE 31

20 C oncurrent en rollment programs have been around for decades, but, until recently studies of program effectiveness were limited in number and rigor Karp et al. (2007) found dual enrollment students in New York City and Florida in career and technical education program s were more likely to enroll in college, persist to the second year, and have higher GPAs and higher credit accumulation. Martin (2013) found that dual enrollment students at one North Carolina community college had higher college grades than non dual enro llment peers. Allen and Dadger (2012) evaluate d the dual enrollment program at the City University of New York and found that dual enrollees earned higher GPAs and more credits once in college. Those studies while finding positive outcomes, were narrow in focus investigating particular colleges or programs and often employed methods that did not adequately control for selection bias ( Giani, Alexander, & Reyes 2014; Taylor, 2015 USDOE, 2017 ). There is one study that meets rigorous quasi experimental desig n standards and is broad in scope An (2013) used a national dataset and found that du al enrollment programs increase degree attainment rates for first generation students (USDOE, 2017). Very recently r esearchers have publish ed quasi experimental evaluatio ns of statewide concurrent enrollment programs These studies were possible due to the recent expansion of statewide longitudinal data systems. t program and found positive effects on college enrollment, particularly for students who are lower performers academically. Taylor (2015) followed graduating class of 2003 to track college entrance and completion rates for dual enrollment studen ts and fo und positive effects overall, though the effect sizes were smaller among low income students and students of color. Haskell (2016) analyzed 2008 and 2009 high school graduates in Utah and found reduced time to college degree completion and potenti al financial savings to families and the state. Giani, Alexander and Reyes (2014) use the statewide longitudinal data system in Texas to track 2004 high school graduates into college. Their study found greater

PAGE 32

21 enrollm ent, persistence and completion among d ual credit students as compared to non dual credit students Importantly, dual credit students enjoyed greater postsecondary benefits as compared to students taking other forms of advanced coursework such as Advanced Placement and International Baccalaurea te courses (Giani et al. 2014). All but one of the above mentioned studies ( Allen & Dadgar, 20 12; An, 2013; Cowan & Goldhaber, 2014; Giani et al., 2014, Karp et al., 2007, Martin, 2013 Taylor, 2015 ) were recently evaluated by the What Works Clearinghous e (WWC) and only two met the WWC design standards with reservations: Giani et al. (2014) (2013) nationally representative study (USDOE, 2017). 5 While the Illinois, Utah, Texas and Washington studies indicate that concurrent enrollment students participating in state wide programs have more positive postsecondary outcomes than their non participating peers, each study is set in its own state policy context. The Texas program design is substantially di The Texas and Illinois studies both use graduating cohorts from the earlier part of the 2000s, which allows them to follow students further into higher education, but also negates the ability to identify more recent trends. With th e exception of Giani et al. (2014), none of the studies use s an intensity measure of concurrent enro l lment participation (e.g. number of credit hours taken). does not include data on Texas high school graduates who attend c ollege out of state in their college matriculation model, which could bias their results. Further, there is value in determining whether concurrent enrollment outcomes are consistent across states. The concluding chapter of this dissertation consider s how results compare to the findings in these other state studies. In summary, additional research beyond the emergent state studies is needed for the field to gain confidence in concurrent enrollment as an effective college readiness intervention 5 Haskell (2016) was not reviewed by the WWC.

PAGE 33

22 p findings across states While prior re sear ch has found dual enrollment programs result in benefits for the average student, it is unclear to whom those bene fits accrue and to what extent. Few studies have examine d the effects on tradit ionally underserved populations, and of those that have the results are inconsistent Taylor (2015) found minority and low income students saw smaller gains in postsecondary ou tcomes when compared to their peers in Illinois, while An ( 2013 ) found higher effect sizes for students from disadvantaged backgrounds It is evident that w ith 47 states having statutes governing concurrent enrollment programs much can still be learned abo ut the effectiveness of these policies Summary & Research Questions ithin five years 91 schools offered concurrent enrollment to some degree. Given the rapid d iffusion of the program, this study will seek to identify the local variables and conditions that affect the decision to adopt concurrent enrollment programs in high schools in an effort to uncover any best practices that could be applied to other states t rying to scale up similar programs. T he first research questi on is stated as follows. RQ 1: What factors influence whether or not high schools adopt concurrent enrollment programs? Additionally b variance and room for innovation at the local level in regards to whether and how the program is implemented. High schools may adopt concurrent enrollment to add another option to an already existing portfolio of college readiness or credit accrual progra ms (e.g. Advanced Placement courses, International Baccalaureate program, honors courses, etc.). Alternatively, a high school may launch concurrent enrollment as a way to provide access to college level cou rses to all or nearly all upper classmen. A

PAGE 34

23 rural school may, for example, enroll all seniors in a concurrent enrollment college level math course. Offering a concurrent enrollment math course to a classroom of seniors at a large high school would only constitute a small percentage of the total senior cla ss, whereas at a small, rural to which students are participating in concurrent enrollment within a high school leads to a sub question: RQ 1a: What fac tors influence the extent to which concurrent enrollment programs are utilized by students within high schools? To date, there is no apparent empirical examination into the school or district level characteristics that lead to faster or deeper program ado ption at certain high schools as compared to others. T his study also seeks to understand if students participating in concurrent enrollment program see improvement in educational outcomes in terms of both their participation in college and their success once in college. The author considers if the program has positive effects underserved students in particular Accordingly, the second and third research question s are as follows: RQ2: How does participation in concurre nt enrollment affect the college going rates of RQ3: How does high school participation in concurrent enrollment affect the college performance and persistence of students? These two research questions combined with the fi rst question are collectively important because in order for state facilitated, voluntary policies to significantly improve educational outcomes the policy needs to be both widely diffused in schools and impactful on individual students A nswers to these questions will be beneficial to policymakers, practitioners and researchers.

PAGE 35

24 CHAPTER II LITERATURE REVIEW AN D HYPOTHESES Several different literature streams inform the r esearch questions set forth in the previous chapter. The first research question con cerning differences in school level adoption of concurrent enrollment is informed by policy innovation and diffusion theory. T he second and third research question s which are concerned with the individual educational outcomes of participating in concurren t enrollment, rely on different strands of education theory. Hypotheses are drawn from the literature and presented throughout the chapter. Policy Diffusion & Innovation Theory Policy diffusion and innovation theory submits that political, economic and soc ial factors, along with competitive and emulative pressures, influence whether or not policy change is adopted (Mok h the diffusion of innovation among American states in which he sought to understand why some because he focused n organizational innovation literature) but also on the role of competitive and emulative pressures among states. Walker (1969) observed that national professional communities served as lear ning opportunities where ideas were spread among state policy makers and administrators. He also noted that some states were seen as leaders, and sought to determine if other states were more likely to emulate policies of the leader states (Walker, 1969). propositions (Berry, 1994b ; Shipan & Volden, 2012 ). Some scholars focused solely on the internal factors, or determinants, that lead states to be early innovators, while ot her scholars focused their

PAGE 36

25 research on regional diffusion and national interaction patterns (Berry, 1994b; Berry & Berry, 2007; Mok h er & McLendon, 2009). In the 1990s, research by Berry & Berry (1990, 1992) significantly advanced the research genre by offe ring a methodology event history analysis that provided a way to empirically test the effects of both internal determinants and external diffusion factors in one model. Since then, policy diffusion and innovation theory has been applied using event history analysis to a wide range of substantive topics, such as health care (Stream, 1999; Volden 2006) hate crime laws ( Soule & Earl, 2001 ) electricity deregulation (Ka & Teske, 2002) and education (Mintrom, 1997; Wong & Shen, 20 02) The majority of empirical applications of policy diffusion focus on state governments, but the theory has also been applied to local municipalities ( e.g. Bingham, 1977; Hoyman & Weinberg, 2006; Lubell et al. 2002). These studies, along with others, considered different mechanisms that drive the diffusion process, including policy entrepreneurs (Ball a 2001; Mintrom 1997 ), learning that occurs from effective policies (Gilardi, Fglister & Luyet, 2009; Volden 2006), competition (Baybeck et al., 2011, Berry & Berry, 1990) and coerciv e forces (Karch, 2006; Welch & Thompson, 1980). More recent diffusion research also has focused on the importance of the characteristics of the policy itself in terms of salience, complexity and compatibility to the diffusion process (Boushey, 2010, Makse & Volden, 2011, Nicholson Crotty, 2009). Makse and Volden (2011), for example, found that compatible policies those that fit seamlessly into current practices are quicker to diffuse than complex policies that require a greater shift in the status quo. Whi le the policy diffusion and innovation literature is wide and varied, the theory is generally focus ed on the following overarching resources that it has to innovate, the obstacles that stand in the way of th is innovation, other co an & Weinberg, 2006, 98 99). The se factors often comprise the central elements of empirical models

PAGE 37

26 used to understand and predict why cer tain public entities are quicker to adopt innovative policies than others. Gi ven the lack of research on what affects policy implementation beyond the initial state of policy adoption, the same theoretical grounding will be used to explore both adoption an d implementation effects in this study (Shipan & Volden, 2012) Motiv ation to I nnovate The policy innovation literature theorizes that if problem severity is high, governmental entities are more likely to be motivated to adopt new policies (Berry & Berry, 1990, 2007; Hoyman & Weinberg, 2006; Mohr, 1969). Berry and Berry (2007), in their review of the literature on policy policy directly by clarifying the nee d for the policy, or indirectly by stimulating demand for the policy found in several studies covering topics such as welfare, school choice and health care re form (Allard, 2004; Mintrom & Vergari, 1998; Stream, 1999). Concurrent enrollment programs can be viewed as a tool for increasing student achievement levels and improving college readiness (Hoffman, 2005). Thus, districts struggling with low academic achi evement levels may have more incentive to innovate and improve outcomes and may be more likely to turn to concurrent enrollment as a potential solution. The severity of low academic achievement in districts and schools should motivate them to adopt concurr ent enrollment, whether that pressure to improve comes from internal leadership, the state (via performance ratings ), or concerned parents. Likewise, if a school has an urgency to improve the achievement of its students, it could be hypothesized that the s chool will encourage higher levels of participation in concurrent enrollment courses. That is, if a high school adopts the program as an improvement strategy, it would follow that the school would actively encourage and recruit students to participate in i t. A hypothesis on the motivation to innovate is proposed as follows.

PAGE 38

27 Hypothesis 1 .1 : High schools that have lower academic achievement levels are more likely to adopt concurrent enrollment and have higher student participation rates Resources to Innovat e Policy diffusion and innovation theory also points to the importance of resources, which are needed both for the innovation itself and to overcome any obstacles to innovate (Berry, 1994b). There are two types of resources that the literature on policy di ffusion and innovation identifies that are particularly relevant to this study: fiscal capacity and organization size. Generally, past studies the resour ces at its disposal (Berry & Berry, 2007). In regards to fiscal capacity, the literature on policy innovation draws from the broader literature on organizational innovation, which has consistently found that financially secure organizations are more likel y to innovate than organizations in fiscal trouble or with fewer slack resources (Berry, 1994a; Bingham, 1977 ; Cyert and March, 1963; Rogers, 1983). Policy innovation theory points to the fact that many new policies and programs require extensive funds to be implemented, and thus agencies with abundant resources may be more inclined to adopt such pro g rams (Berry & Berry, 2007) and may have a greater capacity to widely implement them Therefore, drawing from the literature and theory on financ ial resources a nd innovation, the following hypothesis is proposed. Hypothesis 1. 2: High schools with greater fiscal capacity are more likely to adopt concurrent enrollment and have higher student participation rates Theory and research on organizational innovation has long considered the size of an organization as another key explanatory variable that is positively associated with the likelihood of innovation (Baldridge and Burnham, 1975; Berry, 1994a; Cyert & March, 1963; Mohr, 1969; Rogers, 2003). Size is considered a n important element because it facilitates the presence of other factors that may affect innovation such as the availability of slack, or surplus, resources and specialized staff

PAGE 39

28 (Rogers, 2003). Larger organizations tend to have more resources, and they ar e more likely to have the capacity and structure to hire specialized staff. Having administrators whose job it is to stay attuned to the latest research and innovations in their specific program area makes it more likely that schools will be aware of new p rograms and have the capacity to implement them (Baldridge, 1975; Berry, 1994a). A s school size increases however, it may be more likely that the proportion of eligible students who participate in innovative programs decreases. This is particularly true i f larger schools have a broad array of programs from which students may select. Larger organizations may have greater numbers of participants compared to smaller organizations, but in terms of participation rates it seems likely that organization al size ma y be inversely related to overall par ticipation rates. Based on the organizational innovation literature and general reasoning the following resource hypothesis is proposed. Hypothesis 1. 3: The larger a school, the more likely it is to adopt concurrent en rollment but the share of students participating may be lower External Factors P olicy diffusion scholars have devoted much attention to how competitive and emulative pressures among states affect policy innovation (Berry, 1994b; Berry & Berry, 2007; Walk er, 1969). Policy innovation and diffusion theory points primarily to two avenues through which external pressure for policy change occurs: through regional diffusion that is, through competition with or emulation of neighboring governments or through nati onal interaction, which is the idea that policy ideas spread through networks of policymakers (Berry & Berry, 2007; Walker, 1969). Empirical tests of diffusion models have found varying results. A national study of the diffusion of concurrent enrollment po licies among states, for example, found that regional diffusion pressures were not a statistically significant factor (Mokher & McLendon, 2009). On the other hand, a recent study of the diffusion of charter school legislation among states found the regiona l diffusion variable to have a statistically significant and substantive influence on policy adoption (Lee, 2014). Some scholars

PAGE 40

29 argue that the geographic focus of policy adoption is an outdated concept given that policymakers and leaders can learn about i nnovative programs not just from their geographic neighbors, but from others across the states or increasingly across the world (Shipan & Volden, 2012; Volden, Ting & Carpe nter, 2008). Nonetheless, a s an open enrollment state, schools in Colorado have an incentive to compete with one another for students because those students come with revenue attached. If a high school sees that a neighboring school is offering advanced or enhanced programming, it seems likely that competitive and emulative pressures cou ld influence policy innovation decisions and may also lead to more robust implementation reflected by higher participation rates Thus, although research and theory is unclear on the influence of regional pressures, this study hypothesizes that geographic p roximity wi ll be influential. Hypothesis 1. 4: High schools nearby other schools that have already adopted concurrent enrollment are more likely to offer the program and have higher participation rates Another external factor that could be relevant to this a community college. There is precedence in the literature for focusing on this factor; the influence and presence of community colleges was used in the previously mentioned study of national dual enrollment policy diffusion (Mokher & McLendon, 2009). The nearness of a community college to a high school makes the implementation of the concurrent enrollment policy easier in terms of accessibility to courses delivered at a college campus, credentialing of high school instructors or the provision of community college instructors (in cases where college faculty teach courses at a high While some concurrent enrollment is provided by four year institutions, the large majority of course enrollments are thr ough community colleges. I n Colorado during the 2015 16 school year, for example, 88 percent of concurrent enrollment students took courses through community colleges (C olorado D epartment of H igher E ducation [CDHE] 2017 b ) More over

PAGE 41

30 community colleges are proponents of concurrent enrollment because it is an immediate revenue generator, as well as a recruitment strategy (Crooks 1998; Morest & Karp 2006). Community colleges may exert pressure on schools to offer concurrent courses, in which case, if a community college is geographically close to a particular high school, the po licy may diffuse more readily and, perhaps, more deeply, which would be evidenced by a greater share of students within a high school taking concurrent enrollment courses Hypothesis 1. 5 : High schools with greater proximity to a community college will be m ore likely to adopt concurrent enrollment and have higher participation rates. Education Theory Because education is such a broad field, it is necessary to first situate the research through a specific theoretical lens. Some researchers take an institution al rational choice lens, for example, and place emphasis on how institutional arrangements constrain and aggregate individual choices resulting in organizations of different types and quality (Chubb & Moe, 1990). Other researchers focus on how manipulating inputs within existing organizational and institutional arrangements can alter performance outputs. These scholars, which devote their time to identifying problems and evaluating reforms that occur in and around schools, could loosely be grouped under a t heoretical framework of school improvement. The theoretical reasoning behind concurrent enrollment is grounded in the viewpoint that there are factors within the control of schools that can be manipulated to improve education outcomes. Empirical research has linked a variety of factors to the likelihood that students will attend and be successful in college. Some factors cannot be altered by schools such as socioeconomic status but other factors, including academic preparation and metacognitive skill devel opment, can be manipulated. Thus, the school improvement framework is an applicable lens for this inquiry. T his section of the literature review provides an overview of education theory and

PAGE 42

31 research related to improving student achievement, generally, and college access and success, specifically. School Improvement Framework There are a significant number of education researchers who rely on a theoretical lens that focuses fundamentally on manipulating inputs within existing school arrangements to alter pe rformance outputs (Hanushek, 2003; Purkey & Smith, 1983). Scholars coming from the field of economics refer to these types of studies as education production functions, which are used to their families, and other economics scholars could be loosely grouped together under a theoretical framework of school improvement, which asserts that the the 1966 Coleman Report formally known as Equality of Educational Opportunity (Chubb & Moe, 1990; Hanushek, 1979, 2003; Purkey & Smith, 1983). The Coleman Report (1966) which was the product of a substantial study covering over 600,000 students in 4,000 schoo ls, found that once the family background of individual students and the overall racial composition of schools were taken into account, school characteristics had little effect on achievement le vels (Coleman et al. ). The school characteristics included in the Coleman study were numerous and included such factors as school funding, classroom size, teacher and principal salaries, education levels of teachers, number of free textbooks, and extracurricula r offerings (Coleman et al., 196 6). In a backlash agains spent the following decades attempting to prove that school based elements do matter (Hanushek, 2003; Purkey & Smith, 1983). One rationale for such research is that factors related to family background are not easily ch anged

PAGE 43

32 at least in the short term and, therefore, the many factors that can be manipulated at the classroom, school, or district level must continually be examined in the effort to improve schools (Purkey & Smith, 1983). Some researchers chose to focus on o rganizational and cultural attributes not included in the Coleman Report; this group of scholars contributed to what is known as the 1981; Purkey & Smith, 1983). While encompassing a diverse set of studies and findings, overall, the effective schools research conducted in the 1970s and early 1980s emphasized the following consi derable degree of control by the staff over instructional and training decisions in the school, clear leadership from the principal or other instructional figure, clear goals for the school, and a Empirical evidence that these cultural and organizational elements could affect student achievement (as measured typically through assessment scores) was heralded as proof that schools matter (Edmonds, 1979; Hersh et al. 1981 Purkey & Smith, 1983). Other researchers took a different path and have focused on re examining the variables included in the Coleman Report and have found some nuanced, positive effects on student achievement. Several studies, including the well known Tennessee STAR experiment, for e xample, have found statistically significant, positive effects of small classroom size (below 20 students per teacher) on student learning in Kindergarten through third grade (Centra & Potter, 1980; Mosteller, 1995; Walberg, 1982). Others, however, have fo und little evidence that class size affects learning enough to warrant its financial costs (Hanushek, 1999; Funkhouser, 2009). The debates within the literature on classroom size are representative of the school improvement research as a whole in that ther e are some positive findings but much debate over meaning and replication Analysis of other inputs such as school funding, teacher education and teacher salaries constitutes a large, and

PAGE 44

33 somewhat controversial, literature. The wide variety of variables, t he broad range of statistical methods used, and the differin g levels of emp irical quality have resulted in findings that are often seen as contradictory (Hanushek, 1979). Hanushek (2003) calls for more rigorous research on school ducational policies are to be improved, much more serious attention must In the last 20 years, scholars have indeed continued to analyze a variety of school based pr recent research has been on topics such as standards and curriculum, bilingual education, literacy programs, and improving the transition from high school into college (Hoffman, 2005; Darling Hammond, 2010; Kirst & Venezia, 2004, Hoffman, Vargas & Santos, 2008 a, 2008b ). Initiatives focused on aligning secondary and postsecondary systems are often referred to as K 16 or P 20 initiatives. Research on P 20 syst ems was epitomized in the work of the Stanford Bridge Project in the late 1990s, which made the case that the underpreparation of high school graduates for higher education was a pervasive and critical problem (Venezia, Kirst & Antonio, 2003). The research ers found the coursework offered in high school and college to be disconnected, and they noted that preparatory courses (Venezia, Kirst & Antonio, 2003, 8). The expansion of dual enrollment programs was one of the key recommendations of the Stanford Bridge Project to improve the transition from high school to college. Dual enrollment is seen as an avenue for students to gain stronger academic preparation for college (Kirst & Venezia, 2004; Venezia, Kirst & Antonio, 2003). The next section of the literature review will explore in detail the theory behind how academic

PAGE 45

34 Academic Pr eparation E mpirical research studies over the past half century have consistently identified that disparities in curriculum offerings, including course options, rigor of curriculum, and quality of courses, contribute to and exacerbate achievement gaps. Thi s stream of literature focuses on the inequity of educational opportunity between groups of students, and the detrimental effect that has on academic success. city elementary students in Chicago, for example, found that blac k and white students of similar aptitude performed equally well when exposed to the same instruction high or low quality (Dreeben, 1987). The problem is that high quality instruction is not offered in every classroom, and low income, minority students tend to disproportionately receive inadequate instruction (Darling Hammond, 2010). This contributes to persistent achievement gaps as evidenced through several studies that have found that students who are exposed to rich, challenging curriculum eventually out perform their peers who are placed into less rigorous classes, even after controlling for socioeconomic background (Alexander & McDill, 1976; Gamaron, 1990; Gamaron & Hannigan, 2000; Oakes, 2005). Peterson (1989), for instance, conducted an experimental st udy that randomly placed at risk 7th graders with similar backgrounds into varying levels of math classes. St udents placed into the highest math class (containing a pre algebra curriculum) outperformed the other students on assessments given at the end of the school year (Peterson, 1989). Researchers have found that the disproportionate allocation of high quality instruction to students occurs primarily through two ways. The first is that schools with minority majority populations (i.e. serving mostly Hisp anic, black or Native American students) offer fewer academically rigorous courses. Instead of having a selection of honors, Advanced Placement, lab science and foreign language courses like high schools in wealthy districts do, high schools serving

PAGE 46

35 large numbers of minority and low income students often offer mostly remedial and vocational courses (Pelavin & Kane, 1990; Oakes, 2005; Darling Hammond, 2010). The second way that high quality instruction is allocated away from low income, minority students is through tracking. In schools with socioeconomically diverse populations, white and upper income students tend to be placed into college preparatory classes while minority and low income students are tracked into lower level courses (Darling Hammond, 2010; Oakes, 2005) As Darling (2005) has conducted empirical studies over the past several decades that highlight the r studies, for example, found that after controlling for test scores, white and Asian students were far more likely to be placed into honors courses than their peers (Oakes, 1993). High achieving Latino students who scored at the 90th percentile on standar dized tests had just a 56.3% chance of being assigned to a college preparatory class, as compared to 97.3% of Asian students and 93.3% of white students scoring in the same percentile (Oakes, 1993). The reasons for the underrepresentation of minority and low income students in honors, AP and other challenging secondary courses are many Tracking begins at an early age and by the time students reach high school tracked students often do not have the prerequisite skills or test scores to take advanced cour ses (Darling Hammond, 2010; Oakes, 2005). More directly counselors may advise students from low socioeconomic backgrounds away from challenging postsecondary pathways and towards low status careers (Darling Hammond, 2010). Further, middle and upper incom e parents tend to be more active in pushing for their children to be placed into advanced courses and programs and to be assigned to the best teachers. As Darling Hammond (2010) explains, high quality education is scarce and Scarce resources tend to get a llocated to the students whose parents, advocates or representative have the most political leverage. This typically results in the most highly

PAGE 47

36 qualified teachers offering the most enriched curricula to the most advantaged students. (60) Indeed, other emp irical work has found tracking patterns in place for teachers in which the best (most experienced, most competent) teachers are assigned to the brightest students in upper level classes, while inexperienced and ineffective teachers are assigned to lower le vel classes and students (Finley, 1984; Talbert, 1990). As this section of the literature review demonstrates, disparities in curriculum offerings and ensuring that all students have access to rigorous coursework in high school through programs such as concurrent enrollment could improve education outcomes. Concurrent enrollment programs expand the number of accelerated learning options available to schools, espe cially when programs have clear funding streams (Allen, 2010; Karp, Bailey, Hughes, & Fermin, 2005). Making concurrent enrollment courses more prolific in high schools and targeting them to students of all socioeconomic backgrounds could be a way to ensure traditionally underserved students have access to enriched, advanced curriculum (Karp, Bailey, Hughes, & Fermin, 2005; Venezia, Kirst, & Antonio, 2003). For students who have been tracked into lower level courses from an early age and enter high school fa r behind their peers academically remedial courses to high school seniors to help students become college ready (Allen, 2010; Rutschow & Schneider, 2011). This a newer strategy; dual enrollm ent courses have long been targeted to high achieving students, but schools are now expanding the mission of such programs to serve less academically prepared students (Allen, 2010; Karp et al., 2007; Rutschow & Schneider, 2011; Venezia, Kirst, & Antonio, 2003). Metacognitive Learning Skills When student do not take advanced, college preparatory coursework, it is not content knowledge alone that students miss (e.g. algebra vs. pre algebra) but also the development of

PAGE 48

37 higher order thinking and reasoning sk ills (Darling Hammond, 2010). Researchers have found that classes in the lower tracks often focus on rote memorization, test taking, and behavioral problems (Eckstrom & Villegas, 1991; Good & Brophy, 1987; Oakes, 2005). In contrast, other research studies have shown that in college preparatory tracks, teachers engage students in hands on group activities and projects that encourage students to be creative, to problem solve and to think critically and strategically (Braddock & McPartland, 1993; Garcia, 1993; Wenglinsky, 2002). It is the latter skills that students need to acquire to ultimately be successful in higher education (Conley, 2007, 2010). In fact, research on college readiness has long made the point that content knowledge alone does not predict suc cess in college there is a host of other knowledge, skills, and behaviors students must acquire to be successful in postsecondary education (Attinasi, 1989; Byrd and MacDonald, 2005; Conley, 2005, 2007, 2010; Dickie & Farrell, 1991; Shields, 2002). These o ther a misnomer because the skills and behaviors are directly related to cognition and thinking processes. motivational factor s, experiential and contextual intelligence, social skills and interests, and Students with well developed metacognitive learning skills will be able to manage thei r time effectively, think critically, navigate college resources, maintain study routines, have self awareness of their strengths and weaknesses, analyze and interpret information, and have the confidence to overcome challenges (Conley, 2010, 2013).

PAGE 49

38 Concu rrent enrollment is a venue for students to acquire metacognitive learning skills. As Karp (2012) explains, Dual enrollment can be seen as a social intervention in which potential college students learn about the norms, interpersonal interactions, and beh aviors expected for college exposure and practice, coming to feel comfortable in a college environment and ultimately becoming successful once they matriculate. (22 23 ) Karp (2012) tested facets of sociological theory in a qualitative study that measured the degree to which students in concurrent enrollment courses gained knowledge of college behaviors, norms and processes. The students in her study took dual enrollmen t courses at their high school, taught by a teacher who was credentialed as a college adjunct. Despite not being physically located on a college campus, Karp found that dual enrollment courses still gave students the opportunity to assume the ollege student, which helped them gain a deeper knowledge of what would be expected success by learning before they actually matriculate all aspects of the colle finding supports other research demonstrating that students who take concurrent enrollment courses have higher levels of self efficacy, or confidence, in their academic abilities (Margolis & McCabe, 2004). College Affordability Even if students have acquired the right knowledge, skills and abilities to be successful in college, college affordability remains a significant barrier to matriculation. Numerous studies have found that low and middle income students are more sensitive to tuiti on and aid changes than wealthier students, meaning that when tuition increases or grant aid decreases, there is a bigger decline in enrollment for low income students than for upper income students (see e.g. Heller, 1997; Leslie & Brinkman, 1987; Terenzin i, Cabrera, & Bernal, 2001; St. John, 1990). In other research on college affordability, one study analyzed enrollment rates by income level and standardized tests

PAGE 50

39 scores and found that high income students who scored among the worst on the achievement tes t were as likely to go to college as low income students who performed among the best (Kahlenberg, 2004). Put another way, "the least bright rich kids have as much chance of going to college as the smartest poor kids." (Kahlenberg, 2004, 24). The study als o found that 22 percent of low income students with the highest test scores did not go to college, compared to 11 percent of middle income students and only 3 percent of high income students with the same scores (Kahlenberg, 2004). Concurrent enrollment c an be an opportunity to make college more affordable by allowing students to earn college credit for free, or at a reduced cost, while still in high school, thus reducing the amount of tuition students will have to pay when they matriculate to college (Hof fman, Vargas, & Santos, 2008a, 2008b ; Jobs for the Future, 2006). In some programs, students are encouraged to accumulate enough credits to earn a certificate or associate degree at the same time as they earn their high school diploma From this review of the literature it is theorized here that concurrent enrollment will improve college access and success for its participants for the following central reasons: 1) C oncurrent enrollment provide s for rigorous academic preparation and enhanced content knowledge ; 2) C oncurrent enrollment courses are a venue for students to acquire metacognitive learning skills and exposure to higher education; and 3) Courses provide the opportunity for students to earn free, or low cost, college credit, thus reducing the total amount of a college credential. Following this theoretical framew ork, a hypothesis regarding colle ge access is stated as follows. Hypothesis 2 : High school students who participate in concurrent enrollment programs will have a greater probability of enrolling i n higher education. Similarly, the author expects positive

PAGE 51

40 outcomes in terms of college success based on the theoretical framework, leading to a third central hypothesis. Hypothesis 3 : First year college students who had participated in concurrent enrollme nt programs in high school will have greater academic success and a higher probability of persisting than college students who did not participate. Summary Based on a comprehensive review of the literature, several hypotheses have been identified to guide summary of the research questions and the associated hypotheses. The following chapter will discuss the data and methods used to test the hypotheses. Table 2 Summary of Research Questions and Hypotheses Research Questions Hypotheses RQ 1 : What factors influence whether high schools adopt concurrent enrollment programs ? RQ1a : What factors influence the extent to which concurrent enrollment pr ograms are utilized by students within high schools? H1.1 : High schools that have lower academic achievement levels are more likely to adopt concurrent enrollment and have higher student participation rates. H1.2 : High schools that have greater fiscal capacity are more likely to adopt concurrent enrollment and have higher student participation rates. H1.3 : The larger a school, the more likely it is to adopt concurrent enrollment, but the share of students participating may be lower. H1.4 : High schoo ls nearby other schools that have already adopted concurrent enrollment are more likely to offer the program and have higher student participation rates. H1.5 : High schools with greater proximity to a community college will be more likely to adopt concur rent enrollment and have higher student participation rates. RQ 2 : How does participation in concurrent enrollment affect the college school students? H2 : High school students who participate in concurrent enrollment program s will have a greater probability of enrolling in higher education. RQ 3 : How does high school participation in concurrent enrollment affect the college performance and persistence of students? H3 : First year college students who had participated in concu rrent enrollment programs in high school will have greater academic success and a higher probability of persisting than college students who did not participate.

PAGE 52

41 CHAPTER III DATA & METHODS This chapter describes the data and methods used to e xplore the hypotheses T he study begins with an exploration of factors that influence the adoption of concurrent enrollment programs among and within Colorado high schools using school level, administrative data in event hi story analysis and multivariate regression The author then undertakes propensity score matching and fixed effects regression using s tudent level data to analyze the effects of participating in concurrent enrollment on college matriculation and success Data Sources and Collection Nearly all the d at a used in this study were collected through agencies: the Colorado Department of Higher Education (CDHE) and the Colorado Department of Education (CDE). The author constructed panel data sets by compiling publicly available data fr om the agency websites and by procuring a de identified and secure cro ss section of student level dat a from CDHE. The datasets are timely, comprehensive and most important longitudinal between K12 and higher education due to data sharing agreemen ts in pla ce between the two agencies. Statewide longitudinal data systems are still a relatively new phenomenon having rapidly expanded in states over the past decade. The wealth of information included in these state data systems has the potential to help transfo rm the public administration of schools and colleges into a truly evidence based sector. There are obstacles along the way to reaching that goal, however, including the recent backlash from parents and community members around the perceived overreach of go vernment and businesses in collecting data on students. In fact, between 2013 and 2016, 36 states enacted 74 student data privacy laws, some of which establish important procedures for protecting student data, but some of which also constrain the ability o f state agencies to share student data with researchers (Data Quality Campaign, 2017) Thus, it is within this landscape that

PAGE 53

42 this study occurred, and the author acknowledges that access to such rich and powerful data is not to be taken for granted in educ ation research. Moreover, there is an important contribution that can be made to the debate around the value of making student data available to researchers if studies such as this prove worthwhile to policymakers and practitioners. High School Panel The f irst panel data set was constructed by collecting high school data from CDE, CDHE and the U.S. Census Bureau for the following academic years: 2009 10, 2010 11, 2011 12, 2012 13, 2013 14, and 2014 15. The full data set includes 388 high schools that served at a minimum, grades 10, 11 and 12 during the entirety of the study period. High schools that opened recently and were not in existence for all years of the study, closed during the study period, or only served partial grades (e.g. 9 and 10) during the s tudy period were exclude d from the dataset. 6 The resulting panel is strongly balanced, meaning all schools have data for all years of the study. Aggregate data on multiple indicators for the 388 high schools were obtained from CDE website. The indicator s collected include school performance ratings, school type (i.e. charter, alternative education campus, or traditional school), student count, district setting, prior dual enrollment program participation rates, and free and reduced price lunch informatio n The data that were procured from the CDHE provide details about which high schools and higher education institutions offer concurrent enrollment programs and how many high school students were in enrolled in the program during a given academic year Col lege matriculation rates by high school were also obtained from the CDHE. Lastly, data w ere American Community Survey for median household income. 6 Given the geographical focus of the policy diffusion research question and hypotheses, online high schools (n=19) were also excluded from the data set.

PAGE 54

43 The dataset provide s a sufficient scope of variables and a long enoug h time span to analyze the first research question concerning the diffusion of concurrent enrollment among and within high school s The Colorado legislature passed the Concurrent Enrollment Program Act in spring 2009, and the program was fully operational at the start of the 2010 11 school year. The school panel dataset used in this study spans from fall 2009 through spring 2015, allowing for an analysis of the first five years of program implementation. Student Panel The second panel data set was created from data collected through Student U nit Record Data System (SURDS), which houses comprehensive postsecondary data on students who are enrolled at public colleges and universities in the state, as well as those enrolled at three private institutions : the University of Denver, Regis University, and Colorado Christian University. The CDHE supplements SURDS with data from the National Student Clearinghouse (NSC) to provide information on out of state enrollment and enrollment at private institutions. Th e NSC has a coverage rate of 96 percent of all students enrolled in a U.S. public or private college (NSC, 2013) ; t hus, when this study considers college enrollment patterns the dataset captures nearly all Colorado high school graduates who attend college whether in state or out of state, at a public or a private institution. Further, CDHE has est ablished a partnership with CDE that permits the linkage of the postsecondary data with K 12 data using the State Assigned Studen t Identifier (SASID). Th e SASID linked databases provided the means to create a student level panel dataset that follows cohorts of high school graduates as they move from the K 12 system into higher education. The high school graduating cohorts of 2011, 2012 and 2013 are included in the student level analysis The variables included in the second panel data set provide details, by semester, about what postsecondary institution students are enroll ed in whether they require remedial education, and how they perform in terms of grade point average, credit accumulation, and persistence. The data

PAGE 55

44 that were procured from CDHE around concurrent enrollment include how many credit hours students take and in which high schools and higher education institutions the students are concurrently enrolle d. Research Design The research design begins with an event history analysis of how concurrent enrollment programs expanded among Colorado high schools Using the high school as the unit of analysis the author also uses regression analysis to see if any of the same factors included in the event history analysis affect the magnitude of program participation rates Next, the author conducts multivariate analyses to evaluate the effects of participating in concurrent enrollment on education achievement usin g student level data. Participation in concurrent enrollment is explored both as a dichotomous measure (yes/no) and as an intensity level (i.e. number of credits). The different components of this research design r ely on the same dataset but have different guiding questions and units of analysis. A description of the variables, measure, and methods used in the policy diffusion portion of the study are presented first, followed by an explanation of the variables, measures and methods employed for the student level policy evaluation. The chapter concludes with a summary of the research design. Policy Diffusion Variables and Measures The hypotheses relating to the policy diffusion analysis contain several key concepts related to policy adoption, motivation to i nnovate, resources and obstacle s, and external factors. Table 3 summarizes the indicators that are used to operationalize the explanatory variables in the five diffusion hypotheses, and the following sections provide additional details. Ta ble 4 provides a summary of variable descriptions and data sources. Concurrent enrollment policy adoption. The key dependent variable for all of the hypotheses in the policy innovation and diffusion analysis is adoption of concurrent enrollment

PAGE 56

45 programs. First, this study indicating that the high school has adopted the concurrent enrollment policy and there is at least one student taking a concurrent enrollment course. The study then includes a second measure of adoption to assess how the covariates affect the magnitude of student participation within a high school. That measure is a proportional value calculated by dividing the number of students participating in concurrent enrollment in a given academic year by the total number of students in grades 9 through 12 in that same year. 7 Table 3 Concept Measurement Summary: Policy Diffusion Hypotheses Constructs Indicators 1 .1 : High schools that have lower academic achievement levels are mo re likely to adopt concurrent enrollment and have higher student participation rates. Academic achievement levels College enrollment rates School performance rating (index of ACT scores, graduation rates, dropout rates and achievement on state standardized tests) 1. 2: High schools that have greater fiscal capacity are more likely to adopt concurrent enrollment and have higher student participation rates Fiscal capacity Median household income Free and reduced price lunch eligibility 1. 3: The larger a sch ool, the more likely it is to adopt concurrent enrollment, but the share of students participating may be lower. High school size Student count 1. 4: High schools nearby other schools that have already adopted concurrent enrollment are more likely to offe r the program and have higher student participation rates Proximity to adopters Number of High Schools within 5 miles offering Concurrent Enrollment 1. 5: High schools with greater proximity to a community college will be more likely to adopt concurrent e nrollment and have higher student participation rates Proximity to community colleges Distance in miles from nearest community college Number of community colleges within 10 miles 7 Some of th e schools included in the study serve more grades than 9 12 (e.g. K 12 schools, or secondary schools serving grades 6 or 7 through 12 th grade). In all cases, only the population of grades 9 12 is used as the denominator for calculating the share of student s participating in concurrent enrollment.

PAGE 57

46 Table 4 : Variable Descriptions and Sources Varia ble Description Source High school adoption of concurrent enrollment program Dummy variable (yes = 1; no = 0) indicating whether a high school adopts concurrent enrollment in a given year during the study. Colorado Department of Higher Education College matriculation rate (%) Annual measure of the percent of high school graduates who enroll in a postsecondary institution in the fall immediately following graduation. Colorado Department of Higher Education School performance rating (%) Annual index measur e of the percentage of points earned on state performance framework that includes ACT scores, graduation rates, dropout rates and achievement on state standardized tests. The higher the percentage of points earned, the better the school performed on the me asures. Colorado Department of Education Median household income (logged) Log of the median household income in the past 12 months (in 2014 Inflation adjusted dollars) for the five year period running from Jan. 1, 2010 Dec. 31, 2014. The aggregated 5 y ear survey was used to obtain neighborhood level estimates. U.S. Census Bureau, American Community Survey 5 year estimates (2014) Free and reduced price lunch (FRL) eligible (%) Annual measure of the percentage of students eligible for free or reduced pri ce lunch. Colorado Department of Education Student count (logged) Annual measure of the log of the total student enrollment count in October of each school year. Colorado Department of Education Diffusion of concurrent enrollment Number of high schools w ithin 5 miles offering Concurrent Enrollment. using data from the dependent variable and high school addresses Com munity college distance D istance in miles from the high school to the nearest community college. using data from the dependent variable and high school addresses Concentration of community colleges Number of community colleges within 10 miles of the high school. using data from the dependent variable and high school addresses C harter school Dummy variable (yes = 1; no = 0) indicating whether the school is a charter school. Colorado Department of Education PSEO participation Dummy variable (yes = 1; no = 0) indicating whether the high school previously offered Post Secondary Edu cation Options (PSEO). Colorado Department of Education District Setting Categorical variable: Denver Metro, Outlying City, Outlying Town, Remote, Urban Suburban. Colorado Department of Education

PAGE 58

47 Motivation to innovate. The hypothesized motivatio n to innovate is lower academic performance ratings. College matriculation rates measure the proportion of graduates who enroll in any college in the fall immediatel y following high school graduation, which relies on CDHE data. provides both numerical ratings and categorical ratings (performance, improvement, priority im provement and turnaround) in its annual performance review of school districts and schools. The high school level performance rating is an index that includes graduation rates, dropout rates, ACT scores and achievement and growth on the statewide standardi zed assessment. Both matriculation rates and school performance ratings are available on a yearly basis. The annual data, when included in the event history analysis, are lagged one year to avoid problems with ca us al inference. If the hypothesis is that lo wer performance motivates a school to adopt concurrent enrollment, the data for those indicators need to be from a time prior to the adoption year. Resources. Fiscal capacity and size are two types of resources important to the analysis of policy diffusion and innovation that are included in hypotheses 1. 2 and 1. 3, respectively. Per pupil funding at the school level is not available; only district level data is available. Including the district level data masks important funding variance at the high school level. Instead of including a district level variable, two school level measures are included in an attempt to capture the level of wealth and resources of individual high school communities. Previous studies have found a positive correlation between per p upil spending and the wealth of the local community (e.g. Augenblick, Myers & Anderson, 1997). T hus, as a measure of a household income for the neighborhood immediately surrounding the high school is used in this s tudy. Neighborhood

PAGE 59

48 tract number and county code. The ACS 5 year survey data contains the media n household income in the past 12 months (in 2014 Inflation adjusted dollars) for the five year period running from Jan. 1, 2010 Dec. 31, 2014. The 5 year survey option was used because it offers neighborhood level estimates. An additional measure inclu ded to capture the resources of a school is the proportion of students eligible for free or reduced price lunch (FRL). Annual data is available from CDE on the percentage of students qualifying for free or reduced price lunch; the data are lagged one year in the event history analysis. finance formula awards additional per pupil funds to districts for FRL students. Districts, in turn, use their own formulas to distribute state funds a s well as federal Title I dollars to schools most in need. While the FRL data and the median income indicator are measuring similar information, they are only moderately negatively correlated. There are instances where one measure may more accurately accou nt for the fiscal situation of a school than the other. There are wealthy communities, for example, that have high proportions of FRL students, possibly due to school choice patterns (e.g. median income in one Denver Metro area school is $108,627 and the p ercent FRL is 72.4%). In contrast, there are schools that have very low median incomes in the surrounding neighborhood but also have low FRL counts, most likely due to underreporting by families (e.g. median income in one re mote southwest Colorado school i s $39,476 and the percent FRL is 35.3%). Further, while the FRL indicator captures those schools that may receive additional funding support, not many high schools are Title I served, meaning districts more often direct the funds ear marked for high povert y schools to elementary and middle schools. Schools that are low income but not Title I served, or schools that serve middle income families who just fall short of meeting FRL eligibility will operate differently from schools that serve mostly high income families.

PAGE 60

49 One example of how such differences operationalize in terms of fiscal capacity, is through parent fundraising. According to an investigative report of parent fundraising at Colorado schools, fundraising levels vary dramatically by school and are correlated with the wealth of the community (Schimke, 2016). One school, for example, raised $14,400 through the one fundraising campaign in 2016; within the same school district, a school serving a lower income population raised only $300 (Schimke, 2016). While that is an extreme example, there are concrete differences in fundraising capabilities by school dependent upon the wealth of the parent population, as well as the wealth of the surrounding neighborhood (who are often asked to contribute to school fundraising campaigns). Even in districts that have school choice, controlling for the income levels of the neighborhood immediately surrounding the school building is still critical. Further, funds raised by the school are increas ingly used to support instructional and programmatic needs, as opposed to just supporting extracurricular activities (Schimke, 2016). Consequently, both median household income and FRL eligibility are included in an attempt to capture the different fiscal pressures at play at the school level. The second type of resource important to this study school size is measured by counting the number of students enrol led in the school during the annual October count period, which is the official method CDE relies on to assess pupil membership. Annual data is available for student count data, and the data are lagged one year in the event history analysis to avoid causal inference problems. External factors. Th e main external factors to be measured in hypotheses 1. 4 and 1. 5 are proximity to schools that have adopted concurrent enrollment and proximity to community colleges. The first factor is operationalized by calculating for each individual high school how many other high schools were already offering concurrent enrol lment within a five mile radius. The calculation was done for each school year included in the study so that the values could change as

PAGE 61

50 the program spread. The data are lagged one year in the event history analysis to ensure that the The second proximity factor was measured in two ways: 1) by calculating the distance in miles from the high school to the nearest community college, and 2) by calculating the number of community colleges within 10 miles of the high school. These measures adequately account for those high schools that are located in more populous, urban areas with access to multiple colleges. The author geocoded high school addresses using Texas A&M GeoS ervices. Once the longitude and geodetic distances 8 between individual high schools and community colleges and among the high schools themselves. The geocoded data was also used to match high schools with unique census tract numbers to then link the Colorado administrative dataset with the ACS dataset. Control variables. Lastly, two additional variables were included in the empirical models to control for possible confou nding factors. These include: 1) an indicator for whether a school was a charter schoo l or a traditional school, and 2 ) an indicator for whether the high school offered a different dual enrollment program prior to 2011. First, an indicator for wheth er a sc hool is a charter school is included to control for any potential confounding effects on the variables of interest in the case that charter schools act differently than traditional district run schools, particularly in terms of deciding programmatic offeri ngs. Second Berry & Berry (2007) suggest that diffusion models likely need to include variables that capture whether prior policies were in place that could impact the decision to adopt the policy currently at hand. In this case, there was a dual enrollme nt program available to districts in Colorado prior to the Concurrent Enrollment Programs Act passing in 2009. The program, known as Post Secondary Education Options (PSEO), began phasing out in 2009 and 8 Geodetic distances calculate the length of the shortest curve between two points along the surface of a mathematical model of the earth

PAGE 62

51 was fully phased out by 2011 12. If schools that had PSEO in place wanted to continue to offer similar opportunities, they had to make the transition to Concurrent Enrollment by 2011 12 (CDHE, 2014 ). It was not required that they make the transition, however, and PSEO schools could choose not to offer any d ual, or concurrent, enrollment courses once PSEO was phased out (CDHE, 2014 ). Although substantially different in nature and mechanism, PSEO participation likely is associated with concurrent enrollment adoption and so an indicator will be included to cont rol for if a high school had participated in the PSEO program. Policy Diffusion Methodology This section provides an overview of the two methods used to test hypotheses 1.1 through 1.5 : event history analysis and ordinary least squares (OLS) fixed effect s regression. Event history analysis was the method used for exploring the diffusion of concurrent enrollment across high schools in the state, while OLS fixed effects regression was used to explore any factors related to the intensity of program participa tion within high schools once concurrent enrollment was adopted. Event history a nalysis Event history analysis was conducted to explore the possible existence of explanatory relationships in the diffusion of concurrent enrollment programs among Colorado widely accepted as the most effective way to empirically assess the causes of policy innovation in 9 The method allows researchers to identify what factor s influence events over time. In this study, the event was the adoption of concurrent enrollment by high schools ( i ), and time was measured in discrete units of school years ( t ). The study period ran from the beginning of the 2010 11 school year until the end of the 2014 15 school year. The legislation that created the concurrent enrollment program passed in the spring 2009. In the 2009 2010 school year a small 9 EHA has also been applied to sub state entities, see e.g. Hoyman and Wein berg (2006) and Lubell et al. ( 2002 ).

PAGE 63

52 number of school piloted the program, and concurrent enrollment was officially operational statew ide in the 2010 11 school year. Thus, the window of this event history analysis captures the first five years of the implementation of concurrent enrollment in Colorado. Table 5 displays an overview of the number of high schools in the risk set in each sc hool year, the number of adoptions each year, and the survivor function. In an event history analysis, the survivor function expresses the probability that survival time T is equal to or greater than t, where t represents the actual survival time (Mills, 2 011). S( t t ) The output of S( t ) in this study is, therefore, simply a proportion of the observations that still had not adopted concurrent enrollment following school year t. Table 5 : Concurrent Enrollment Adoptions and Survivor Functions, by School Year School Year Number of Adoptions Cumulative Adoptions Risk Set Survivor Function Std. Error 95% Confidence Interval 2010 11 195 195 388 0.50 0.03 [0.45, 0.55] 2011 12 76 272 193 0.30 0.02 [0.26, 0.35] 2012 13 65 337 11 7 0.13 0.02 [0.10, 0.17] 2013 14 6 343 52 0.12 0.02 [0.09, 0.15] 2014 15 10 353 46 0.09 0.01 [0.07, 0.12] As Table 5 depicts, t he first year saw the largest number of adoptions with fifty percent of the high schools implementing the program that year. The diffusion of the program continued rapidly after that, with eighty seven percent of high schools offering concurrent enrollment by the third year. As the survivor function indicates, j ust 9 percent of high schools (n=36 ) had not adopted concurrent enro llment by the end of the study period in the 2014 15 school year. The high schools that did not adopt concurrent enrollment by the end of the study period are considered to be right censored. Event history analysis is preferred over other regression models for policy diffusion studies because it can account for both censored and non censored observations when producing

PAGE 64

53 estimates of the likelihood than an event will occur at a specified point in time (Mills, 2011; Mokher & McLendon, 2009). Following Berry a nd Berry (1990), many public affairs scholars used discrete time logit or probit models to perform event history analysis (Allison, 1984; Berry & Berry, 1990; Buckley & roach to testing diffusion theory with discrete event history analysis is straight forward, computationally discrete time models assume that the probability of policy adoption in one year is unrelated to the probability of adoption in previous years, when that may not be the case in actuality (Berry & Berry, 1992; Buckley & Westerland, 2004). The odds, for example, that a high school adopts concurrent enrollme nt in the first year of the study are likely different from the odds that a school adopts the program in the last year of the study when t he policy is more popular and eighty five percent of other high schools have already adopted it. As a result, scholars have turned to the Cox proportional hazards model, which allows for the probability of policy adoption to change over time while not having to specify the functional form (Buckley & Westerland, 2004; Jones & Branton, 2005, Mills, 2011). The semi parametri c nature of the model lends itself to being robust to different data, even if the author does not know the precise underlying shape of the probability distribution (Mills, 2011). The Cox method also permits the incorporation of time dependent variables. Fo r these reasons, this research design used Cox proportional hazards models to analyze the diffusion of concurrent enrollment. The Cox proportional hazards regression model relies on maximum partial likelihood estimation when computing hazard rates. The ha zard rate is the likelihood that the event of interest will occur in a specified unit of time given that the observation has survived any prior time periods. The Cox model estimates changes in hazard rates as a function of a set of covariates. While the

PAGE 65

54 mo del does not assume a particular shape of the baseline hazard rate, it does make a strong assumption that the ratio of hazard between any two observations is proportional across time (Box Steffensmeier and Jones 2004; Mills, 2011). If the proportional haza rds assumption is violated, the relative risk may be improperly estimated. To test the proportional hazards assumption, Schoenfeld residuals were estimated and ime dependency. The variable for district setting clearly violated the proportional hazards assumption. As a remedy, the Cox model was stratified on the variable, which essentially sets a separate baseline hazard function for each value of district setting Once the stratification was conducted, there were no further violations of the proportional hazards assumption in individual covariates or for the model as a whole. The final Cox model specification can be expressed as: h i ( t ) = h 0 ( t ) exp( x j ) where t he proportional hazard of high school i adopting concurrent enrollment in school year t is the result of an unspecified baseline hazard function h 0 ( t ) and a vector of the exponent of the varying cova riates ( x j ) in the model (Mills, 2011). The Efron approximation was used in estimating the Cox model, which more appropriately time period, than the typically u sed Breslow approximation. Additional model diagnostics that were conducted include the estimation and plotting of Cox Snell residuals to assess overall model adequacy and the plotting of martingale residuals to assess any nonlinearity in the covariates. As a result of analyzing the martingale residuals, two variables were log transformed to im prove linearity: median income and student count. The results of the Cox Snell residual analysis indicated that the full model specification was an adequate fit.

PAGE 66

55 Or dinary least squares fixed effects r egression A progression of ordinary least squares (OLS) regression models were conducted to investigate if any of the covariates included in the event history analysis model serve as predictors for how deeply a high sch ool implements concurrent enrollment. The dependent variable in the OLS models is a school level participation rate created by dividing the number of students within a high school taking at least one concurrent enrollment course by the total number of stud ents enrolled at the high school for each academic year within the study. The first OLS model includes the same variables used in the event history analysis as predictors The subsequent models include a series of time and unit fixed effects with and witho ut a lagged dependent variable. If the dependent variable and one or more independent variables trend in a direction over time, including time fixed effects in the regression model is often a necessary precaution (Wooldridge, 2006). If the dependent vari able and one of the key covariates both are trending upward, for example, the two time series processes may appear to be correlated when they are actually both trending for reasons related to factors unaccounted for in the model. E mploying dummy variables for each year of the study ( e xcluding the baseline year) controls for spurious trend relationships. If the time dummy variables end up being statistically significant, and the coefficients of other variables change in a meaningful way, that is e vidence of the need to include time fixed effects in the regression model (Wooldridge, 2006). After including time fixed effects, regression models were run with the addition of unit fixed effects to control for possible omitted variable bias. Fixed effect s wer e included, separately, at the district and school levels to control for district or school specific variation, which could have a confounding effect on the high school level model. High schools within districts are likely influenced by district leve l factors such as administrative capacity, the presence of a college preparatory culture, history of pursing partnerships within the district, or fiscal characteristics. Including district

PAGE 67

56 fixed effects results in a within regression analysis where the cha nge in covariates is only analyzed within each district. This could provide a good amount of control for any unobserved confounding factors, while also allowing for some across school variation (within districts that have more than one high school). Only including district fixed effects, however, still leaves room for doubt that the model is accounting for all unobserved variables. School level characteristics such as leadership, culture, academic systems (e.g. curriculum and instructional model), and teac her capacity likely also effect the implementation of concurrent enrollment. One disadvantage of s chool fixed effects is that they absorb nearly all is very little change in the time varying predictors within individual schools over the five years of the study (all time invariant predictors are dropped from a school fixed effects model). Thus, the author runs both a district and school level OLS fixed effects regression model which can be formerly expressed as: Y it = 0 + n 1 + F n 1 + X it + it where the dependent variable ( Y ) is the concurrent enrollment participation rate for each high school ( i ) in time period ( t ) and is a function of time fixed effects ( T n 1 ), school or district fixed effects ( F n 1 ), a vector of the coefficients of parameters ( ) for the time varying covariates ( X it ) and the error term ( it ). Two prominent issues with running OLS regression on time series data are serial correlation and heteroskedasticity. Serial correlation, or autocorrelation refers to the correlation of error terms among observations and is often present in time series data since the same unit is being measured in repeated time periods (Wooldridge, 2006). If the value of a covariate in one time period is related to its value in the previous time period, for example, then the error terms are likely to be correlated. Serial correlation does not bias the estimates but it does result in an underestimation of the

PAGE 68

57 standard errors, which in turn inflates the t statistic and leads to overestimation of statistical s ignificance. Heteroskedasticity or a violation of the OLS assumption that error variance is constant, likewise affects the calculation of standard errors and results in misleading claims of statistical significance (Wooldri dge, 2006). Diagnostics tests were run on the time series data set and found that both serial correlation and heteroskedasticity were present. To correct for both issues, robust school clustered standard errors are used in the regression models. Another is sue with running OLS models on this data set in particular stems from the functional form of the dependent variable, which is a percentage bound between 0 and 100. Using OLS simplifies the interpretation of the regression results, and post estimation analy sis found that only 20 of 1820 (1.10%) of predicted values fall outside of the 0 to 100 range (see Papke, 2005 for a similar approach). The regression models were also run using fractional logit to ensure that the results from OLS m odels are robust and not affected by misspecification error ( Papke & Wooldridge, 1996) The fractional logit models substantiated the statistical significance and direction of the OLS results. Dynamic panel data model Even though the OLS model s control for year and unit fixed ef fects there remains a need to investigate the effect of participation rate on the current year. enrollment participation rate for one year would be highly predi participation rate. Including a lagged dependent variable with fixed effects in the same OLS model, however, may lead to biased estimates as a result of correlation between the error terms and the covariates (Allison, 2009; Wo oldridge, 2010). Economists refer to models that include lagged dependent variables as dynamic panel data models, and there are several approaches that can be used for estimation (Allison, 2009; Williams, Allison & Moral Benito, 2016; Wooldridge, 2010). Th is study employs an approach that uses maximum likelihood estimation and allows for the inclusion of

PAGE 69

58 time invariant predictors while still retaining the benefits of fixed effects ( Williams, Allison & Moral Benito, 2016). Other approaches to modeling dynami c panel data, including traditional fixed effects methods and generalized method of moments (GMM), exclude time invariant predictors. As explained above, there is not much within school variation on the time varying predictors, so the inclusion of addition al covariates in the model is beneficial to understanding patterns in concurrent enrollment participation rates. The dynamic panel data model using maximum likelihood estimation was run with the dependent variable lagged one year. Full information maximum likelihood (FIML) was used to treat missing data. About 5 percent of schools are missing data on matriculation rates, and using FIML allows for those schools to remain in the estimation by using the data that is available for those schools rather than usin g list wise deletion and losing those observations altogether ( Arbuckle, 1996) Policy Evaluation Variables and Measures The follow ing section provides details on the dependent, explanatory and control variables used in the evaluation of concurrent enrollmen t. Table 6 su mmarizes the indicators used to operationalize the variables in the two policy evaluation hypotheses, and Table 7 provides a summary of variable descriptions. Table 6 Concept Measurement Summary: Policy Evaluation Hypo theses Construct s Indicators H 2 : High school students who participate in concurrent enrollment programs will have a greater probability of enrolling in higher education. College access Concurrent enrollment participation Immediate enrollment in college following high school graduation Concurrent enrollment partic ipation (y/n) Number of concurrent enrollment credit hours H 3 : First year college students who had participated in concurrent enrollment programs in high school will have greater academic succe ss and a higher probability of persisting than college students who did not participate. Academic success & persistence Concurrent enrollment participation Need for remedial education First year college grade point average (GPA) Fall to fall college persi stence C oncurrent enrollment partic ipation (y/n) Number of concurrent enrollment credit hours

PAGE 70

59 Table 7 Descriptions of Pre College Independent Variables and College Outcome V ariables Variable Description Concurrent enrollment p a rticipation Concurrent e nrollment credit hours Dummy variable (Took any concurrent enrollment course = 1) Categorical variable for number of concurrent enrollment credit hours attempted ( 0, 1 3, 3 6, 6 12 or 12+) Student academic characteristics ACT compo site score English language l earner Special e ducation Continuous variable (min=12; max=36) Dummy variable (Students designated as ELL=1) Dummy variable (Students designated as SPED = 1) Student and family background White African American Hispanic Asia n Other race Gender F ree or reduced price lunch (FRL) Dummy variable (white = 1) Dummy variable (African American = 1) Dummy variable (Hispanic = 1) Dummy variable (Asian = 1) Dummy variable (other race = 1) Dummy variable (male = 1) Dummy variable (FRL e ligible students = 1) School environment Rural/urban school district Dummy variable (Rural = 1) College outcomes College enrollment Remedial education need First year college grade point average (GPA) College Persistence Dummy variable (Enrolled in college anywhere in the fall immediately following high school graduation = 1) Dummy variable (Needed remedial education in at least on e math, reading or writing course = 1) Cumulative grade point average in the spring semester of a first year in college Dummy variable (If enrolled in year one and enrolled in year two of college anywhere = 1) Dependent variables The dependent variable in H ypothesis 2 is college enrollment, which was measured by considering those students who enrolled in college in the fall immediately following high school graduation. Students who enrolled in college anywhere at an in state, out of state, public or private institution are captured. This is a dichotomous variable; students who enrolled in college were coded as a 1. There are several dependent variab les that are operationalized from H ypothesis 3. First, need for remedial education in college is included as a measure of academic performance. The measure includes both students assessed as needing remediatio n and those enrolled in remedial courses who did not have an assessment score on file. This is a dichotomo us variable;

PAGE 71

60 students who need remedial education in college are coded as a 1. While college level concurrent enrollment courses require students to b e college by an appropriate assessment), students could need remedial education in a different content area. For example, a student may take a concurrent enrollment, college level literature course, but tha t student may require remediation in math. Further, some students only take career and technical education (CTE) concurrent enrollment courses that do not have the same academic prerequisites as core subject areas. Thus, remedial education is considered to be a worthwhile measure of academic preparation and success to include in the model. Second, student postsecondary academic success is assessed by considering the cumulative grade point average after the spring semester of the first year in college. Cou rse level data is only available through the state s administrative data collection and thus only students who enrolled in a public college in Colorado are captured in the calculation of remedial rates and grade point averages (CDHE, 2016 ). While th is is a limitation of the dataset, approximately 75 percent of high school graduates who enrolled in college did so at an in state, public institution and are captured in t he state s data system The third dependent variable that is measured in the second hypothesis is coll ege persistence, which is measured using a dummy variable indicating whether a student who enrolled in year one of college returned to enroll in year two of college (returned to any institution not just the original institution). Because enrollment data is available through both the state administrative system and the National Student Clearinghouse, the data for this variable includes all students who enrolled in college anywhere, not just those who enrolled in Colorado. Concurrent enrollment participation The key explanatory variable for both hypotheses is participation in concurrent enrollment. This study employed two measures of participation: 1) a dichotomous measure of participation, in which students who graduated high school having taken at

PAGE 72

61 least on e concurrent enrollment course were coded as a 1; and 2) a categorical measure based on the attempted number of concurrent enrollment credit hours. There are five categories for the credit hours measure: no credit hours, 1 3 credit hours, 3 6 credit hours, 6 12 credit hours and more than 12 credit hours. Zero credit hours is set as the baseline category in the analysis. The remaining four categories were selected after viewing the descriptive statistics and seein g natu ral breaks between each category that e quate roughly to quartiles. Demographic pre c ollege i ndependent v ariables The empirical an alysis included demographic and geographic control variables that, based on prior research, are thought to influence concurrent enrollment participation, college goi ng behavior and postsecondary outcomes. These measures include gender, high school free and reduced price lunch (FRL) status, special education (SPED), English Language Learner (ELL) status, rac e/ethnicity, and ACT scores. The data for FRL, SPED, and ELL students are reported by high schools to the Colorado Department of Education and indicate whether a high school graduate received free or reduced price lunch, was identified as special education, or was identified as ELL, respectively. Race/ethni city is self reported by students to schools and was measured here using dummy variables for African American students, Hispanic students, white students (the baseline group) includes American Indian/Alaskan Native or Hawaiian/ Pacific Islander students. 10 Gender, FRL, SPED, ELL and race/ethnicity fields are required components of the datasets schools submit to the C olorado Department of Education and there are no missing data. Lastly a dummy variable for rural schools was includ ed ( when school level effects were not utilized ) to capture school level differences attributable to geographic setting. 10 The categories of race/ethnicities used in this study are representative of the largest groupings of students and were necessary to accurately run the propensity score matching (PSM) analysis. Including separate, sma ller sub groups of students in the analysis did not change the end results but substantially reduced the likelihood of achieving non biased matches during the PSM analysis.

PAGE 73

62 Academic pre c ollege independent v ariable The composite ACT score was used as a proxy control for academic achievement. ACT scores we re an important variable to include because performance on the college entrance examination is highly correlated with college attendance. Further, ACT scores are strongly correlated with assessment scores from the statewide accountability tests administere d in 9 th and 10 th grades, indicating the scores are a good control for overall academic aptitude. 11 ACT subject scores (reading, writing, math and science) also were collected, but in the effort to achieve a more parsimonious model, composite scores were us ed. 12 During the period of this study, Colorado r equired all high school juniors to take the ACT since (ACT, Inc., 2009). The test wa s provided free to students, and one day of the academic year for juniors was devoted to taking the ACT. Nonetheless, some students opt ed out of taking the assessment. In addition, when the data from the ACT w ere matched with postsecondary data from the CDHE, some records w ere not matched successfully As a result, across the three high school graduating cohorts 15.0 percent had missing ACT scores. Multip le imputation was initially used as a treatment for the missing data, and the results did not alter significantly from those that are presented later in the chapter. Thus, l ist wise deletion was ultimately used to eliminate the obs ervations with missing AC T data for ease of interpretation. This method does have dis advantages because the population of students with missing data differed from the general population. T hose with missing data were less likely to attend college (2 4 .6%) than the students with ACT scores (63.2%), and they were less likely to participate in concurrent enrollment (6 .8 % compared to 15.4%). 11 The statewide assessments administered in 9 th and 10 th grades had slightly fewer missing values than the ACT assessment and could be used an alternative measure for academic aptitude. However, students taking concurrent enrollment are most likely to do so in the 11 th and 12 th grade of high school, and with the ACT being administered i n the 11 th grade and being designed to assess college readiness, it is a timelier source of performance data and, arguably, a more reliable and valid control variable for the hypotheses being tested here than the statewide grade level assessments. However, regression models were estimated using the 9 th and 10 th grade data, which confirmed the statistical significance and direction of the resulted presented here. 12 Models that were run with subject scores did not vary from the models run with the composite s core.

PAGE 74

63 However, nearly one third (32.7%) of students with missing ACT scores were attending an Alternative Education Campus (AEC). On average throughout t he study, only 2.3 percent of students were enrolled at an AEC. It is common for AECs to have missing data given that they serve a highly mobile and transient population of at risk students. AECs typically also have very low college matriculation rates (16 .4% compared to 57.7% at traditional high schools in 2014). It is likely that the AEC school status (and what that represents) is a bigger predictor of college outcomes than the ACT scores would be if the data were not missing. As explained later, school f ixed effects are used in the final model of the regression analysis to control for such unobserved bias. Further, as stated above, the results from regressions that were estimated following multiple imputation for the missing ACT scores confirmed the findi ngs presented below. Policy Evaluation M ethodology Multivariate, fixed effects regression and propensity score matching (PSM) were used to determine relationships between concurrent enrollment participation (the dichotomous measure) and the key dependent v ariables. The average treatment effect was calculated for both methods, which allows for comparisons of the techniques and provides a triangulation of the findings to assess how college outcomes for students who participate in concurrent enrollment are aff ected as compared to non participating students. When cons id ering how interactions between race/ethnicity and concurrent enrollment participation affect college outcomes, the author conducted the analysis solely with fixed effects regression given the sign ificant complications posed by including interaction terms in PSM ( Garrido et al., 2014 ; Imbens, 2000). The same methodological complications arise when using categorical predictor variables in PSM; thus, fixed effects regression was used when analyzing the effect of the number of concurre nt enrollment credits on college outcomes Overview R andomized controlled trials are the ideal method because both observed and unobserved factors are accounted for through the process of randomly assigning individua ls to the

PAGE 75

64 treatment group (Schneider, Carnoy, Kilpatrick, Schmidt & Shavelson, 2007; Singleton & Straits, 2010). In nonrandomized designs, selection bias when the participants in the treatment group differ systematically from those who did not participate in the treatment is a threat to internal validity (Singleton & Straits, 2010). In education research, as in the social sciences generally, it is often difficult to adhere to experimental designs due to practical, situational, or ethical considerations (Tit us, 2007; Winship & Morgan, 1999). As a result, researchers have developed and refined analytical techniques that attempt to create quasi experimental conditions that control for selection bias (Heckman, 1979; Rubin, 1974, 1997; Schneider et al., 2007; Win ship & Morgan, 1999). Both multivariate regression and PSM are techniques that fill that role. Multivariate regression allows the author to control for confounding or intervening variables that affect the relationship between the treatment and the outcomes (Singleton & Straits, 2010). PSM, developed by Rosenbaum and Rubin (1985), has a different focus and controls for those observed variables that affect whether an individual participates in the treatment or not. The theory behind PSM asserts that if assig nment to the treatment is driven solely by observable factors, then after the m atching i s c o n d u c t e d analysis of the treatment effect can proceed as if assignment was random (Rosenbaum & Rubin, 1985; Winship & Morgan, 1999). PSM begins by generating propensity scores based on observable variables that predict the likelihood that an individual will receive the treatment. Those scores are used to match individuals who received the treatment with similar individuals who did not receive the treatment. Through the ma tching process a control group is essentially created, which then allows the author to mimic experimental conditions, establish a counterfactual, and estimate the treatment effect (Guo & Fraser, 2010; Rosenbaum & Rubin, 1985; Rubin, 1997; Winship & Morgan, 1999). PSM has become popular amongst applied researchers dealing with observational data (Caliendo & Kopeinig, 2008; Dehejia & Wahba, 2002, 1999; Heckman, Ichimura & Todd, 1997). As

PAGE 76

65 the number of studies using PSM has increased, the debate around whethe r PSM is any better of a method than standard regression analysis persists (Brand & Halaby, 2006; Shadish & Steiner, 2010; Shadish, Clark & Steiner, 2008; Shah, Laupacis, Hux & Austin, 2005; Smith & Todd, 2005). Both PSM and multivariate regression analysi s condition only on observable variables and several studies have found that regression estimates tend to be similar to PSM estimates (Cook, Shadish, &Wong, 2008; Shadish & Steiner, 2010; Shah, Laupacis, Hux & Austin, 2005). Nonetheless, there are some adv antages to PSM. First, PSM estimations avoid bias caused by a misspecification of the functional matching assumes selection on observables, it does not assume linear selection as does covariate Second, there are several balancing tests that can be performed using PSM that provide information regarding the validity of causal inferences from the data set. A data set with cov ariates that are balanced between the treatment and control groups is not enough to fully eradicate selection bias concerns if there are u nobservable con founders present; but having a dataset with balanced observable covariates and areas of common support is a minimum requirement for making causal inferences (Shadish & Steiner, 2010). Regression analysis does not typically limit estimates to areas of common support, or to parameters of variables where both treatment and control observations exist ( Brand & Halaby, 2006). Third, there are diagnostic tests allowing the author to assess the sensitivity of the treatment effects to unobservable bias. These tests do not conclusively determine the level of unobserved bias, but provide ways for the resear ch to lend credence to findings in the case of theorized selection bias (Caliendo & Kopeinig, 2008). Propensity score m atching The pre college exogenous variables described in Table 2 were used to generate propensity scores, after which the results were e valuated to ensure there was an

PAGE 77

66 even distribution of propensity scores across treatment and comparison groups. PSM requires there to be an individual in the comparison group with a similar propensity score for each individual in the treatment group to make inferences from the results (Garrido et al., 2014). The sample is divided into a number of blocks that is sufficient enough to ensure equal mean propensity scores between treatment and comparison groups. Figure 2 displays the visual output from the result s and indicates that there was a satisfactory overlap in propensity scores between the treatment and control groups and an appropriate range of propensity scores. Figure 2 Distribution o f Propensity Score A cross Treatment and C omparison G roups Another important verification step is to check if the propensity scores were accurately specified by ensuring that the covariates are individually balanced in each block of the propensity score for both the treatment and comparison group s (Garrido et al., 2014; Imbens, 2004). Typically, initial specifications are not balanced and variables need to be dropped or transformed. In this study, several iterations of generating propensity scores were performed until balance was achieved. Initial ly, dummy variables for each high school (n=423) were included, but balance in the covariates could not be reached given the large number of high schools and the inability to produce

PAGE 78

67 matched samples on each covariate within each high school. In the final s pecification, school fixed effects were not included, but an indicator for rural or urban high schools was included to control for some school level effects. The race/ethnicity categories were redefined from seven categories to four categories (Hispanic, w hite, African American, other race) Additionally, the variable for ACT scores was converted from a continuous variable to a categorical variable. 13 After making those changes, balance in the covariates was satisfactorily achieved across the blocks using t tests. Some level of imbalance is expected and acceptable (Austin, 2009), but to further ensure balance had been achieved standardized differences were computed for the covariates across the blocks. The PSM literature has set forth an acceptable amount of imbalance as being maximum standardized differences of 10 to 25 percent (Garrido et al., 2014). The results of this study achieved standardized differences no larger than 2.5 percent. Thus, it was concluded that sufficient balance was achieved. Treated ind ividuals were matched with comparison individuals who had the most similar propensity score within a certain range of scores referred to as a caliper Keeping the matches within a range, or caliper, prevents poor matches from occurring. Here, a caliper of 0.2 of the standard deviation of the logit of the propensity score was used based on prior research that has found that range to produce optimal results (Austin 2011; Rosebaum & Rubin, 1985). One to one matching produces the least biased estimates since t he first match is always the strongest match, especially given the restriction of the caliper which prevents poor matches. One to many matching, on the other hand increases bias (from poorer matches) but decreases the variance of the estimates by includin g more counterfactual information for each treated subject (Caliendo & Kopienig, 2008). This study used both one to one mat ching and one to four matching, in which each 13 ACT scores were divided into the following categories: Low ACT score = 5 to 16; Medium Low ACT score = 17 to 20; Medium High ACT score = 21 to 24; High ACT score = 25 36. The analysis was also run using a continuous variable for ACT scores and the average treatment effect sizes did not vary, but there was more bias present during the initial stages of the PSM process due to the inability to find precise matches with the continuous variable.

PAGE 79

68 treatment unit i s matched to four control units. Other matching techniques such as kern el, radius and stratification matching were tested, but the one to one and one to four caliper matching with replacement produced the best balance in covariates across the treatment and control samples, as measur ed by standardized differences Regression a nalysis R egression models were estimated separately, for both concurrent enrollment independent v ariable s (dichotomous and credit hours) on the four college outcome dependent variables Initially, bivariate mode ls with the dependent variable and the key independent variable were run. Then demographic control variables (gender, FRL status, race/ethnicity) were added, followed by the addition of the controls for academic achievement (ACT composite score, ELL, SPED). The final models included all of the prev ious controls and added fixed effects for high school and graduation year ; this was considered the preferred model specification for both main research questions. Adding fixed effects helps alleviate concerns about omitted variables. In particular, school specific features such as the availability of college guidance counselors, the presence of a college preparatory culture, school leadership and school location (e.g. rural, urban, suburban) vary and could have a confounding effect on the model. Additionall y, adding in time fixed effects for each year of the study (minus the baseline year) controls for any spurious trend relationships occurring over the three year time period of the study. When investigating the effects of concurrent enrollment participatio n on minority and low income students, interaction terms were added to the fixed effects regression models by crossing each race/ethnicity variable and the FRL variable with concurrent enrollment participation (yes/no). The author used logistic regression for the dichotomous dependent variables (college enrollment, remedial education need, and persistence) and OLS regression for the single continuous dependent variable (first year GPA). Another way to investigate differential effects by disaggregated groups of students is to divide the sample by student group and run separate regression models. Using a

PAGE 80

69 pooled regression analysis with i nteraction te rms was selected as a more parsimonious and efficient way to approach analyzing the effects of concurrent enroll ment on different groups of students, but both approaches should, in theory, obtain the same outcomes. The findings chapter reports results from the average treatment effects calculation that was run after the teffects and margins suite of commands (Williams, 2012; Wooldridge, 2010) Data & Methods: Summary This chapter described the data and methods used to explore the research questions and associated hypotheses. Table 8 provides a summary of the methodological approache s for eac h research question. Table 8 : Methodological A pproaches with Associated Research Questions Analytic Methods Unit of Analysis Research Questions Event History Analysis High School RQ1: What factors influence whether high schools a dopt concurrent enrollment programs? Fixed Effects Regression & Dynamic Panel Data Model High School RQ1a: What factors influence the extent to which concurrent enrollment programs are utilized by students within high schools? Propensity Score Matching & Fixed Effects Regression Student (high school graduate) RQ2: How does participation in concurrent enrollment affect the college Student (college student) RQ3: How does high school participation in concurre nt enrollment affect the college performance and persistence of students? T he study begins with an exploration of factors that influence the adoption of concurrent enrollment programs among and within Colorado high schools using school level, administrat ive data in event hi story analysis and multivariate regression. The research then focuses on using student level data in propensity score matching and fixed effects regression to analyze the effects of participating in concurrent enrollment on college matr iculation and success. All components of the

PAGE 81

70 longitudinal data system, but the research questions are explored using different units of analysis and methods. This research design allows for a comprehensive inves tigation of the effects of olicy

PAGE 82

71 CHAPTER IV POLICY DIFFUSION FIN DINGS & DIS CU S SION This section presents findings from the descriptive statistics of the dataset and from the inferential statistics used to analyze the fi rst research question regarding what f actors le a d some schools to adopt concurrent enrollment more quickly and implement the program more intensely as compared to other schools The results from the event history analysis are presented following the descri ptive statistics; the OLS fixed effects regression results are discussed thereafter. Descriptive Statistics The diffusion of concurrent enrollment throughout Colorado high schools was rapid and nearly complete by the end of the five year study in 2015 Fi gure 3 provides a visual of program adoption by school district s and high schools during the first 5 years after the state legislation passed. Table s 9 and 10 contain descriptive st atistics for the data As Table 9 depicts, th e means of the time varying c ovariates did not alter dramatically from 2011 to 2015, with the exception being the diffusion of concurrent enrollment indicator, which captured the rapid increase in the number of high schools with concurrent enrollment (and the corresponding increase in the number of observations that had nearby high schools offering the program). All time varying covariates were lagged one year to ensure that they were being accurately captured as predictor variables. As an example, for the 2010 11 school year, the time dependent covariates reflect values from the 2009 10 school year. Table 10 displays the mean values of the covariates by year of concurrent enrollment adoption. In 2011, for example, 195 high schools adopted concurrent enrollment. The average matriculati on rate for those high schools for the year prior to adoption was 59.1 percent, which was higher than the average rates for high schools adopting later and nearly twice the mean for high

PAGE 83

72 schools that did not adopt concurrent enrollment during the study per iod (31.1%). The mean distance to the nearest community college was higher for schools adopting earlier (2011 2013) than for schools adopting la ter, or not adopting at all. School districts 2010 11 (Year 1 of study) High Schools 201 4 15 (Year 5 of study) Figure 3 Adoption of Concurrent Enrollment Programs from the 2010 11 School Year to the 2014 15 School Year, by School Districts and High Sc hools. The upper two maps indicate that school districts have at le ast one high school offering concurrent enrollment if the district boundary is shaded. In the bottom two maps, each dot represents an individual high school that offers concurrent enrollment. The variable for charter schools also reveals differences betwee n cohorts. Of the 195 schools adopting concurrent enrollment in 2011, only 4 percent were charter schools even though charters comprise about 13 percent of all high schools in the study. Charter schools are overrepresented, though, in the later adoption ye ars of 2014 and 2015. Additionally, of those high schools that did not adopt concurrent enrollment, 43 percent were charter schools. Another point of interest is the variable for prior participation in dual enrollment (PSEO), which indicates that 63

PAGE 84

73 percen t of the first cohort of adopters had previously had the PSEO program in place; the mean of that variable declines as the study progresses. Only 8 percent of the 37 schools that did not adopt concurrent enrollment had PSEO. Lastly, Denver metro area and ru ral schools (outlying city, outlying town and remote schools) tended to adopt concurrent enrollment on the earlier side of the study period, whereas urban suburban schools adopted in the later years. Of the non adopters, about half are located in the Denve r metro area. Tabl e 9 : Descriptive Statistics for All High Schools, Beginning and End of Study Time Varying Covariates M (SD) 2011 Min Max M (SD) 2015 Min Max College matriculation rate (%) 53.33 (21.93) 0 100 53.07 (20.01) 3. 8 92.3 School performance rating 63.50 (17.32) 25 100 68.74 (13.71) 27.6 100 Student count (logged) 5.76 (1.30) 1.39 8.16 5.76 (1.32) 1.61 8.16 Diffusion of concurrent enrollment (# of CE High Schools within 5 miles) 0.15 (.611) 0 5 4.56 (5.98) 0 26 Free/reduced price lunch students (%) 37.6 8 (23.08) 0 100 42.64 (23.12) 0 100 2011 2015 Fixed Covariates M (SD) Min Max Median household income (logged) 10.89 (.37 ) 9.68 11.81 Community college distan ce 1 1.72 (1 3.00 ) 0 67.11 Co ncentration of community colleges 1.91 (1.35) 1 6 Charter school 0.13 (. 33) 0 1 PSEO participation 0.52 (.50 ) 0 1 District Setting Denver Metro Area 0.34 (.47) 0 1 Outlying City or Town 0.22 (.42) 0 1 Remote 0.23 (.42) 0 1 Ur ban Suburban 0.21 (.40) 0 1 N 388 Standard Deviations (SD) in Parentheses

PAGE 85

74 Table 10 : Comparison of Variable Means, by High School Adoption Year If adoption year was: 2011 2012 2013 2014 2015 Did Not Adopt Time Varying Co variates Lagged M M ** College matriculation rate (%) 5 9 .06 53.15 52.89 54.50 33.53 31.14 School performance rating 65.36 65.95 70.76 77.68 56.43 62.01 Student count (logged) 5.96 5.95 5.72 5.60 5.13 4.67 Diffusion of concurrent enrollment (CE) (# of C E High Schools within 5 miles) 0.22 1.66 1.80 5.83 4.3 2.33 Free/reduced price lunch students (%) 38.06 3 6 .55 38.71 37.47 45.15 45.06 Fixed Covariates Fixed M (2011 2015) M ** Median household income (logged) 10.90 10.88 10.88 10.84 10.94 10.86 Com munit y college distance 11.65 13.06 1 2 .18 9.24 7.56 9.95 Concentration of community colleges 1.97 1.68 1.77 1.83 1.90 2.33 Charter school 0.04 0.08 0.22 0.33 0.40 0.43 PSEO participation 0.63 0.57 0.43 0.17 0.20 0.08 District Setting Denver Metro Are a 0.40 0.35 0.12 0.00 0.20 0.51 Outlying City 0.26 0.24 0.09 0.17 0. 4 0 0.19 Remote 0.24 0.21 0.31 0.17 0.00 0.16 Urban Suburban 0.10 0.21 0.48 0.67 0.40 0.14 N 195 76 65 6 10 36 Means for time varying covariates for high schools adopting concurrent e nrollment in 2011 through 2015 are lagged one year to reflect values for the year prior to adoption ** Mean s for time varying covariates for high schools that did not adopt are an average of values from 2011 2015 Figure 4 depicts the mean percentage of hi gh school students participating in concurrent enrollment over time, by adoption year cohorts. In the first year the concurrent enrollment program was fully operational (2010 2011), 195 high schools adopted concurrent enrollment, and, on average, about 12 percent of the students in those high schools participated in the program. By 2015, for those same 195 high schools, the mean participation rate increased to 18.2 percent. In 2012, 76 high schools adopted concurrent enrollment, and in those schools about 8 .1 percent of students, on average, took at least one concurrent enrollment course. Participation rates were similar at the 65 high schools that adopted the program in 2013. Both the 2012 and 2013 cohorts of

PAGE 86

75 adopters saw increases in average participation rates over time. Very few high schools adopted concurrent enrollment in 2014 and 2015, and the mean participation rates at those 16 high schools was quite small ranging from 1.9% to 2.4%. Figure 4 : Average Percentage of High S chool Students P articipating in Con current Enrollment (CE) within High S chools, b y Adoption Year C ohort from 2010 11 to 2014 15. If a high school adopted CE in 2011 (n=195) it represented in the dark shaded bar on the far left of the series. The dotted lin e represents the statewide average of the percent of students participating in CE during a school year. While all adoption year cohorts increased average participation rates over time, regardless of the year of adoption, there are apparent differences in t he level of participation by year of adoption with early adopters having higher starting levels of participation rates as compared to those adopting later. This indicates that the degree to which a school uses the program is dependent upon not only how ma ny years the program has been in place, but also whether or not the school was an early adopter. It also likely suggests that earl y adopters had prior dual enrollment programs in place. This pattern is also depicted in Figure 5, which presents descriptive statistics of participation rates in map form for key variables of interest: percent FRL, matriculation rates, PSEO participation and year of program adoption. 11.97% 13.54% 15.93% 16.64% 18.18% 8.07% 8.73% 8.24% 10.10% 8.55% 9.33% 10.98% 1.87% 2.40% 1.87% 0.00% 10.00% 20.00% 2010-11 2011-12 2012-13 2013-14 2014-15 Mean % of High School Students in CE School Year 2011 n=195 2012 n=76 2013 n=65 2014 n=6 2015 n=10 Year of Concurrent Enrollment Adoption State Mean % CE Participation

PAGE 87

76 Figure 5 Maps of Colorado high schools and Concurrent Enrollment (CE) participation rates by covariates of interest. Each circle repr esent a high school; the size of the circle reflects the percent of students participating in CE in 2014 15 school year. Shading in the top left map indicates % of students who were FRL eligible in the 2013 14 school year, with darker shading indicating hi gher FRL rates. The top right map reflects matriculation rates with high rates designated by darker shading; bottom left map displays schools that did not offer PSEO (dark blue) in contrast to those that did offer the prior dual enrollment program (light b lue); bottom right map depicts the year of program adoption with darker circles representing high schools who adopted earlier in study.

PAGE 88

77 In the Figure 5 maps, each dot represents a high school, with larger circles representing higher participation rates. The maps generally reveal there is variation across the state on these indicators. Some schools, for example, have high concurrent enrollment participation rates and high percentages of FRL students, while other schools with high participation rates have l ow FRL percentages. Interestingly, the PSEO map reveals several instances of high schools that have high participation rates that did not previously have PSEO in place (i.e., large, dark shaded circles). The year of adoption map appears to be dominated by darker shaded circles representative of early adopters that have higher participation rates, which was the pattern seen in Figure 4. Event History Analysis The findings of the multivariate event history analysis are presented in Table 11. The Cox proporti onal hazards model results are displayed as exponentiated coefficients, referred to as hazard ratios. A ratio that is greater than 1 indicates the high school is more likely to adopt concurrent enrollment as the values of the covariate increase. Hazard rat ios that are less than 1 indicate that a high school is less likely to adopt concurrent enrollment at higher values of the covariate. A ratio of 1 is interpreted as there being no association between the covariate and the hazard of adopting concurrent enro llment. Each hypothesis is tested separately with the control variables included, and then the full model is specified. The results are largely robust across the models. Three o f the five hypotheses contain statistically significant findings and both of th e control variables are statistically significant. Looking at the H ypothesis 1.1 the coefficient for matriculation rates is statistically significant at p<0.01, but the direction is the opposite of what was hypothesized. For every 1 percentage increase i n matriculation rates, the likelihood of adopting concurrent enrollme nt increases by an estimated 1.5 percent. I t was hypothesized based on the literature, that schools with lower college going rates might be quicker to adopt the program given its link to improving

PAGE 89

78 matriculation a While the direction is not what was hypothesized, it is also not surprising; it would be easy to justify having written the hypothesis in the opposite direction given that high schools with already established cult ures of college readiness would also be more likely to take advantage of the concurrent enrollment program. Table 11 : Cox Proportional Hazards Model Results *** p<0.01, ** p<0.05, p<0.1 Robust standard errors in parentheses Coefficients are expressed as hazard ratios All models are stratified by District Setting (Remote, Outlying City, Urban Suburban, Denver Metro Area) The second hypothesis concerning fiscal capacity has mixed findings. H igh schools with higher percentages of free and reduced p rice lunch eligible students have higher hazard ratio s. This statistically significant coefficient indicates that for every 1 percentage point increase in the Relevant Full Model Hypotheses Variable Model 1 Model 2 Model 3 Mod el 4 Model 5 Hypothesis 1 .1 Matriculation Rates (%) 1.015*** 1.015*** (0.00308) (0.003) School Performance 0.992* 0.998 Rating (%) (0.00424) (0.005) Hypothesis 1. 2 Median Household 1.125 1.035 Income (logged) (0.168) (0.159) FRL eligibility (%) 1.003 1.012*** (0.00264) (0.003) Hypothesis 1. 3 Student Count 1.316*** 1.224*** (logged) (0.0679) (0.082) Hypothesis 1. 4 Diffusion of 0.967 0.976 Concurrent Enrollment (0.026) (0.027) Hypothesis 1. 5 Community college 0.994 0.998 distance (0.005) (0.006) Community college 1.064 1.015 concentration (0.059) (0.056) Controls Charter school 0.437*** 0.393*** 0.447*** 0.435*** 0.469*** (0.0905) (0.0820) (0.0916) (0.088) (0.111) PSEO participa tion 1.571*** 1.860*** 1.667*** 1.833*** 1.494*** (0.171) (0.206) (0.183) (0.200) (0.164) Observations 727 782 789 790 722 Likelihood ratio 79.15 75.48 111.6 79.51 108.1 df 4 4 3 5 10 Prob>chi2 0.0319 0.0276 0.0368 0.0265 0.0395

PAGE 90

79 percent age of FRL elig ible students, the likelihood of adopting concurrent enrollment increases by an estimated 1.2 percent. 14 The coefficient on median household income is not significantly different from 1. H ypothesis 1.3 is statistically significant and in the predicted direc tion. L arger schools have a higher chance of adopting concurrent enrollment more quickly than smaller schools Hypotheses 1. 4 and 1. 5 were both rejected with nonsignificant result s on the diffusion of concurrent enrollment measure and the proximity to comm unity colleges measures The model indicated no association between having more neighbors offering concurrent enrollment and time to adoption. One reason for this finding may be how quickly the program diffused throughout the state that the impact of not find statistically significant effects of regional diffusion (Mokher & McLendon, 2009). The coefficients for distance to the nearest community college and number of nearby community colleges were also nonsignificant. This is not a surprising finding given the descriptive statistics, which reveal that high schools that were first to adopt concurrent enrollment were as likely to be located in rural areas where community colleges are few and far between as in urban areas where community colleges are more prolific. Lastly, as expected, both control variables were statistically significant. The chance of adopting concurrent enrollment was about 1.5 times grea ter for those high schools that previously offered the PSEO program compared to those schools that did not have PSEO in place. C harter schools were half as likely to take up the concurrent enrollment program as compared to traditional schools a s indicated in the results and in Figure 5 The h azard function in Figure 5 provides a visual display of the large magnitude of the effect of charter schools on the likelihood of adopting concurrent enrollment. As a means of comparison, the hazard function for matricu lation rates is also 14 Percent changes were calculated using the formula (e b 1)*100.

PAGE 91

80 displayed. While the coefficient was highly significant for the matriculation rates indicator the effect size was not as large as that of the charter school var iable, which is seen in Figure 5 Both graphs also indicate a relative pro portionality in the hazard functions, supporting the Cox proportional hazards assumption. Figure 6 Cox Proportional Hazards Regression S moothed Hazard Functions for Charter Schools and College Matriculation R ates. Non charte r schools (dotted line, left) and high schools with high college matriculation rates (75 th percentile dotted line, right) had a higher likelihood of adopting concurrent enrollment as compared to charter schools (solid line, left) and high schools with lo w matriculation rates (25 th percentile solid line, right), respectively. OLS Fixed Effects Regression Analysis The findings of the OLS regression a nalysis are presented in Table 12 and Figure 6 The first model consists of the same variables that were i ncluded in the event history analysis (EHA) C ox proportional hazards regression model. While the C ox regression was stratified by district setting, the OLS regression version includes district setting categories as dummy variables, with Denver Metro Area s erving as the baseline, or reference, group. The second model builds off of the first and also includes time fixed effects by adding dummy variables for school year, with the 2010 11 school year serving as the baseline group. The third model expands on Mod el 2 by including district fixed effects. The d istrict fixed effects are absorbed due to the large number of districts (n=178); consequently, coefficients are not displayed.

PAGE 92

81 Table 12 Predictors of Student Participation Rates in C oncurrent Enrollment (CE) ( Model 1) ( Model 2) ( Model 3) (Model 4) Variable EHA Model Year Fixed Effects (FE) Year FE & District FE b Year FE & School FE c College matriculation rates (%) 0.187*** 0.197*** 0.095** 0.017 (0.032) (0.032) (0.038) (0.037) School performance rating (%) 0.064 0.029 0.037 0.067** (0.046) (0.048) (0.053) (0.030) Median Household Income (logged) 0.474 0.834 1.149 (1.341) (1.342) (1.782) FRL eligibility (%) 0.141*** 0.131*** 0.096** 0.019 (0.036) (0.037) (0.043) (0.050) Student Count (logged) 2.840*** 2.950*** 1.173 5.473* (0.577) (0.576) (0.754) (3.263) Diffusion of CE 0.155* 0.114 0.224** 0.173** (0.092) (0.117) (0.101) (0.081) Community college distance 0.060 0.058 0.006 (0.062) (0.062) (0.1 59) Community college concentration 0.624* 0.080 0.625 (0.359) (0.376) (0.578) Charter school 0.589 0.444 0.716 (2.553) (2.550) (3.239) PSEO participation 4.695*** 4.795*** 2.193 (1.024) (1.020) (1.355) Outlying a 2.134 1.268 (1.51 9) (1.527) Remote 0.999 0.305 (2.188) (2.191) Urban Suburban 3.025** 3.428*** (1.260) (1.258) 2011 12 2.268*** 2.417*** 2.700*** (0.589) (0.581) (0.547) 2012 13 5.013*** 5.284*** 5.654*** (0.774) (0.804) (0.810) 2013 14 5.384* ** 5.799*** 6.317*** (0.920) (0.919) (0.885) 2014 15 6.078*** 6.452*** 7.257*** (0.953) (0.968) (0.893) Constant 10.238 13.397 14.414 43.278** (14.862) (14.945) (18.662) (18.330) Observations 1,828 1,828 1,828 1,828 Adj. R squared 0.219 0.239 0.538 0.753 *** p<0.01, ** p<0.05, p<0.10 a District Setting baseline group is Denver Metro Area b District fixed effects included but not reported; 178 categories absorbed c School fixed effects included but not reported; 383 categories absorbed Ro bust school clustered standard errors in parentheses for all models

PAGE 93

82 The fourth, and final, model includes school rather than district, fixed effects, which are also absorbed (n=383). Including school fixed effects omits several indicators that do not vary over time (i.e., during the study) within high schools, including median household income, c ommunity college distance c ommunity college concentration c harter schoo l and PSEO participation. S chool fixed effects are important to include, however, since th e treatment is at the school level and characteristics specific to schools such as leadership, culture, academic syste m, and teacher capacity likely a ffect the implementation of concurrent enrollment. None of the covariates remain statistically significan t throughout all models. College m atriculation rates and FRL eligibility are statistically significant in Models 1 through 3 but are not statistically significant at p<0.1 when school level fixed effects are added (Model 4), as a result of the fixed effect s controlling for unobserved variables and absorbing across school and across time variation. Regarding the college matriculation variable, while it is lagged one year, there is still a threat of endogeneity if an increase in concurrent enrollment partici pation rates leads to an increase in matriculation rates. The variable for the size of the high school is statistically significant at p<0.10 in M odel 4. The coefficient indicates that a 10 percent increase in the count of students in a high school leads t o a 0.5 2 percent decrease in the participation rate, meaning that smaller high schools are more likely to have a larger share of their students participating in concurrent enrollment, although the effect size in this model is substantively small. To put i t in more relative terms, an increase of 60 stud ents at the average high school (going from the mean of 592 students to 652 students) is associated with approximately 3 fewer students participating in concurrent enrollment (decreasing from about 72 student s participating to 69). The charter school and PSEO covariates are not statistically significant in Model 3, which includes district fixed effects; however, Model 4 is the preferred specification and b oth variables are

PAGE 94

83 omitted in that model because there i s no variance within a school on those indicators. Therefore, it is unclear from these fixed effects models what the effect of being a charter school or having had PSEO in place prior to 2009 is on participation rates. The diffusion of concurrent enrollmen t measure, which captures the number of neighboring high schools offering the program, is statistically significant in both the district and school fixed effects models ( M odel 3 and Model 4). The coefficient from Model 4 can be interpreted as an increase o f 1 additional neighboring high school offering concurrent enrollment reduces the share o f students participating by 0.17 percent. While the result is not in the predicted direction, the effect size is very small in substantive terms. I n the first two mod els, before district and school fixed effects were added, a variable for district setting was included. T he coefficient for urban suburban high schools was statistically significant at p<0.1, indicating those schools were associated with a 3.4 percentage p oint lower participation rate when compared to Denver metro area high schools. The descriptive statistics showed urban suburban high schools were slower to adopt concurrent enrollment, and it appe ars that those schools also experience a lower overall parti cipation rate once they do adopt con current enrollment. Lastly, t he year fixed effects also reflect what was described in the descriptive statistics; average school wide participation rates in concurrent e nrollment have increased over time. The 2014 15 sc h ool year is associated with a 7.3 percentage point increase in the mean participation rate as compared to the 2010 11 school year. Dynamic Panel Data Model Table 13 displays the results from the dynamic panel data model, which includes a lagged value of t he dependent variable and school and year fixed effects, and uses the maximum likelihood estimator to produce estimates. The model was estimated with the dep endent variable lagged one year, which removes one year (2011) from the dataset. T he coefficient fo r the lagged effect

PAGE 95

84 indicates that a high school will see a on rate for every 1 percent increase Table 13 Dynamic Panel Data Model using M aximum L i kelihood for Concurrent Enrollment (CE) Participation Rates in High Schools Variable Model 1 CE Participation Rate t 1 0.75 0 *** (0.054 ) Matriculation Rates (%) 0.026 (0.0 37 ) School Performance Rating (%) 0.04 7 (1 .17 1 ) Median Household Income (logged) 4. 63 3 ** (2.334 ) FRL eligibility (%) 0.018 (0.0 3 5 ) Student Count (logged) 7.06 0 *** (1.528 ) Diffusion of CE 0.052 (0.153 ) Community college distance 0.324 *** (0.099 ) Community college concentration 0.34 6 (0. 674) Charter sc hool 6.22 4 *** (1 895 ) PSEO participation 5.162 *** (1. 352 ) Observations 388 Likelihood ratio 1 05 .11 df 65 Prob>chi2 0.00 1 2 *** p<0.01, ** p<0.05, p<0.10 Robust school clustered standard errors in parentheses As the results for the other variab les in Table 13 show, even with the large effect size of the lagged dependent variable, other predictor variables remained statistically significant, including student count, which was statistically significant in the school fixed effects model. In the dyn amic panel data model, however, the student count coefficient is much higher than in the fixed effects model; here, a one percent increase in student count is associated with a 7.1 percent decrease in concurrent enrollment participation rates, all else hel d constant. Other differences between the

PAGE 96

85 dynamic panel data model and the district fixed effects model are the non significance of the school performance rating and the diffusion of concurrent enrollment variables. An important advantage of the dynamic p anel data model is that it allows for the effects of time invariant indic a tors to be observed. The following variables were omitted from the school fixed effects model and are statistically significant (p<0.01) in the dynamic panel data model: median house hold income, distance to the nearest community college, charter school, and PSEO. A one percent increase in median household income is a ssociated with approximately a 4.6 percent increase in participation rates within a high school, accounting for the prior and holding all else constant. Adding an additional mile to the distance from the nearest community college results in a small decrease of 0.32 percent in the participation rate, while being a charter school is associated with a l arge decrease of 6 .2 percent in the share of students in concurrent enrollment as compared to non charter schools. High schools that participated in PSEO before concurrent enrollment see a 5.2 percent increase in participation rates as compared to those scho ols that did not have PSEO, all else equal. Taken altogether the results provide moderate support for H ypotheses 1.1 through 1.3, and less support for Hypotheses 1.4 and 1.5 (see Table 14 ). The following section provides further discussion of the results and draws conclusions.

PAGE 97

86 Table 14 : Summary of Statistically Significant Results across Methods and Hypotheses Method Event History Analysis (Cox PH Model) Fixed Effects Analysis (OLS with School & Year FE) Dynamic Data Panel Model (Lagged DV & School/Year FE) Dependent Variable HS Adopted CE (Y/N) CE Participation Rate CE Participation Rate Hypotheses Variable Results B (SE) 1. 1: L ower academic achievement levels are more likely to adopt CE and have hi gher student participation rates. College matriculation rates 1.015*** (0.003) Non significant Non significant School Performance Rating Non significant 0.067** (0.030) Non significant 1 2 : G reater fiscal capacity are more like ly to adopt CE and h ave higher student participation rates. Median household income Non significant Omitted 4. 63 3 ** (2.334 ) % FRL Eligibility 1.012*** (0.003) Non significant Non significant 1. 3: More students have a greater likelihood of adopt ing CE but the share of stud ents participating may be lower. Student count 1.224*** (0.082) 5.473* (3.263) 7.06 0 *** (1.528 ) 1. 4: O ther schools nearby that have already adopted CE are more likely to offer the program and have higher student participation rates. Number of high sc hools w/in 5 miles offering CE Non significant 0.173** (0.754 ) Non significant 5: G reater proximity to a community college will be more likely to adopt CE and have higher student participation rates. Distance from CC Non significant Omitted 0.324 *** (0 .099 ) # of community colleges w/in 10 miles Non significant Omitted Non significant Control variables Charter 0.469*** (0.111) Omitted 6.2 2 4 *** (1 895 ) PSEO 1.494*** (0.164) Omitted 5.162 *** (1. 352 ) *** p<0.01, ** p<0.05, p<0.10

PAGE 98

87 C onclusion & Discussion This study sought to understand influences on the adoption and utilization of concurrent enrollment among Colorado high schools using a panel data set spanning the first five years following the enactment of the Concurrent Enrollment Programs Act of 2009. This pursuit was particularly compelling given that recent research has shown a positive association between nearly fully diffused within five year s. The goal was to use ev ent history and regression analysis to see if there were any best practices that could be gleaned from the Colorado case study and applied to other states trying to scale up similar programs. There were several important findings t hat resulted from this research, although questions remain. Academic achievement levels The first hypothesis posited that schools with lower academic achievement levels would be more motivated to adopt concurrent enrollment, which is seen as a strat egy for boosting achievement and postsecondary readiness. The variable for school performance rating, which encompasses information on student academic achievement and growth, is nonsignificant across models with the exception of the school fixed effects m odel in which the effect size is very small The college matriculation rate variable is nonsignificant in the participation rate models but is statistically significant in the event history analysis model The directio n of the coefficient is not what was hypothesized, but as noted previously, the finding is not surprising high schools with already established cultures of college readiness are likely to take advantage of an additional college access program. Given the eventual, widespread diffusion of the program, it is evident that even high schools with historically low college going rates have implemented concurrent enrollment as well

PAGE 99

88 Further, as Table 14 summarize s schools with higher percentages of FRL students were also more likely to adopt concurr ent enrollment. While that variable is technically included as part of the school resources hypothesis, there is a well known i nverse relationship between income status and college going rates. Thus, the two findings are seemingly contradictory at first gl ance. U pon reflecting on the results further though, it is understandable that both findings occurr ed simultaneously. It seems natural that there would be a higher propensity for the adoption of college preparatory programs in schools where a majorit y of students matriculate to college Such schools already have a colle ge going culture established, and likely have p arents who push for the inclusion of opportunities that would advance their Concurrent En rollment Programs Act was specifically passed to expand access to those students who typically had not been included that is, low income and minority students. Given the rapid diffusion of the program, and the positive, statistically significant coefficien t on the FRL variable, one could conclude that high schools have taken up the opportunity to expand access to new groups of students as the law intended. Future research could seek to further untangle these effects, and also explore to what degree those sc hools that had initially low college going rates and high FRL rates are seeing positive gains in their postsecondary outcomes as a result of offering concurrent enrollment. Fiscal Capacity From the results for Hypothesis 1.2, there is some evidence that f iscal capacity relates to concurrent enrollment participation. The EHA results show that schools with higher proportions of low income students were quicker t o adopt concurrent enrollment, but the median household income was not statistically significant When considering effects on participation rates, the opposite case is seen. The FRL variable is not statistically significant, but the median household income variable is statistically significant and substantively large in the lagged dependent variable mo del. The results of that model suggest that high schools in wealthier communities have higher

PAGE 100

89 levels of participation in concurrent enrollment. These somewhat conflicting results mirror the previous discussion. While schools that have lower income populati ons were quick to adopt the program, perhaps motivated by the intent of the program to expand access, it appears that those schools with more advantaged populations (likely with strong college going cultures and higher numbers of eligible students ) have higher shares of students actually partici pating in courses. Size of High School School size was a sta tistically significant variable in both the event history and the regression analyses Larger high schools were quicker to adopt concurrent enrollment, leading one to conclu de that organizational resources matter in the initial adoption of a program, as hypothesized The hypothesis regarding school size also posited that the number of a students in a school may be inversely related to the share of students participating in concurrent enrollment. T he results confirmed that supposition, with the regression models indicating that smaller high schools were more likely to have high proportions of students participating in concurrent enrollment than large high schools. That is not to say there are n ot examples of large high schools that have high percentages of students in concurrent enrollment, but rather, on average and holding all else constant, it is more likely for smaller schools to more intensely use the program. As discussed earlier, this can be due to the fact that when a small high school offers a concurrent enrollment course, for example a math class for seniors, that class comprises a large percentage of its overall enrollment, whereas offering one class at a large high school will only co nsist of a small share of the overall student body. There is also the consideration that for smaller schools, many located in remote areas, concurrent enrollment provides access to content the school is not able to deliver on its own. A high school, for i nstance, can go through a community college and use its instructors to offer college level chemistry, whereas without concurrent enrollment that school would not be able to offer the

PAGE 101

90 course at all. Lastly, in smaller schools with fewer administrative resou rces, once a program has begun in the school it may be the case that the program is more highly utilized because there are fewer competing opportunities. In large high schools, it is often the case that students have the choice of various college readiness and credit accrual programs; for example, in more resourced schools students often have the choice of concurrent enrollment or advanced placement. In smaller schools, it may be the case that only one college readiness program is offered, and thus a higher Type of School: Charters vs. District run Schools Charter schools have slower rates of adopting concurrent enrollment and lower overall participation rates. After investigating further t he results of the charter control variable, several explanations surfaced as to why charters were half as likely to adopt concurrent enrollment. Under the Concurrent Enrollment Programs Act, charter schools are considered to be Local Education Agencies (LE As) and, as such, are treated the same as school districts (also referred to as LEAs). The legislation permits charter schools that are authorized by their local school district to either enter into their own memorandum of understanding (MOU) with communit y colleges, or to be a part of he Colorado Department of urrent enrollment administrator, there could be instances where the district chooses to exclude charters from th eir MOU (M. Camacho Liu, personal communication, October 8, 2016) Setting up an MOU with a college requires negotiating financial arrangements, course offerings, teacher and faculty assignments, and other logistical details. Charter schools, many of which have low er administrative capacity, may conclude that the process is too cumbersome and for e go participating in concurrent enrollment. districts handle the financial tr ansactions on behalf of its schools. One large urban school district, for

PAGE 102

91 example, covers the tuition payments to colleges for students (at district runs schools) who are attending concurrent enrollment courses on the college campus. Sending students to a college campus for courses is typically a cost that would have to be absorbed by the high school O ffering can be less costly or cost free, but charter schools may have limited instructional staff with the necessary academic credential s to t each concurrent enrollment courses Thus, if the district is picking up the off campus tuiti on tab it is a significant benefit, an d it appears that the benefit may not be extended to charter schools. Lastly, an explanation for the results could also be that charter schools, on average, prefer to pursue other college readiness programs that are often perceived to be more elite or rigorous (e.g. Advance d Placement or International Baccalaureate). Some charter schools may prefer to partner with more selective, four year institutions instead of community colleges, and while that is permissib le under concurrent enrollment, it is more costly to do so. Prior Dual Enrollment (PSEO) Offerings T he statistically significant effect of the control variable for a prior dual enrollment program PSEO may indicate that some high schools are more dispo sed to pursuing postsecondary readiness programs. In many cases, there were likely relationships that had been established with community colleges through the PSEO program that made implementing concurrent enrollment an easy transition. Besides any par tnership advantages PSEO high schools had, though, there could also be an established college going culture in place at those high schools that supports the implementation of new college preparation programs when they become available. In the diffusion lit erature, Berry and Berry (2007) refer to that as a softening of the environment when one innovation takes place, it makes it easier for subsequent innovations to occur. Prior research on the importance of establishing college going cultures in high schools would also support the claim that states putting resources into promoting college access programs may see an increasing return on

PAGE 103

92 that investment (Hoffman, Vargas & Santos, 2008a ; Roderick, Nagaoka, Coca & Moeller, 2008). That is, once one program is in p lace it may open the door to continuous improvement and increase the dedication to ensuring that all students have access to high quality college preparation curriculum and supports, which should be at the forefront of secondary school reform.

PAGE 104

93 C HAPTER V P OLICY EVALUATION FINDINGS The findings from the inferential statistics used to analyze the final two research questions are divided into three sections The first section investiga tes the effects of concurrent enrollment participation on individual outcomes related to college access and success by focusing on the dichotomous measure of participation The results from the propensity score matching (PSM) are presented first, followed by the fixed effects regr ession results. A comparison of the results from the different methods is provided The second section take the analysis further by considering the effects of concurrent enrollment participation on the outcomes of low income and mi nority students, specifically. Interaction terms are added to the regression analysis to explore relationships between race/ethnicity, income and concurrent enrollment participation. The third section presents the findings from the analysis conducted using the categorical credit hours variable. Including different levels of concurrent enrollment credit hours in the research to understand the differential im pacts of treatment intensity. Descriptive Statistics Table 15 provides descriptive statistics for the overall sample, for those students who participated in concurrent enrollment and for those who did not participate. Overall, 15.40 percent of high school students graduating in 2011 2012 or 2013 particip ated in concurrent enrollment. Concurrent enrollment students have a higher mean college matriculation rate, college GPA and college persistence rate than students who did not participate. They also have a lower average remedial education rate than non participating students.

PAGE 105

94 Table 15 : Descriptive Statistics for Overall Sample and by Concurrent Enrollment (CE) Participation Treatment Variables Overall Sample M (SD) Min Max C E Students M (SD) Non CE students M (SD) CE participation (Y/N) 0.15 (0.36) 0 1 1 0 CE attempted hours 1.46 (4.80) 0 79.5 9.45 (8.61) 0 Outcome Variables College matriculation 0.57 (0.50) 0 1 0.73 (0.44) 0.56 (0.49) Remedial education 0.35 (0.47) 0 1 0 28 (0.45) 0.36 (0.48) First year GPA 2.75 (0 .86 ) 0 4 2.81 (0.83) 2.74 (0.87) College persistence 0 .81 (0.39) 0 1 0.83 (0.38) 0.81 (0.39) Additional Covariates FRL 0.26 (.44) 0 1 0.27 (0.44) 0.26 (0.44) ELL 0.05 (0.22) 0 1 0.03 (0.17) 0.06 (0.23) SPED 0.07 (0.25) 0 1 0.03 (0.17) 0.07 (0.26) ACT Composite Score 20.64 (5.25) 12 36 21.30 (4.65) 20.52 (5.34) White 0.65 (0.48) 0 1 0.66 (0.47) 0.65 (0.48) Hispanic 0 .26 (0.44) 0 1 0.25 (0.43) 0.26 (0.44) Black 0.05 (0.22) 0 1 0.05 (0.21) 0.05 (0.22) Asian 0.02 (0.15) 0 1 0.02 (0.14) 0.02 (0.15) Other Race 0.01 (0.11) 0 1 0.01 (0.09) 0.01 (0.11) Female 0.50 (.50) 0 1 0.55 (0.50) 0.49 (0.50) Rural 0.19 (0.39) 0 1 0.30 (.46) 0.17 (0.38) N 2011 Graduation Year N 2012 Graduation Ye ar N 2013 Graduation Year N Total 43,716 43,688 44,488 131 892 3,957 7,101 9,255 20 313 39,759 36,587 35,233 111 579 The percentage of students eligible for free or reduced price lunch (FRL), a proxy for income, remains nearly the same across the d isplayed groups. Similarly, the racial/ethnic composition of students who took concurrent enrollment closely mirrors the composition of the population as a

PAGE 106

95 whole. Rural students are overrepresented in the concurrent enrollment population (30%) in compariso n to the population mean (19%) while English language learners (ELL) and special education (SPED) students are underrepresented amongst concurrent enrollment participants. Figure 7 provides an additional display of descriptive statistics, breaking out co ncurrent enrollment participation rates for each graduation year by gender and race/ethnicity. Participation rates have been higher for female students than male students during each year of the study. Hispanic students had higher participation rates than white students in 2011 and 2012 and were one percentage point below them in 2013. All groups have seen increased participation rates each year of the study. Figure 7 Participation in Concurrent Enrollment, by Graduation Year, G ender and Race/Ethnicity Effects of Concurrent Enrollment Participation on College Outcomes This section presents findings from the research conducted using the dichotomous measure of concurrent enrollment participation in estimating effects on the four c ollege outcomes. The results from the propensity score matching (PSM) are presented first, followed by the fixed effects regr ession results. Lastly, a comparison of the results from the different methods is provided P ropensity S core Matching Findings As d escribed in the methods section, PSM involves multiple steps that must be taken before treatment effects can be estimated. An important first step is the matching process wherein propensity scores are generated based on a set of covariates, and concurrent enrollment students (the treatment group) are matched with students who have similar propensity scores but did not 0.0% 5.0% 10.0% 15.0% 20.0% 25.0% Female Male AfricanAmerican Hispanic Other Race White 2011 2012 2013

PAGE 107

96 participate in concurrent enrollment (the control group). Thus, the first set of findings from the PSM analysis considers how well the matchi ng process performed in terms of leveling biases that were present in the unmatched sample. Matching results Overall, the sample mean bias was reduced from 9.3 to 0.0 through the PSM process which, practically speaking, means that certain groups of stud ents were over or under represented in the unmatched treatment group, but after the matching process occurred, the treatment and control groups are equally balanced Figure 8 presents the standardized bias for each individual covariate, which is a measure of the difference in means between treated and matched control groups before and after the matching occurs. 15 The graphic displays the significant reduction of bias amongst all covariates in the matched and comparison groups. M ore specifically, what Figure 8 coveys is that in the unmatched sample ELL, male, and SPED students and those with very low ACT scores (<16) are underrepresented (the dots are to the left of the vertical line that indicates there is no difference in sample means between groups). On th e other hand, r ural students, students eligible for free and reduced price lunch (FRL), and Hispanic students were all overrepresented for their populations in the concurrent enrollment treatment group pri or to matching (the dots are to the right of the ze ro bias line). Prior research indicates that being of low income status, of minority status, or from a rural school is negatively associated with college enrollment (Fry, 2011; Kahlenberg, 2004; Terenzini, Cabrera, & Bernal, 2001). If selection biases in t his case assuming students who opt in to concurrent enrollment were the type of students who would be more likely to go to college are present in this study, these descriptive statistics mean there is at least some evidence that the bias of how students ar e selecting enroll ment into concurrent programs is not a clear case of the most 15 More technically, the standardized bias is calculated as a percentage of the square root of the average of sample variances in both groups (see, e.g., Caliendo & Kopeinig 2008).

PAGE 108

97 academically proficient students opting in One plausible explanation could be that while low income, Hispanic and rural students are less likely to attend college, of those wh o do plan to attend they will be more inclined to select into concurrent enrollment programs because they are able to earn college credits for free. Another explanation is that concurrent enrollment is sufficiently different from programs such as Advance d Placement in that it targets students who are of more modest academic standing and come from diverse backgrounds. Therefore, the selection bias problems you would see in a study of high achieving students from higher income backgrounds participating in vol untary college readiness programs may not apply at least as fully in this case because the original treatment population appears to be more representative of students who are typically less likely to attend college. Regardless, as Figure 8 depicts, these i mbalances in the original sample are eradicated the matched sample. Figure 8 Standardized bias differences (%) across all covariates in original and matched samples Treatment effects Once the matching process was performed, th e author estimated average treatment effects (ATE) for each outcome model. The ATE is the difference between the average outcomes of those who participated in concurrent enrollment and those who did not

PAGE 109

98 participate. Table 16 displays ATEs from the PSM anal ysis both 1:1 matching and 1:4 matching for the four college outcome models. The results from the two matching models are very similar. Table 16 Propensity Score Matching Average Treatment E ffects Outcome Matched 1:1 caliper wi th replacement Matched 1:4 caliper with replacement College Enrollment 10.16 *** (.00 4 ) 10.27*** (.003) Remedial Education 7.54 *** (.004) 7.57*** (.004) First year college GPA .08*** (.009) .08*** (.009) Persistence 2.17 *** (.003) 2.15%*** (.003 ) Robust standard errors in parentheses *** p<0.01, ** p<0.05, p<0.10 The PSM results using 1:1 matching suggest that taking concurrent enrollment courses increases the probability of college matriculation by 10.16 percentage points and reduces the cha nce of needing remedial education in the first year of college by 7.5 4 percentage points When considering the sample means, those effect sizes are substantively large the predicted probability of attending college increases from 57 percent to over 67 perc ent for concurrent enrollment students, while their probability of needing remedial education decreases from 35 percent to 27.5 percent. Students who took concurrent enrollment in high school have first year college GPAs that are on average, 0.08 points h igher than their peers who did not participate in the program. Substantively, the effect size is on the small side; students with a mean GPA of 2.75 would see their GPA increase to 2.83, for example. The effect of concurrent enrollment participation on per sistence rates is also statistically significant but substantively small ; the probability that freshmen college students will retur n for a second year of college is on average, 2.1 percentage points higher for concurrent enrollment participants than for n on

PAGE 110

99 probability of returning for a secondary year of college increases from 81 percent to 83 percent based on the sample mean as a result of concurrent enrollment participation. Sensitivity analysis After the P SM ATE results were tabulated, a sensitivity analysis was performed to see how robust the findings were to potential unobserved confounders. Rosenbaum (2002) proposes a bounding approach to determine to what extent an unobserved variable would have to infl uence the selection process in order to mitigate any statistically significant findings. Bounds for the significance levels and confidence intervals of the results are defined by varying the odds ratios that two individuals with the same observed covariate s are assigned to the treatment group at a different rate due to an unobserved variable. The results of the sensitivity analysis suggest that the ATE estimates for all outcomes are fairly robust to changes in the likelihood of treatment assignment due to h idden bias. The college matriculation treatment effect, for example, is statistically significant until an unobserved variable caused the odds ratio of participating in concurrent enrollment to differ between treatment and comparison groups by a factor of 1.75. That finding indicates that the unobserved variable would need to exert a large influence one that that is larger than the effect sizes of any individual covariate included in the PSM model to undermine the positive effect of concurrent enrollment o n college matriculation. The ATE on persistence rates had a similar sensitivity factor of 1.65, while the ATE for remediation rates was robust to hidden bias that would more than double the odds of participation variable triples the odds that concurrent enrollment participation differs between the treatment and com parison groups. Overall, the sensitivity analysis performed for each model found that the influence of concurrent enrollment on each outcome is resilient to the presence of moderate and even high

PAGE 111

100 amounts of selection bias. As mentioned before, the sensiti vity analysis cannot determine the presence of hidden selection bias, but it lends credence to these findings since the unobserved bias would have to be quite strong to negate the statistical significance of the results (An 2013; Rosenbaum, 2002) Regress ion Findings A progression of regression models were estimated to assess the robustness of the PSM findings. The regression models also permit the addition of school level fixed effects, which could not be included in the PSM analysis. Table 17 displays on e example of the progression of models estimated with logistic regression for the dichotomous college matriculation and concurrent enrollment variables as the dependent and key independent vari ables, respectively. The same progression of models was run for all four outcome variables. High school clustered, r obust standard errors are reported for all models to correct for serial correlation and h etero scedasticity. Initially, a bivariate model is run indicating a positive, statistically significant relationsh ip between the two va riables. The second model adds demographic (FRL, ELL, SPED race/ethnicity and gender ) control variables, and the thi rd model adds the control for academic achievement (ACT composite score). The fourth model, and final, model includes all of the previous covariates in addition to graduation year and high school fixed effects. In the regressi on results displayed in Table 17, the coefficient for concurrent enrollment participation is statistically significant (p<.01) across all models a nd the direction of the coefficient indicates that p articipation in concurrent enrollment in high school results in positive gains in college matriculation rates As expected, the pseudo R squared increases across the model specifications as variables are added. The bivariate model explains 0.8 percent of the variance, while the final model accounts for 16.5 percent of the variation. The results from the final model that includes fixed effects are used to generate the average treatment effects are discussed in the following section.

PAGE 112

101 Table 17 Progression of Logistic Regression Models Estimating the Effect of Concurrent Enrollment Participation on College Matriculation (1) (2) (3) (5) Variables Bivariate Includes demographics Inclu des ACT score Full model w/fixed effects a Concurrent Enrollment 0.632*** 0.635*** 0.553*** 0.594*** (0.053) (0.045) (0.038) (0.037) FRL 0.676*** 0.382*** 0.348*** (0.036) (0.031) (0.029) ELL 0.784*** 0.192*** 0.178*** (0.049) (0.0 52) (0.043) SPED 1.060*** 0.223*** 0.314*** (0.039) (0.035) (0.034) Hispanic 0.474*** 0.060* 0.051** (0.036) (0.031) (0.024) Black 0.007 0.522*** 0.611*** (0.053) (0.053) (0.052) Asian 0.590*** 0.574*** 0.575*** (0.055) (0.061) ( 0.067) Other Race 0.556*** 0.258*** 0.178*** (0.065) (0.065) (0.066) Male 0.316*** 0.347*** 0.369*** (0.015) (0.016) (0.016) ACT Composite Score 0.166*** 0.152*** (0.003) (0.003) 2012 Graduation Year 0.058*** (0.018) 2013 Graduation Year 0.110*** (0.018) Constant 0.455*** 0.988*** 2.581*** 3.771*** (0.044) (0.043) (0.070) (0.076) Observations 131,681 131,681 131,681 131,601 b Pseudo R squared 0.008 0.065 0.141 0.165 Correctly classified 63.27% 67.49% 7 1.49% 72.76% Robust, high school clustered standard errors in parentheses *** p<0.01, ** p<0.05, p<0.10 a Fixed effect model includes graduation years 2012 and 2013 (2011 as baseline) and dummies for 423 high schools (not displayed) b 15 high schools pr edicted failure perfectly; a total of 80 students were dropped as a result Table 18 provides average treatment effects for the dichotomous concurrent enrollment variable from the multivariate regression models for the different outcome variables The coef ficient for the concurrent enrollment variable is statistically significant on all four college outcomes On average, and controlling for confounding variables, the probability of going to college is 10. 57 percentage points higher for students who particip ated in concurrent enrollment than for those

PAGE 113

102 who did not participate. Concurrent enrollment students see their probability of needing remedial education decrease, on average by 6.21 percentage points when compared to their peers. Participating in concurre nt enrollment increases first year college GPAs by an average of one tenth of a point, and results in an increase in retention rates of 3.36 percentage points, on average. Table 18 Average Treatment Effects Outcome 1 Outcome 2 Outcome 3 Outcome 4 College Matriculation Remedial Education First year College GPA College Persistence Concurrent Enrollment 10.57*** 6.21*** 0.10*** 3.36*** (0.00 6 ) (0.005) (0.014) (0.004) FRL 6.65*** (0.006) 2 33 *** (0.004) 0.09* ** ( 0.013 ) 5 19 *** (0.005) ACT Composite Score 2 8 1*** (0.0004) 6 .18*** (0.0003) 0.06 *** (0. 001) 1 56 *** (0.0004) Hispanic 0.95** (0.004) 0.16 (0.005) .06*** (0.0110 0. 06 (.004) Robust, high school clustered standard errors in parentheses *** p<0.0 1, ** p<0.05, p<0.10 Table 18 also provides the treatment effects for other variables of interest to help compare the magnitude of the effects. FRL students have college matriculation rates that are, on average, 6.65 percentage points below those of non FRL students, while Hispanic students have a college matriculation rate that is 0.95 percentage points below that of white students, all else equal. A one point increase in composite ACT score is associated with a 2.81 percentage point increase in college matriculation rates. To get a similar effect size as concurrent enrollment, a student would need to increase his or her ACT score by roughly 4 points. Prior research has found FRL eligibility, ACT scores and Hispanic ethnicity to be correlated with colleg e matriculation; the fact that the effect size for concurrent enrollment is substantively larger than the effect size for those variables lends support to the assertion that the findings have practical significance The treatment effect for concurrent enr ollment in the remedial education model is similar to ACT score; both are larger than the FRL effect. In the GPA

PAGE 114

103 concurrent enro llment, and a similar effect size is seen for the FRL variable, although it goes in the other direction ( 0.09). The effect size for FRL students is larger than the concurrent enrollment treatment in the persistence model, but both are larger than the trea tment effect for ACT scores. In sum, concurrent enrollment has the largest comparative effect size on matriculation rates, with small comparative effects seen for the other outcome variables. Comparison of Results Table 19 provides a comparison of average treatment effects across four methods: 1) PSM using 1:1 matching with caliper; 2) a multivariate regression using the same covariates that are included in the PSM model; 3) fixed effects regression; and 4) the difference in means between students who part icipated in concurrent enrollment and those who did not. The multivariate regression results include the same conditioning effect the PSM includes, but there is not a matching process conducted before estimating the regression. The fixed effects regression uses the results provided in the previous section ( Table 6) derived from regression models that add school and time fixed effects to the list of covariates included in the PSM models. The unmatched difference in means provides a baseline of what the treat ment effect would be without controlling for any exogenous factors or adjusting for sample biases. Overall, the results in Table 19 are fairly consistent between the PSM analy sis (matching 1:1), multivariate regression analysis and the fixed effects analy sis on t he college matriculation and remedial education outcome results All three methods provide similar and more modest results for college matriculation and remedial education t han the unmatched, unconditioned difference in means. This indicates that w hether using PSM or regression analysis, confounding effects are being controlled for in those two outcome models In the college matriculation model, the PSM method

PAGE 115

104 produced the most conservative estimate of the average treatment effect, while the fixed e ffects regression produced the most conservative estimate for the remedial education model. Table 19 Comparison of Average Treatment E ffects Outcome PSM Matched 1:1 Multivariate regression (using PSM model) Fixed effects regress ion Unmatched Difference in Means College matriculation 10.16 *** (.003) 10.40 *** (.003 ) 10.57 *** (.006 ) 13.66 *** (.003) Remedial e ducation 7.54 *** (.004) 7.52 *** (.004) 6.21*** (.005) 7.76 *** (.005) First year college GPA .08*** (.009) .07 *** ( .0 0 9 ) 0.10*** (0.014) .07*** (.009) College persistence 2.17 *** (.003) 2.13 *** (.003) 3.36*** (.004) 1.74 *** (.003) Robust standard errors in parentheses *** p<0.01, ** p<0.05, p<0.10 The college GPA average treatment effects range from 0.07 to 0.1 w ith the multivariate regression results mirroring those of the unmatched difference in means. The fixed effects regression produced the largest average treatment effect (.10) for the GPA outcome model, and for the college persistence model (3.36). The aver age treatment effect of concurrent enrollment on persistence is larger when using PSM, multivariate regression or fixed effects regression than when looking at the unmatched difference in means. Thus, those methods are adjusting for unobserved bias in both the college GPA and the college persistence models, although in a way that increases effect size rather than decreases. Concurrent Enrollment Effects for Low Income Students and Minority Students While participating in concurrent enrollment appears to ha ve beneficial impacts on college outcomes for the average student, the author is also interested in understanding if and how the effects of program participation vary depending on the race/ethnicity or income level of students.

PAGE 116

105 Robustness Checks Once again a progression of regression models were estimated for each dependent variable to check robustness As an example, Table 20 displays the progression of models estimated with logistic regression for the dichotomous college matriculation and concurrent enro llment variables as the dependent and key independent variables, respectively. The same progression of models was run for the other three dependent variables (remediation, GPA, and persistence) Initially, a bivariate model is run indicating a positive, st atistically significant relationship between the two variables. The second model adds demographic (FRL, ELL, SPED, race/ethnicity, and gender) control variables, and the third model adds the control for academic achievement (ACT composite score). The fourt h model includes the interaction terms of race/ethnicity and FRL eligibility with participation in c oncurrent enrollment. The fifth, and final, model includes all of the previous covariates in addition to cohort year and high school fixed effects. In the regressi on results shown in Table 20, the coefficient for concurrent enrollment participation is statistically significant (p<.01) across all models and the direction of the coefficient indicates that p articipation in concurrent enrollment in high school results in positive gains in college matriculation rates The coefficients on the Hispanic and FRL interaction terms indicate a positive, statistically significant relationship The progression of regression models for the additional dependent variables al so demonstrated robustness across specifications. The results from the final model that includes fixed effects and interaction terms for each dependent variable are displayed and discussed in the following section.

PAGE 117

106 Table 20 Prog ression of Logistic Regression Models Estimating the Effect of Concurrent Enrollment Participation on College Matriculation DV: College Matriculation (1) (2) (3) (4) (5) Variables Bivariate Includes demographics Includes ACT score Includes interaction t erms Full model w/fixed effects a Concurrent Enrollment 0.632*** 0.635*** 0.553*** 0.471*** 0.523*** (0.053) (0.045) (0.038) (0.044) (0.043) FRL 0.676*** 0.382*** 0.403*** 0.366*** (0.036) (0.031) (0.032) (0.029) ELL 0.784*** 0.192*** 0.185*** 0.172*** (0.049) (0.052) (0.052) (0.044) SPED 1.060*** 0.223*** 0.222*** 0.314*** (0.039) (0.035) (0.035) (0.034) Hispanic 0.474*** 0.060* 0.084*** 0.069*** (0.036) (0.031) (0.031) (0.025) Black 0.007 0.522*** 0.535*** 0 .619*** (0.053) (0.053) (0.055) (0.054) Asian 0.590*** 0.574*** 0.562*** 0.562*** (0.055) (0.061) (0.064) (0.069) Other Race 0.556*** 0.258*** 0.263*** 0.178*** (0.065) (0.065) (0.064) (0.065) Male 0.316*** 0.347*** 0.347*** 0.369*** (0.015) (0.016) (0.015) (0.016) ACT Composite Score 0.166*** 0.166*** 0.152*** (0.003) (0.003) (0.003) Hispanic CE 0.158** 0.126** (0.073) (0.058) Black CE 0.103 0.050 (0.088) (0.089) Asian CE 0.091 0.100 (0.11 1) (0.123) Other CE 0.039 0.009 (0.230) (0.229) FRL CE 0.123** 0.114** (0.060) (0.057) 2012 Graduation Year 0.058*** (0.018) 2013 Graduation Year 0.108*** (0.018) Constant 0.455*** 0.988*** 2.581*** 2.568** 3.749*** (0.044) (0.043) (0.070) (0.070) (0.077) Observations 131,681 131,681 131,681 131,681 131,601 b Pseudo R squared 0.008 0.065 0.141 0.141 0.165 Correctly classified 63.27% 67.49% 71.49% 71.52% 72.77% Robust, high school clustered stan dard errors in parentheses *** p<0.01, ** p<0.05, p<0.10 a Fixed effect model includes graduation years 2012 and 2013 (2011 as baseline) and dummies for 423 high schools (not displayed) b 15 high schools predicted failure perfectly; a total of 80 student s were dropped as a result

PAGE 118

107 Interaction Effects Table 21 displays results from the multivariate regression models using the dichotomous independent variable of concurrent enrollment participation The interaction between income (FRL status) and concurrent e nrollment participation was statistically significant in all outcome models at p<.10 and in three of the four models at p<.05. The interaction between Hispanic students and concurrent enrollment was statistically significant at p<.05 in the college matricu lation model, while the Asian student interaction term was significant in the remedial education and GPA models (p<.01). The other interaction terms were not statistically significant. Interpreting interaction terms in logistic regression models is not as straightforward as in linear regression. Thus, the findings are presented for each dependent variable as the change in the probability that the outcome will occur based on the interactions. College matriculation interaction results. Figure 9 displays the predicted probability of college matriculation for students who were FRL eligibl e compared to those who were not FRL eligible by concurrent enrollment p articipation. The probability of immediately attending college for FRL students who participate in concu rrent enrollment increases by 12. 7 8 percentage points compared to their FRL peers who do not participate in the program. When looking at higher income students (not FRL eligible), the difference in matriculation rates between concurrent enrollment particip ants and non participants is 9.9 2 percentage points. Those differences along with the variance between those differences (2.86) are statistically significant at p<0.01 In logistic regression models, it is important to consider, in particular, the statist ical i n If the variance in differences is negligible and statistically insignificant then there is not an interaction effect present (Mitchell & Chen 2005). This is the case for the interactions between Asian students, black students and students of other race with concurrent enrollment participation

PAGE 119

108 in the college matriculation outcome model. On the other hand, if the variance between differences is statistically significant, as is the case with the FRL interaction term, then that indicates there is a meaningful interaction effect between FRL status and concurrent enrollment participation occurring in this outcome model. Table 21 Regression M odels Estimating the Interaction E ffects of Concurrent Enrollment Participation on College O utcomes Outcome 1 Outcome 2 Outcome 3 Outcome 4 Variables College Matriculation Remedial Education First year college GPA College Persist ence Concurrent Enrollment 0.523*** 0.517*** 0.091*** 0.200*** (0.043) (0.054) (0.016) (0.042) ACT Composite Score 0.152*** 0.566*** 0.057*** 0.114*** (0.003) (0.008) (0.001) (0.003) Hispanic CE 0.126** 0.047 0.009 0.011 (0.058) (0.096) (0.024) (0.072) Black CE 0.050 0.047 0.033 0.147 (0.089) (0.139) (0.042) (0.091) Asian CE 0.100 0.468*** 0.123*** 0.170 (0.123) (0.161) (0.046) (0.169) Other CE 0.009 0.260 0.115 0.194 (0.229) (0.376) (0.122) (0.226) FRL CE 0.114* 0.241*** 0.052* 0.152** (0.057) (0.092) (0.027) (0.065) Constant 3.749*** 11.189*** 1.627*** 3.267*** (0.077) (0.159) (0.025) (0.072) Observations 131,601 a 61,100 b 52,242 c 83,291 d Pseudo R squared 0.136 0.426 0.0982 Adj R squared 0.15 0 Correctly classified 72.77% 84.87% 81.64% Robust high school clustered standard errors in parentheses *** p<0.01, ** p<0.05, p<0.10 All models include demographic controls and school and year fixed effects Outcome model 3 is estimated with OLS ; all other models are estimated with logistic regression a Includes all high school graduates b Includes high school graduates who immediately enrolled in college at an in state, public institution c Includes high school graduates who immediately enrolled in college at an in state, SURDs institution d Includes high school graduates who immediately enrolled in college anywhere

PAGE 120

109 FRL Non FRL Hispanic White Variance in Differences Non Participant 56.42% 63.57% 60.70% 62.01% FRL v. Non FRL: 2.86 *** Hispa nic v. White: 2.25** Concurrent Enrollment Participant 69.19% 73.49% 72.91% 71.97% Difference in Probability 12.78 *** 9.9 2*** 12.21%*** 9.96% *** p<0.01 **p<0.05 Figure 9 Probability of College Matriculation, by Concurren t Enrollment Participation and Free or Reduced Price Lunch (FRL) Status and Race/Ethnicity (Hispanic or white) The other interaction term that was statistically significant in the college matriculation model was with Hispanic students and concurrent enrol lment participation. H ispanic students who take concurrent enrollment have, on average, a 12.21 percentage point increase in the probability of going to college over their non concurrent enrollment Hispanic peers, as compared to the baseline group of white students, who see a 9.96 percentage point increase when taking concurrent enrollment versus not participating in the program (see Figure 9 ). While the interaction term for white students is not statistically significant, the difference in the interaction terms (2.25) is statistically significant at p<0.05, indicating that Hispanic students who take concurrent enrollment courses see a greater impact on their likelihood of going to college than do white students who participate in the program. College remedi al education interaction results As displayed in Figure 10 the probability of 56.42% 63.57% 60.70% 62.01% 69.19% 73.49% 72.91% 71.97% 30.00% 40.00% 50.00% 60.00% 70.00% 80.00% FRL Non FRL Hispanic White Probability of College Matriculation College Matriculation Non-Participant Concurrent Enrollment Participant

PAGE 121

110 needing remediation when in college for FRL students who participate d in concurrent enrollment decreases by 8 2 5 percentage points compared to their FRL peers who did not parti cipate in the program. When looking at higher income students, the difference in remedial education rates between concurrent enrollment participants and non partici pants is 5.52 percentage points. Those differences, along with the variance between those d ifferences (2.72 ), are statistically significant at p<0.01 FRL Non FRL Variance in Differences Non Participant 38.32% 35.31% 2.72 *** Concurrent Enrollment Participant 30.07% 29.79% Difference in Probability 8.25 *** 5.5 2*** *** p<0.01 Figur e 10 Probability of College Remediation, by Concurrent Enrollment Participation and Free or Reduced Price Lunch (FRL) Status While the coefficient for the interaction term Asian*Concurrent Enrollment was statistically significant in the college remediation logistic regression model, the difference in predicted probabilities for Asian students taking concurrent enrollment compared to Asian students not participating is not statistically significant, nor is the variance in differenc es between the Asian students interaction term and the white students interaction term. 38.32% 35.31% 30.07% 29.79% 0.00% 5.00% 10.00% 15.00% 20.00% 25.00% 30.00% 35.00% 40.00% 45.00% FRL Non FRL Probability of College Remediation Need for College Remedation Non-Participant Concurrent Enrollment Participant

PAGE 122

111 College GPA interaction results. The first year college GPA outcome model is estimated with linear regression making the interpretation of the interaction terms more s traightforward. White students who take concurrent enrollment have, on average, a positive increase in their first year GPA of about one tenth of a p oint compared to white students without concurrent enrollment (p<.01). The coefficient on the Asian students interaction term is statistically significant at p<.01 and negative. Asians students who were in concurrent enrollment do not have a statistically different GPA than Asian students who did not participate, but the difference between their change in GPA a nd the change in GPA for white students (i.e. the difference in differences) is statistically significant, and that is what is reflected in the coefficient for the interaction term. The change in GPA for Asian students is, on average, 0.12 points lower tha n the change in GPA for white students, indicating that white students participating in concurrent enrollment see a larger effect on their first year GPA than do Asian students. The coefficient on the FRL interaction term is statistically significant at p< .10 and positive. The change in GPA for FRL students is, on average, 0.05 points higher than the change in GPA for white students, indicating that FRL students participating in concurrent enrollment see a slightly larger effect on their first year GPA than do non FRL students. College persistence interaction results Figure 11 displays the predicted probability of college persistence for students who were and were not FRL eligible by concurrent enrollment participation. The probability that FRL students wh o participated in concurrent enrollment will return for a second year of college increases by 5.50 percentage points when compared to their non concurrent enrollment, FRL peers. When looking at higher income students, the difference in persistence rates be tween concurrent enrollment participants and non participants is only 2.72 percentage points. Those differences, along with the variance i n those differences (2.78), are statistically significant at p<0.01 The predicted probability of college persistence for low income

PAGE 123

112 stude nts who take concurrent enrollment (81.7%) is nearly identical to that of higher income students who do not take concurrent enrollment (82.0%) FRL Non FRL Variance in Differences No n Participant 76.19% 81.69% 2.78 *** Concurrent Enrollment Participant 81.96% 84.68% Difference in Probability 5.50 *** 2.7 2*** *** p<0.01 Figure 11 Probability of College Persistence by Concurrent Enrollment Participation a nd Free or Reduced Price Lunch (FRL) Status In summary, the results from the interaction effects show that while all students benefit from concurrent enrollment, low income students see a n even greater increase in positive college outcomes when compared t o higher income students. This is a critical finding given the policy goal of increasing outcomes for traditionally underserved students. The interaction terms for minority students were not as consistent. Nonetheless, one important finding is that Hispani c students who take concurrent enrollment courses see a greater impact on their likelihood of going to college than do white students who participate in the program. Effects of Concurrent Enrollment Credit Hour Levels on College Outcomes In the last part of this study, the author uses a categorical measure of concurrent enrollment participation to understand if there are differential effects on college outcomes depending on how 76.19% 81.96% 81.69% 84.68% 70.00% 72.00% 74.00% 76.00% 78.00% 80.00% 82.00% 84.00% 86.00% FRL Non FRL Probability of College Retention College Persistence Non-Participant Concurrent Enrollment Participant

PAGE 124

113 many credits students take while in high school. Five categories were created f or the number of credit hours attempted, with zero credit hours set as the baseline category in the analysis. The remaining four categories were selected after viewing the descriptive statistics (see Table 22 ) and seein g natu ral breaks between each categor y that equate roughly to quartiles. Table 22 Credit Hours Descriptive Statistics for Concurrent Enrollment Students Number of Credit Hours N Percent Cumulative Percent 1 3 Credit Hours 5,382 26.5 26.5 3 6 Credit Hours 5,299 26.0 9 52.58 6 12 Credit Hours 4,876 24 76.59 12+ Credit Hours 4,756 23.41 100 Total 20,313 100 Table 23 displays sample means of the college outcome variables by credit hours category. Generally, there are improved outcomes when moving across the table f rom no credit hours to 12+ hours. An upward trend in means is considered an improvement for college matriculation, GPA and persistence, while a downward trend in means is an improvement when looking at the need for remedial education. Fixed effects regress ion models were estimated to see if those trends hold while controlling for a set of covariates. Table 23 Sample Means of Key College Outcomes by Concurrent Enrollment Credit Hours Outcome Variables Concurrent Enrollment Credit Hou rs No Credit Hrs 1 3 Hrs 3 6 Hrs 6 12 Hrs 12+ Hrs College Matriculation 61.12% 73.52% 72.90% 75.06% 77.96% Remedial education 36.21% 35.42% 29.39% 27.93% 20.59% First year college GPA 2.74 2.75 2.77 2.80 2.92 College persistence 80.96% 82.16% 83.25% 82.84% 82.58% Cell values represent means of the sample as a whole for each outcome, by credit hour level. Table 2 4 presents the results from the regression models using the categorical credit hours measure. All four college o utcomes generally improve whe n students attempt high numbers of credit hours as compared to no credit hours even while controlling for demographic and academic factors and including school and year fixed effects.

PAGE 125

114 Table 24 Progression of Regression Models E stimating the Effect of Concurrent Enrollment Participation on College Matriculation Outcome 1 Outcome 2 Outcome 3 Outcome 4 Variables College Matriculation Remedial Education First year college GPA College Persistence 1 3 Credit Hours 0.518** 0.245*** 0.063*** 0.147*** (0.048) (0.058) (0.017) (0.046) 3 6 Credit Hours 0.480*** 0.475*** 0.061*** 0.245*** (0.048) (0.072) (0.019) (0.049) 6 12 Credit Hours 0.650*** 0.749*** 0.105*** 0.309*** (0.051) (0.077) (0.019) (0.050) 12+ Credit H ours 0.828*** 1.116*** 0.209*** 0.387*** (0.073) (0.097) (0.028) (0.083) FRL 0.348*** 0.213*** 0.092*** 0.357*** (0.029) (0.040) (0.013) (0.031) ELL 0.175*** 0.322*** 0.215*** 0.327*** (0.043) (0.087) (0.032) (0.058) SPED 0.315*** 0.407*** 0.037 0.238*** (0.034) (0.104) (0.029) (0.046) Male 0.368*** 0.180*** 0.273*** 0.433*** (0.016) (0.025) (0.008) (0.020) Hispanic 0.050** 0.016 0.062*** 0.004 (0.024) (0.042) (0.011) (0.030) Black 0.610*** 0.048 0.128*** 0.236*** (0.05 2) (0.054) (0.025) (0.047) Asian 0.575*** 0.555*** 0.000 0.516*** (0.067) (0.072) (0.017) (0.057) Other Race 0.180*** 0.040 0.102** 0.032 (0.066) (0.128) (0.047) (0.097) ACT Composite Score 0.152*** 0.565*** 0.056*** 0.114*** (0.003) (0.008) (0.001) (0.003) 2012 Graduation Year 0.060*** 0.011 0.018* 0.024 (0.018) (0.033) (0.010) (0.023) 2013 Graduation Year 0.114*** 0.171*** 0.028*** 0.068*** (0.018) (0.034) (0.009) (0.026) Constant 3.737*** 11.222*** 1.630*** 5.419*** (0.077 ) (0.161) (0.026) (0.073) Observations 131,601 a 61,100 b 52,242 c 83,291 d Pseudo R squared 0.165 0.467 0.151 (Adj. R sq) 0.098 Correctly classified 72.76% 84.91% 81.63% Robust, high school clustered standard errors in parentheses *** p<0.01, ** p< 0.05, p<0.10 All models include demographic controls and school and year fixed effects Outcome model 3 is estimated with OLS; all other models are estimated with logistic regression a Includes all high school graduates b Includes high school graduates wh o immediately enrolled in college at an in state, public institution c Includes high school graduates who immediately enrolled in college at an in state, SURDs institution d Includes high school graduates who immediately enrolled in college anywhere

PAGE 126

115 When looking at the college matriculation and GPA models in Table 24, the coefficients for 1 3 credit hours and 3 6 credit hours are very similar, but both are positive and statistically significant indicating that taking 1 to 6 credits of concurrent enrollmen t, typically the equivalent of one or two courses, has a positive effect on outcomes as compared to students who take zero credits. The remedial education and college persistence models show steady increases in the size of the coefficient as the number of credit hours attempted increases. Table 25 provides treatment effects for each credit hour level, by outcome, to further explore the substantive differe nces in levels. Recalling the average treatment effects (ATEs) from the fixed effects regression using the dichotomous variable for concurrent enrollment participation (Table 18) one will note that each of those ATEs falls in between the results for the categories of 3 6 credit hours and 6 12 credit hours. Table 25 Average Treatment Effects of Credit Hours Levels on Coll ege Outcomes Credit Hours Outcome 1 College Matriculation Outcome 2 Remedial Education Outcome 3 First Year College GPA Outcome 4 College Persistence ATE 95% CI ATE 95% CI ATE 95% CI ATE 95% CI 1 3 9.08 *** (0.00 8 ) [7.55, 10.61] 2.64 *** (0.006 ) [ 3.84, 1.44] 0 .06 *** (0.017) [0.03, 0.10] 1.93 *** (0.006 ) [0.79, 3.08] 3 6 8.45*** (0.008) [6.89, 10.01] 5.06*** (0.007) [ 6.51, 3.60] 0.06 *** (0.019) [0.02, 0.10] 3.15*** (0.006) [2.00, 4.31] 6 12 11.21*** (0.008) [9.63, 12.80] 7.85*** (0.007) [ 9.31, 6.38] 0.11 *** (0.019) [0.07, 0.14] 3.91*** (0.006) [2.77, 5.06] 12+ 13.95** (0.011) [11.83, 16.07] 11.46*** (0.009) [ 13.24, 9.67] 0.21 *** (0.028) [0.15, 0.26] 4.81*** (.009) [3.00, 6.64] Robust, high school clustered standard errors in pare ntheses *** p<0.01, ** p<0.05, p<0.10 Omitted category is 0 credit hours. The ATE for the dichotomous participation variable in the fixed effects model, for example, was 10.57, which falls in between the ATE of 8.45 for 3 6 credit hours and the ATE of 11.21 for 6 12 credit hours. One could also note, however, that an ATE of 10.57 falls within the 95% confidence

PAGE 127

116 interval for 1 3 credit hours, albeit at the bounds of the interval. The college matriculation outcome treatment effect does not vary much betwe en 1 3 and 3 6 credit hours, while for the remedial education outcome the difference between the two categories is more pronounced. The ATE for the dichotomous participation variable was 6.21, which is included in the confidence intervals for 3 6 and 6 12 credit hours. The GPA outcome shows similar effects for taking 1 3 credit hours and 3 6 credit hours (0.06) as compared to zero credit hours, but there are substantively larger effect sizes for 6 12 credits and over 12 credits (0.11 and 0.21, respectively ). In comparison, the effect size for the dichotomous participation model was 0.10 of a GPA point. The effect sizes for the college persistence outcome increase by just around 1 percentage point from one category to the next. One overall inference from th e results is that while those students who take six or more credits see greater positive effects on college access and success, even those students with only one or two concurrent enrollment courses (between 1 and 6 credits) see benefits accrue. The colleg e matriculation outcome, in particular, is still substantively meaningful even for those students taking fewer credits. The effect sizes for college remediation, GPA and persistence do get considerably smaller when looking just at those students take 1 3 c redits, but the confidence intervals do not include zero in any of the models. Selection bias con cerns increase w hen considering the type of student that would take more than 6 credit hours, since s uch a student is likely to already be college bo und with or without the treatment. Thus, looking at the treatment effects for the 1 3 and 3 6 credit hour categories, while providing more conservative estimates, may provide higher confidence in the validity of the results. Moreover, even though the 1 3 a nd 3 6 credit hour categories have smaller effect sizes, the ATEs from the dichotomous college participation model are within the 95% confidence intervals for at least one of those bottom two credit hour categories. This lends some support to the validity of the results from the prior analysis conducted using the dichotomous predictor variable.

PAGE 128

117 Conclusion & Discussion The PSM analysis found positive, statistically significant and substantively large effects of concurrent enrollment participation on college matriculation for Colorado high school graduates an increase of 10.16 percentage points in the probability of college enrollment immediately following high school graduation. The PSM results also suggest that taking participating in concurrent enrollment reduces the chance of needing remedial education in the first year of college by 7.5 percentage points. Other research has found that freshmen who start in college level courses have higher chances of completing a degree (e.g. Ru tschow & Schneider, 2011 ) so increasing the rates of college lev el course placement is critical. Participating in concurrent enrollment is associated with an increase in first year college GPAs of 0.08 points, on average. Lastly, the PSM analysis found that participating in concurr ent enrollment increases the probability of persisting in college from freshmen to sophomore year by 2.2 percentage points. The findings from the fixed effects regression analysi s and the PSM analysis are quite similar, which corroborates other studies tha t have found minimal differences between PSM and regression output ( Brand & Halaby, 2006; Shadish & Steiner, 2010) When interaction effects were added to the fixed effects regression, the results indicated that there is a greater increase in positive coll ege outcomes for free and reduced price lunch (FRL) students than non FRL students, which is what one would hope to see given that a primary goal of concurrent enrollment is to increase college access and success for traditionally underserved groups of stu dents, including low income students. Additional results from the interaction term models indicate that Hispanic students who take concurrent enrollment courses see a greater impact on their likelihood of going to college than do white students who partici pate in the program. While such findings are important contributions to research and practice, the study does have some limitations. The dataset did not include information on w hether the concurrent

PAGE 129

118 enrollment course was taught at the high school, at a pos tsecondary campus, or online. Taking a course on campus could have more influence on postsecondary readiness than taking a course at dependent on the whether the cou rse is taug ht by high school teachers or by college faculty Researchers with access to course setting and instructor data should considering exploring if there are relationships between those variables and postsecondary outcomes. Another significant limi tation of this study is the potential for omitted variables to threaten internal validity. Individual motivation is not a directly observable variable and could have a confounding effect on the analysis. A student, for instance, may be intrinsically motiva ted to both select into concurrent enrollment courses and to attend college. The presence of selection bias means that the effect sizes may be overestimated. However, the sensitivity analysi s performed after the PSM estimations provide s credibility to the results and indicates they a re robust enough to withstand the presence of moderate and even high levels of unobserved bias The sensitivity analysis found that the unobserved variable would need to exert a large influence one that that is larger than the e ffect sizes of any individual covariate included in the PSM model to undermine the positive effect of concurrent enrollment on each outcome. Furthermore, t he final part of this study considered a categorical independent variable to assess the effects of c oncurrent enrollment credit hour levels on college outcomes. Generally, as would be expected, the results show that college o utcomes improve when students attempt higher numbers of credit hours as compared to no credit hours. The top two categories of cred it hour levels (6 12 hours and 12+ hours) likely include more selection bias than the lower two categories of credit hour levels (1 3 hours and 3 6 hours), because students taking a large number of credit hours may be intrinsically motivated to pursue adva nced educational opportunities, including higher education.

PAGE 130

119 The lower two categorical levels perhaps provide a more accurate level of treatment effects, and even in viewing those findings, one can see meaningful, positive effects on college outcomes. Over all, across the different methods and model specifications in this study, the findings remained robust and conveyed the same narrative, which is that p articipation in concurrent enrollment in high school results in positive gains in college enrollment rate s, first year grade point averages, and college persistence rates, and results in a decrease in the need for remedial education. These are promising findings that contribute to the collective knowledge about what programs improve postsecondary and workforc e readiness for all students.

PAGE 131

120 CHAPTER VI CON C LUSION The purpose of this study was to investigate how state level policies, particularly those that create voluntary programs for schools and students, can meet the intended policy goals and affect educat ional outcomes This research was motivated by two realities: 1) that low income and minority students, on average, lag behind their peers on nearly every important education milestone including co llege enrollment and completion ; and 2) that a higher edu cation credential is increasingly necessary to have a productive career and earn family sustaining wages State policymakers, having observed those same realities, have implemented countless policies to better prepare stude nts for life after high school. W ith the development of statewide longitudinal data systems over the past decade, researchers are now able to investigate if those policies are having their intended effect of improving outcomes for traditionally underserved students. One of the prominent approaches among states to expanding college access is concurrent enrollment, which provides high school students the opportunity to enroll in a college course for which they may receive both high school and college credit. While concurrent enrollment prog rams have been available in public high schools for at leas t the last half century, they were typically seen as a n enrichment opportunity for academically advanced students. Programs have grown exponentially since the early 2000s when policymakers began ex panding concurrent enrollment opportunities to those students who are traditionally underserved, including students o f color and low income students, as well as to those students who are not high academic performers (Hoffman, Vargas & Santos, 2008a) Prop onents of concurrent enrollment argue that it increases academic preparation for college and provides momentum toward degree attainment by giving students the opportunity to enter college with credits already accumulated (An, 2013; Swanson, 2008). Providin g students

PAGE 132

121 exposure to college is also thought to be a strategy for developing metacognitive skills readying students for the demands of college life and increasing college aspirations Moreover, p olicymakers are also drawn to concurrent enrollment as a way to increase college affordability by offering college courses at low or no cost to families. With the foregoing as a backdrop, t his stud y set out to understand, first what factors le a d some schools to adopt concurrent enrollment more quickly and implement the program more intensely a s compared to other schools After understanding the key factors at play at the school level, the study also sought to evaluate how effective concurrent enrollment is at improving college access and success for all students, including low income and minori ty students. These research questions are collectively impor tant because the only way state policies can significan tly impr ove educational outcomes is if program s are both widely diffused in schools and impactful on individual students The d ata used in th is study to evaluate the research questions were collected primarily through state education agencies: the Colorado Department of Higher Education (CDHE) and the Colorado Department of Education (CDE). The author constructed panel data sets by com piling publicly available data from the agency websites and by procuring a de identified and secure cross section of student level data from CDHE. The datasets are timely, comprehensive and longitudinal between K 12 and higher education due to data sharin g agreemen ts in place between the two agencies. The research began with an event history analysis of the diffusion of concurrent enrollment acro ss high schools in the state. OLS fixed effects regression and a dynamic panel data model were used to explore any factors related to the intensity of program participation within high schools once conc urrent enrollment was adopted. F ixed effects regression and propensity score matching (PSM) were used in a student level analysis to determine relationships between concurrent

PAGE 133

122 enrollment participation and the college access and success outcomes. Participation in concurrent enrollment was explored both as a dichotomous measure (yes/no) and as an intensity le vel ( number of credits). The different components of the research design relied on similar administrative data but had different guiding questions and units of analysis. Taken altogether, there were several important findings from the multi level analysis, which were presented in chapter s 4 and 5. Key findings a re highlighted here, followed by a discussion of the implications of the findings for research and practice. This chapter concludes with a description of the limitations and suggestions for future research. Key Findings The diffusion of concurrent enrollm ent throughout Colorado high schools was rapid and nearly complete by the end of the five year event history analysis which ran from 2010 to 2015 Results from the analysis found those schools that were more likely to adopt concurrent enrollment immediate ly after legislation passed in 2009 had higher college matriculation rates and also higher rates of free or reduced price lunch (FRL) students. Early adopters were also more likely to be larger high schools, and schools that had previously offered the Post Secondary Education Options (PSEO) adopt concurrent enrollment as traditional district run schools. Results from the dynamic panel data model, which included a lagged value of the dependent variable and school and year fixed effects, found that the still growth in participation rates that occurs year over year. The lag ged dependent variable indicates for every 1 percent increase in t he previous The variables for median household income and PSEO participation had a positive, statis tically significant association with

PAGE 134

123 concurrent enrollment participation rates S chool size, distance to the nearest community college and the charter school indicator were statistically signi ficant and negatively associated with participation rates The student level analysis found a positive, statistically significant and substantively large effect of concurrent enrollment participation on college matriculation for Colorado high school gradua tes The effect sizes for the matriculation outcome are still meaningfully large even when considering outcomes for students who only took 3 credits (e.g. 1 course) or 6 credits (e.g. 2 courses). The propensity score matching analysis and regression analys is also found that p articipation in concurrent enrollment in high school results in positive gains in first year college grade point averages and college persistence rates, and results in a decrease in the need for remedial education. While concurrent enro llment, on average, improves college outcomes for all students, low income students experience a greater positive impact on their outcomes than higher income students. Additionally, Hispanic students who take concurrent enrollment courses see a greater imp act on their likelihood of going to college than do white students who participate in the program. Implications fo r Research and Practice Policy Diffusion Research The vast majority of the numerous studies conducted using policy innovation and diffusion theory focus on the adoption of a policy without considering what occu rs after adoption in the implementation stage Scholars have identified this gap in the literature and have called for studies to apply policy diffusion analysis beyond a simple dichotom ous measure of adoption and to analyze policy implementation (Shipan & Volden, 2012). Further, policy diffusion research has largely focused on states as the unit of analysis. There have been studies conducted of local governments, but the body of research is much smaller and focuses on municipalities. Thus, as this study considers

PAGE 135

124 both policy adoption and policy implementation using high school s as the unit of analysis the findings make an important contribution to the literature. This study began with a dichotomous measure of adoption, but followed it up with an analysis of policy implementation using a proportional dependent variable measuring the share of students taking concurrent enrollment courses within a high school. The results showed that some v ariables affected both policy adoption and the level of implementation (e.g. charter, PSEO), while other variables had effects on policy adoption but not on whether a school widely implemented the program (e.g. matriculation rates and FRL status). The find ings also lend some support to one of the prominent determinants of policy diffusion according to the literature the importance of s ize and organizational capacity as s maller schools were slower to adopt concurrent enrollment compared to larger ones Regarding the effects of fisc al capacity, the findings were mixed. Schools with higher percentages of FRL students were quicker to adopt, but schools with higher median incomes were more likely to fully implement the program. Distinctions such as this between policy adoption and polic y implementation are precisely what Shipan and Volden (2012) called for in their review of the policy diffusion literature. In this case study, it appears that the characteristics of the policy itself specifically, the salience, clarity and compatibility of the law i nfluenced adoption more than traditional diffusion factors such as fiscal capacity. This supports m ore recent diffusion research that has highlighted the importance of policy characteristics on the diffusion process (Boushey, 2010, Makse & Vold en, 2011, Nicholson Crotty, 2009) Regarding the clarity of the law, c oncurrent enrollment does require work on the part of both school s and districts, but the law is clear and uniform in structure with enough local flexibility embedded in the law to allow schools and coll eges to implement the program in a way that fits with their local needs and current practices. Further, the financial provisions of the law permitted both school s and colleges to set up systems to fund and operate the programs in a way

PAGE 136

125 that does not unduly burden them making the law compatible with current practice T hese policy attributes likely affected how quickly the program was adopted and support findings from Makse and Volden (2011) that found compatible policies those that fit seamlessly into curre nt practices are quicker to diffuse than complex policies that require a great shift in the status quo. Thus, the study makes a modest contribution to the newer stream of policy diffusion literature that focuses on the importance of policy characteristics in the adoption stage. Interestingly, though, this study reaffirmed the significance of some of the more traditional diffusion factors such as fiscal capacity in the imple mentation stage of the analysis, as sc h ools with higher median incomes had higher imple mentation levels. Supporting High S chools If fiscal capacity, organizational capacity, school type and prior program offerings are key predictors of the adoption and implementation of dual enrollment programs, the question becomes what can be done with th is information in practice ? While policy makers and education practitioners cannot easily change the size or location of schools, or their governance structures, targeted support and outreach can be provided to schools that are smaller or less resourced En suring all schools have ac cess to personnel and support either a t the state level, at a regional level (e.g. through their Board of Cooperative Education Services) or through the district to administer concurrent enrollment programs could be a way to improve access across all d ifferent types of high schools. Implementing concurrent enrollment, for example, requires staff to fill out paperwork on behalf of students and negotiate MOUs with community colleges c oncerning the courses that will be offered a nd who will teach them (i.e. a high school teacher or community college faculty). If a high school teacher is leading the course, there needs to be collaboration with the partnering college around course syllabi, curriculum, scope and sequence professional development supervision, and expectatio ns. School counselors need to ensure that the courses students are selecting are aligned

PAGE 137

126 with their individual career and academic plans. If schools with less organizational capacity had a liaison to help them with the logistical and administrative details it could make program implementation easier. While some of the larger school districts have district staff who are dedicated to concurrent enrollment, small districts a re less likely to have such positions ( M. Camacho Liu, personal communication, October 8, 2016 ) While supporting initial program adoption a nd implementation is important for any new educational initiative schools also need sustained support to ensu re the program can continue to be offered. Many of the concurrent enrollment implementation tasks are on going particularly aroun d student course enrollment paperwork, negotiating with colleges over course offerings and instructors, and counseling students and parents on course options. If the local funding arrangements put too much of a burden on school districts, concurrent enrollment programs are at risk of being scaled back when resources are tight. Some school districts and colleges have developed str ong partnerships that make adjustments when needed to promote student access. Eagle County School District, for example, partners with Colorado Mountain College (CMC) to offer concurrent enrollment opportunities. Prior to 2015, the district paid CMC tuitio n for courses located on high school campuses, and CMC reimbursed the school for the cost of the instructor if it was a high school teacher. To ease operations and help the district increase concurrent enrollment opportunities, CMC now waives tuition for a ll courses on high school campuses, and the district is only responsible for covering the costs of a CMC faculty member if there is not a high school teacher available to teach the course (CDE, 2016). The Eagle County Schools and CMC arrangement is an exa mple of a best practice from the model state policy elements discussed in the introduction. While there are a number of ways to creatively fund concurrent enrollment program s the goal should be to ensure that neither system

PAGE 138

127 (K 12 or higher education) is u nduly burdened by the program (Zinth, 2014b, 2015b; Ward & Vargas, 2012) Given the p ositive outcomes in this study and others, showing that concurrent enrollment results in improved college access and success for students (in particular, low income studen ts), the hope would be that such evidence would make a case for sustained state funding and support for concurrent enrollment. States, however, continue to face budget ary challenges, and there is no shortage of com petition for public services that require funding. Thus, as the Education Commission of the States and Jobs for the Future o ften recommend if funding cannot be guaranteed at the state level, local arrangements should be developed to be as predictable and adequate as possible to ensure traditionally unde rserved students continue to have access to dual enrollment opportunities (Zinth, 2014b, 2015b; Ward & Vargas, 2012). Returns on I nvestment I nvesting i n supports for schools with lower capacity during the implementation of concurrent enrollment, or similar college readiness progr ams, is also important given the results from the participation rates analysis. Across all schools, participation rates increased year over year, indicating that schools and families are taking advantage of the opportunities po sed by concurrent enrollment. Given the positive effects of participation on college outcomes, it is encouraging that participation rates continue to increase. While larger schools were quicker to adopt concurrent enrollment, which was expected given the a dvantages of organizational capacity, s maller schools that offer concurrent enrollment often have a high er percentage of their students taking concurrent enrollment as compared to larger schools As explained in Chapter 4, it is likely that when a small high sc hool offers a concurrent enrollment course, for example a math class for seniors, that class comprises a large percentage of its overall enrollment, whereas offering one class at a large high school will only consist of a small share of the overall student body. Further, many small schools located in remote areas use

PAGE 139

128 concurrent enrollment to provide access to content the school is not able to deliver on its own. A rural high school, for instance, can go through a community college to offer college level che mistry, whereas without concurrent enrollment the school would not be able to offer the course at all. Another explanation is that there are fewer opportunities in small schools for advanced coursework, whereas in large high schools, students often have th e choice of various college level courses, including Advanced Placement. The implication for practitioners and policymakers is that it is worth investing in concurrent enrollment support for schools that are small or remote because the se schools can take full advantage of the opportunities available to them Additional returns on the state investment may accrue if students from rural areas are able to earn college credits while in high school and apply those to their college degree so that they graduate more quickly. While this study did not explore the financial savings to families and the state, it is a worthwhi le question for future research, particularly if such research were to compare the cos t effectiveness of concurrent enrollment to other college access programs. Lastly, the findings of the PSEO indicator were, as expected, statistically significant and predictive of concurrent enrollment adoption and the student participation rate. Having had PSEO in place made it easier to transition to concurrent enrollment in terms of already having college partnerships in place. Logistically, there was still significant change that had to occur to tran sition to concurrent enrollment though, and the inte ntion of the program is different. The Concurrent Enrollment Programs Act has a clear intent of expanding access to students who are traditionally underserved. Thus, while PSEO was included as a necessary control variable, there is also an impo rtant linkage to be made to the literature on policy diffusion. Berry and Berry (2007) refer to a softening of the environment to describe wh en one innovation takes place and makes it easier for subsequent innovations to occur. The findings

PAGE 140

129 support that t heoretical aspect as schools t hat had PSEO were quicker to adopt concurrent enrollment. A related implication for practitioners and policymakers is that once a college going program is established it may not only continue to grow in terms of student participatio n year over year, but it may also soften the environment and the make the school more open to other promising practices. Prior research on the importance of establishing college going cultures in high schools would also support the claim that states puttin g resources into promoting college access programs may see an increasing return on that investment (Hoffman, Vargas & Santos, 2008a ; Roderick, Nagaoka, Coca & Moeller, 2008). Expanding Access: Charter Schools ow participation levels in concurrent enrollment is driven by their lack of access to funding, support, and programming or by a difference in the nature of charter high schools, many of which are have their own thematic focus that may not cohere with concu rrent enrollment. It could be the case that charter schools prefer to pursue other college readiness programs that are perceived to be more elite or rigorous (e.g. Advance d Placement, International Baccalaureate). Or, it may be that some charter schools prefer to par tner with more selective, four year institutions instead of community colleges, and while that is permissible under concurrent enrollment, it is more costly to do so. Another explanation could be that charter schools are not required to employ licensed tea chers, and that may decrease opportunities for offering concurrent enrollment courses at the high school if there is a shortage of instructors with the necessary qualifications The author did not collect data on such information, so further research would need to be conducted to test those speculative claims. If the low participation at charter schools is related to the lack of shared resources that could be a learning p oint for Colorado and for other states. Some districts handle portions of the logistic al and financial transactions of concurrent enrollment on behalf of thei r district operated

PAGE 141

130 schools. Not including charter schools in those arrangements could reduce the likelihood that charter schools will offer the program. Charter schools may also have lower admin istrative capacity than larger district high schools and without support from the district or another entity they may not be able to offer concurrent enrollment Ensu ring equitable arrangements within the state for all schools, regardless of charter status, could increase access for students to concurrent enrollment opportunities. Meeting Policy Goals ? ed to expand access to low income and minority students who had not typically been included in similar programs In the descriptive analysis it was noted that free and reduced price lunch ( FRL ) students were slightly overrepresente d in the concurrent enrollment population. When considering race/ethnicity, Hispanic students had higher participation rates than white students in 2011 and 2012 and were one percentage point below them in 2013. All groups have seen increased participation rates each year of the study. The event history analysis performed in this study found that even after controlling for confounding factors, high school s with higher percentages of low income students were quicker to adopt concurrent enrollment. From these results, one can conclu de that the law is functioning as intended by reaching groups of traditionally underserved students. Expanding access, however, is only on e half of the equation; ensuring successful outcomes for students is also a key goal of the policy Did the Policy Improv e O utcome s for L ow I ncome and M inority S tudents ? This s tudy explored the effects of participating in concurrent enrollment for low income and minority students The r esults indicate that all students benefit from concurrent enrollment, but low inc ome students see a n even greater increase in positive college outcomes when compared to higher income students. This is a critical finding given the policy goal of increasing outcomes for

PAGE 142

131 traditionally underserved students. It c orroborates other research that found students of low socioeconomic status (SES) see greater benefits from dual enrollment than high SES students (An, 2013). While it makes intuitive sense that students who are al ready disposed to go to college will l ikely enroll regardless of their high s chool programming, there has been a lack of evidence around what happens to students who are on the margins of attending. The U S education system has historically provided advanced coursework opportunities (e.g. Advanced Placement, International Ba ccalaureate, honors courses, and dual enrollment) to academically advanced students (Darling Hammond, 2010; Hoffman, 2005; Hoffman, Vargas, Santos, 2009a; Oakes, 2005) Now, w ith the expansion of dual enrollment, researchers are investigating the effects of providing dual enrollment to those students who are not a t the top of their classes Given the pervasive gaps in academic achievement rates by race/ethnicity and SES, expanding access to advanc ed coursework to stud ents who are middle achievers t hose who can meet the minimum academic prerequisites for dual enrollme nt but are not top performers will n aturally reach students who have typ ically been underrepresented in higher education This dissertation contribut es to the literature on c ollege access because prior research on concurrent enrollment is limited and the findings have conflict ed at times While An (2013) found dual enrollment had a greater effect on college degree attainment for low SES students when compared to high SES students nationa lly Taylor (2014) found smaller effect sizes for low income students in Illinois The findings from this study were consistent across all four college outcomes college matriculation, remedial education, college GPA, and persistence indicating that low inc ome students receive a greater benefit from participating in concurrent enrollment than wealthier students. The findings of this study, however, were not consistent across all outcomes for minority students. T he research f ound that Hispanic students wh o take concurrent enrollment courses see a

PAGE 143

132 greater impact on their likelihood of going to college than white students w ho participate in the program. This is a promising finding since the college going rate for Hispanic students lags 18 percentage point s below that of white students in Colorado (CDHE, 2017 a ). Hispanic students also lag behind their peers in college GPA, persistence and completion and given that the college success outcomes for Hispanic students were not statistically significant in this study, i t is uncle ar if the positive benefits related to college access for Hispanic students extend to college success and degree completion. When additional years of longitudinal data are available, future research should analyze the effect of the progra m on college success and degree attainment for minority and low income students. N onetheless, this study met m any of its intended goals around expanding access to concurrent enrol lment courses and incr easing college enrollment rates for both low income and Hispanic students. Burgeoning Empirical Support for Concurrent Enrollment Education researchers often struggle with controlling for selection bias due to limitations on available data and analytical methods, and prior research on concurrent enrollment specifically has left room for improvement (Allen & Dadgar, 2012; An, 2012; Le, Casillas, Robbins, & Langley, 2005). T hough concurrent enrollment programs have been around for decades, only in the past several years have researchers begun publishing quasi experimental evaluations of statewide concurrent enrollment programs, due in par t to the recent expansion of statewide longitudinal data systems. We are now seeing burgeoning evidence that concurrent enrollment has decidedly positive effects on student s, which lends support to the school improvement framework literature that contends that educa tion inputs can affect student outcome s. While t his e valuation concurrent enrollment program found positive outcomes, the effect sizes are smaller than those found in similar studies conducted in Illinois, Texas and Ut ah, but they are larger than the effect sizes found in a study conducted in Washington (see Table 26).

PAGE 144

133 All studies under comparison used a quasi experimental design to estimate the effect of concurrent enrollment participation (using a dichotomous measure) on college matriculati on rates. participation intensity by including a continuous measure of credit hours earned. L ike Giani et al. (2014) this research found increased e ffect sizes as concurrent enrollment credit hours increased. G iani et al. however, did not track college enrollment outside of the state of Texas due to limitations in their data. Cowan and Goldhaber (2014) assert t hat excluding out of state students c ould be a serious measurement error. Their analysis found that up to the half of the effect size in studies that exclude out of state matriculation is due to bias from measurement error. This dissertation included out of state students in the college matri culation analysis, which could be a reason why the effect size is smaller than other studies although further investigation would need to be conducted to confirm that claim. Table 26 Comparison of Statewide Evaluations Assessing E ffect of Dual Enrollment Programs on College Matriculation Study Citation State of Focus Increased % Probability of College Matriculation Method HS Graduation Year Cohort(s) Haskell (2016) Utah 30.7% PSM 2008 & 2009 Taylor (2015) Illinois 34.0% PSM 2003 Giani et al. (2014) Texas 36.0% PSM 2001 Cowan & Goldhaber (2014) Washington 5.9% Fixed effects regression 2010 Findings presented within this study Colorado 10.21% PSM 2011, 2012 & 2013 10.57% Fixed effects reg ression Note: All displayed results use a dichotomous treatment variable. Lastly, while rigorous statewide studies are becoming more commonplace on this education topic, as well as others, there is a constant need to ensure the sustainability of statewide longitudinal data systems. With stu dies such as the ones listed in Table 26, it is hoped that

PAGE 145

134 policymakers and practitioners will see the value in continuing to support data systems and the sharing of data with researchers. Setting a Foundation for a K 14 System? In closing, one additional implication of the findings is that dual enrollment programs can be viewed as setting a foundation for a public K 14 education system. It is a fact: a high school diploma alone Nearly all postsecondary credentials, even technical c ertificate s or s, can make an individual far more valuable in the job market As dual enrollment programs have proliferated across states, school districts and high schools, some have seen the programs as an opportunity to create a near s eamless system at the end of which students exit with public K 14 system has grown in popularity in rece nt years One report, referring to the increase in dual enrol lment programs in states notes, rhetoric of K 16, something historic is beginning to emerge in these states: the creation of an or postsecondary education 2005, 26) Shifting the public education paradigm to be one in which all students earn degree through a free and compulsory system of education may not be politically palatable on a larger scale. While it is one matter for a state to pro mote a voluntary policy that is, in many cases, fi nan ced within existing means, it is quite another for a state to create a system where the default is all students may stem would surely require more public revenue and states ar e already under severe budget constraints. Even though the return on the investment could be considered economically worthwhile, it would be a challenging political feat to accomp lish. As Paul Reville comments, at some point a fundamental shift needs to occur. We claim to want a system that educates all our students to a high level so that they can successfully participate in our high skills/high knowledge 21st century economy, thereby assuring the growth of that economy and prosperity for them and their families. But we h aven't built an engine to drive such an enterprise. We just keep tinkering with the old

PAGE 146

135 engine, trying to get it to do a job that is fundamentally different from that for which it was designed. ( Reville, 2014, 24) Perhaps concurrent enrollment i s just mo re tinkering, but it could lay the foundation for a fund amentally different system wherein all students have the opportunity to receive a rigorous secondary education and graduate wi e prepared to enter the workforce in a productive c areer or pursue additional postsecondary degrees. L imitations and Fu t u r e Research While this research makes important contributions to the field, there are several limitations that the author acknowledges. From these limitations ideas for future research may be generated. Omitted Variable Bias A significant limitation of this study is the potential for omitted variables to threaten internal validity. Individual motivation is not a directly observable variable and could have a confounding effect on the analy sis. A student, for instance, may be intrinsically motivated to both select into concurrent enrollment courses and to attend college. Other unobserved variables that could influence both selection into concurrent enrollment and the college outcomes under i counselors, and peer effects. Additionally, concurrent enrollment is one program among many intended to increase college going rates. While the school level fixed e ffects help isolate the impact programs are offered within the same high school in the same time period, it would be challenging to disentangle the effects of th e disparate programs. The presence of omitted variable bias means that the effect sizes may be overestimated. The sensitivity analysi s performed after the quasi experimental PSM estimations however, lends credibility to the practical significance of the r esults by indicating they a re robust enough to withstand the presence of moderate and even high levels of unobserved bias

PAGE 147

136 Further, t he final part of this study considered a categorical independent variable to assess the effects of concurrent enrollment c redit hour levels on college outcomes. The top two categories of credit hour levels (6 12 hours and 12+ hours) may reflect more selection bias than the lower two categories of credit hour levels (1 3 hours and 3 6 hours), because students taking a large nu mber of credit hours may be intrinsically motivated to pursue advanced educational opportunities, including higher educ ation. The lower two categories could provide a more accurate level of treatment effects, and even in viewing those findings, one can see meaningful, positive effects on college outcomes. The college matriculation outcome, in particular, is still substantively l arge even for those students taking fewer credits. Even so, future research should continue to identify ways to better cont rol for selection bias. If there is ever a n instance where a dual enrollment program is capped and students become waitlisted, there could be a n opportunity to conduct a natural experiment. Concurrent Enrollment Instructors Another area of exploration for future research that was a limitation of this current study is how teacher specific variables affect the research questions. It would be valuable to include, for example, a measure that controls for the type of postsecondary degree teachers have within hi gh schools. If a high school wants to offer concurrent enrollment courses in their building with their level math course). That level of granular data is not available for this study, but researchers with access to such data could explore how the credentialing of teachers within a high school corresponds with the adoption of concurrent enrollment as well as the intensity of implementation. In Colorado, finding high school teachers with the needed credentials to run a concurrent enrollment course is a challenge for all schools but especially for rural schools (Zinth, 2014 a ). In 2016, for example, Montezuma Cortez High School stopped offering concurrent enrollment because

PAGE 148

137 they did not have a teacher available who had the necessary credentials. Another area for future research is to follow trends in concurrent enrollment offerings at r ural high schools to see if program access declines due to barriers around teacher availability. Course Characteristics The dataset did not include information on w hether the concurrent enrollment course was taught at the high school, at a postsecondary campus, online or in a hyb rid for m a t. Taking a course on campus could have more influence on postsecondary readiness than taking a course at the given the benefits of being in a physical college environment Further, the quality of concurrent enrollment courses could vary dependent on whether the course is taug ht by high school teachers or by college faculty If courses are taught by high school teachers, quality can also be dependent upon the strength of the relationship between the high school and its partneri ng college. Despite not having access to data on course characteristics, the findings are reas suring in that there are positive outcomes occurring from the program Th e majority of s tudents participating in dual enrollment in Colorado and nationally take courses at their high school (Borden, et al., 2013) From this research and others, it is evident that taking courses on high school campuses produces benefits (An, 2013; Giani et al., 2014, Taylor, 2015). In a qualitat ive investigation, Karp (2012) found that d espite not being physically located on a college campus, dual enrollment cour ses still gave students the opportunity to assume the role of a college student, which helped them gain a deeper knowledge of what would be expected of them in college courses. Nonetheless, the existing state level research studies have not explored course setting and instructor characteristics due to limited data availability ; r esearchers w ith access to such data should consider if there are relationships between those variables and postsecondary outcomes. While positive benefits result, on avera ge, from concurrent enrol l ment participation, it would be worthwhile to see if there are

PAGE 149

138 v arying effects from taking concurrent enrollment on a college campus as compared to at a high school or online. Policy Diffusion Pressures Lastly, it is also impo rtant to note that during the time of the diffusion study (2010 2015), the economy was recovering from the Great Recession of 2008. There is reason to believe that economic factors influenced the diffusion of concurrent enrollment, particularly as communit y colleges rely on the program for consistent revenue. Enrollments at community colleges peaked during and immediately after the recession but have been declining nationwide since 2010 (Smith, 2016). Concurrent enrollment provides community colleges with r eliable revenue, and that could have factored into why the program diffused so quickly. While there were indicators included in the study to account for the prese nce of nearby community colleges, as well as year fixed effects there was not a separate control for the larger economic forces that may have affected the rapidity with which concurrent enrollment diffused. Additionally it could be beneficial to delve deeper into the emulative pressures that did or did not influence high schools in their decisions to adopt the program. The indicator f or regional diffusion (number of high schools offering concurrent enrollment within a 5 mile radius) was not statistically significant in the event history analysis model but there could be a more appropriate indicator to capture emulative pressures ( for example rival high schools based on sports associations or school choice patterns ) More recent literature on policy diffusion has found the geographic focus of policy adoption to be an outdated concept (Shipan & Volden, 2012 ; Volden, Ting & Carpernter, 2 008 ). Policymakers and leaders can learn about innovative programs not just from their geographic neighbors, but from others across the states or increasingly across the world A recent publication of the best practices in education from other countries f or example, is making its way across state legislatures; as a result, one can expect to see legislation enacted emulating

PAGE 150

139 educational practices from Finland, Singapore and other high performing countries (National Conference of State Legislatures, 2016). G eographic patterns of the diffusion of concurrent enrollment may not be as critical to investigate as other factors that influence policy learning, such as informal personal networks or competitive pressures (Binz Scharf, Lazer & Mergel, 2012; Shipan & Vol den, 2012). Principals, for instance, may influence one another through informal personal networks. The lack of a measure to account f or school leadership was a limitation of this study, although school fixed effects were used in an attempt to a ddress it It could be the case that entrepreneurial principals are more likely to push for offering concurrent enrollment in their schools. A lso, perhaps, they are more likely to influence one another. An informal network could ex ist within and across districts, and would support findings in the literature related to both policy entrepreneurs and informal learning networks (Binz Scharf, Lazer & Mergel, 2012; Mintrom, 1997). Thus, the measure used here in for the diffusion of concurrent enrollment may not have accurately captured the emulative pressures that e xisted C onsidering the rapidity with which the program spread, it stands to reason that there was some amount of external pressure high schools felt to adopt the program f or which the model d id not directly account whether that pressure was from other high schools or from community colleges motivated, at least in part, by economic factors One indicator that did reveal differences in diffusion patterns was the urban suburban indicator. High schools in urban suburban areas were far slower to adopt concur rent enrollment than Denver Metro Area schools or rural schools. In Colorado, the urban suburban designation refers mostly to Colorado S prings and Greeley. An interesting idea for future research could be to conduct a qualitative case study analysis of dis tricts within those cities to understand why they were slower to adopt, what eventually led to the program adoption, and how the concurrent enrollment programs are faring now that they are in place there

PAGE 151

140 As states continue to expand and promote college ac cess and success programs, research questions such as the ones presented in this section will be useful to practitioners and policy makers. The empirical support for concurrent enrollment is growing, but states still have a long road ahead of them to close achievement gaps. The research presented in this dissertation contributes to advancing collective knowledge about how states can improve postsecondary outcomes for students, particularly those who have been traditionally unders erved, but future research c an and should build off of these findings to provide further guidance to researchers, policymakers and practitioners

PAGE 152

141 REFERENCES Adelman, C. (2006). The toolbox revisited: Paths to degree completion from high school through college. Washington, DC: U.S. Department of Education. Allard, S.W. (2004). Competitive pressures and the emergence o United States. Policy Studies Journal 32 521 544. Allen, D. (2010). Dual Enrollment A Comprehensive Literature Review & Bibliography Ne w York: City University of New York. Allen, D., & Dadgar, M. (2012). Does Dual Enrollment Increas Evidence from a Quasi Experimental Analysis of Dual Enrollment in New York City. New Directions for Higher Education 158, 11 19. Allison, P. D. (1984). Event history analysis: Regression for longitudinal event data Beverly Hills, CA: Sage. An, B.P. (2013). The Impact of Dual Enrollment on College Degree Attainment: Do Low SES Students Benefit? Educational Evaluation and Poli cy Analysis 35 (1): 57 75. Angrist J D & Krueger A B. ( 1992 ) Estimating the payoff to school ing using the Vietnam era draft lottery NBER Working Paper 4067, National Bureau of Economic Research Cambridge, MA. Arbuckle, J. L. (1996). Full information e stimation in the presence of incomplete data. Advanced structural equation modeling: Issues and techniques 243 : 277 Attewell P. & Lavin D.E. (2007). Passing the Torch: Does Higher Educa tion for the Disadvantaged Pay Off Across the Generations? New York: Russell Sage Found. Attinasi, L. C. (1989). Getting In: Mexican tions of University Attendance and the Implications for Freshman Year Persistence. Journal of Higher Education 60 : 247 277. Austin, P. C. (2009). Balance Diagnostics for Com paring the Distribution of Baseline Covariates between Treatment Groups in Propensity Score Matched Samples. Statistics in Medicine 28 : 3083 107. Austin, P.D. (2011).Optimal Caliper Widths for Propensity Score Matching When Estimating Differences in Mean s and Differences in Proportions in Observational Studies. Pharmaceutical Statistics10 : 150 61. Bailey, T. R., and Karp, M. M. (2003). Promoting College Access and S uccess: A Review of Dual Credit and Other High School/College Transition Programs Washin gton, DC: U.S. Department. Baldridge, J.V. and Burnham, R.A. (1975). Organiza tional Innovation: Individual, Organizational, and Environmental Impacts. Administrative Science Quarterly 20 (June), 165 176.

PAGE 153

142 Ba l la, Steven J. ( 2001 ). Interstate Professiona l Associations and the D iffusion of Policy Innovations. American Politics Research 29 (3): 221 45. Baser, O. (2006). Too much ado about propensity score models? Compar ing methods of propensity score matching. Val and Outcomes Research 9 (6), 377 85. doi:10.1111/j.1524 4733.2006.00130.x B aybeck, B. Berry, W. D., & Siegel D. A. ( 2011 ) A Strategic Th eory of Policy Diff usion via Intergove rnmental Competition. Journal of Politics 73 (1): 232 47. Berry, F. S. (1994a). Innovation in Public Management: The Adoption of Strategic Planning Public Administration Review 54 (4), 322 330. Berry, F. S. (1994b). Sizing Up State Policy Innovation Resea rch. Policy Studies Journal 22 (3), 442 456. doi:10.1111/j.1541 0072.1994.tb01480.x. Berry, F. S., & Berry, W. D. (1990). State Lottery Adoptions as Polic y Innovations: An Event History Analysis. The American Political Science Review 84 (2), 395 415. Ber ry, F. S., & Berry, W. D. (1992). Tax Innovation in the States: Capital izing on Political Opportunity. American Journal of Political Science 36 (3), 715 742. Berry, F. S., and Berry W.D. (2007). Innovation and Diffusion Mo dels in Policy Research. In P. A. Sabatier (Ed.), Theories of the Policy Process (2 nd ed., pp. 223 260). Boulder, CO: Westview. Bettinger, E. P., & Long, B. (2005). Remediation at the Community College: Student Participation and Outcomes. New Directions For Community Colleges 129 : 17 26. Bingham, R. D. (1977). The Diffusion of Innovation Among Local Governments. Urban Affairs Review 13 (2), 223 232. doi:10.1177/107808747701300205 Binz Scharf, M. C., Lazer, D. M., and Mergel I ( 2012 ) Searchin g for Answers: Networks of Practice among P ublic Administrators. American Review of Public Administration 42 (2): 202 25. Biondi Zoccai, G., Romagnoli, E., Agostoni, P., Capodanno, D., Castagno, D. G., Modena, M.G. (2011). Are propensity scores really superior to standard multivariable analysis? Contemporary Clinical Trials 32 (5), 731 40. Blanco, C. (2006). Policy Implications. In WICHE (Ed.) Accele rated Learning Options: Moving the Needle on Access and Success (pp. 73 82). Boulder: Western Interstate Commission for Hig her Education. Borden, V. M. H., Taylor, J. L., Park, E., & Seiler, D. J. (2013). Dual credit in U.S. higher education: A study of state policy and quality assurance practices Chicago: Higher Learning Commission. Braddock, J. & McPartland, J. M. (1993) Education of ea rly adolescents. In L. Darling Hammond (Ed.), Review of research in education (Vol. 19, pp. 135 170). Washington, DC: American Educational Research Association.

PAGE 154

143 Brand, J. E. (2010). Civic returns to higher education. Soc iological Forces 8 9 :417 34. Brand, J. E., & Halaby, C. N. (2006). Regression and matching est imates of the effects of elite college attendance on educational and career achievement. Social Science Research 35 : 749 770. doi:10.1016/j.ssresearch.2005.06.006 Brand, J.E. & Xi e, Y. (2010). Who benefits most from college? Evid ence for negative selection in heterogeneous economic returns to higher education. Am. Sociol. Rev. 75 :273 302 Buckley, J. and Westerland, C. (2004). Duration Dependence, F unctional Form, and Corrected Sta ndard Errors: Improving EHA Models of State Policy Diffusion. State Politics and Policy Quarterly 4 (1), 94 113. Byrd, K. L., and G. MacDonald. (2005). Defining College Readin ess from the Inside Out: First Generation College Student Perspectives. Community College Review 33 : 22 37 Caliendo, M., & Kopeinig, S. (2008). Some practical guid ance for the implementation of propensity score matching. Journal of Economic Surveys 22 (1), 31 72. Carnevale, A.P., Smith, N. & Strohl, J. (2010). Help Wanted: Pro jections of Jobs and Education Requirements through 2018 Washington, D.C.: The Georgetown University Center on Education and the Workforce. Carnevale, A.P., Smith, N. & Strohl, J. (2013). Reco very: Job Growth and Education Requirements through 2020 Washington, D.C.: The Georgetown University Center on Education and the Workforce. Centra, J. A. & Potter, D. A. (1980). School and teacher effe cts: An interrelational model. Review of Educational Research, 50 (2), 273 291. Chubb, J. E. & Moe, T. M. (1990). Politics, Washington, D.C.: The Brookings Institution. Coleman, J. S., Campbell, E. Q., Hobson, C. J., McPartland, J., Mood, A., Weinfeld, F. D., and York, R. L. (1966). Equality of Educational Opportunity Washington, D.C.: U.S. Gov ernment Printing Office. Colorado Department of Education. (2010). HB 09 131 9 and SB09 285 Fact sheet: The concurrent enrollment programs act of 2009 Denver: CDE. Colorado Department of Higher Education. (2014 ). Annual report on concurrent enrollment 2012 2013 school year Denver: CDHE. Colorado Department of Higher Education. (201 6 ). 201 5 Leg islative report on remedial educa tion Denver: CDHE. Colorado Depart ment of Higher Education. (2017 a ). 2017 Legislative report on the postsecondary progress and success of high school graduates Denver: CDHE.

PAGE 155

144 Colorado Depar tment of Higher Education. (2017 b ). Annual report on concurrent enrollment 2015 2016 school year Denver: CDHE. Colorado Department of Education. (2016). Concurrent Enrollment Policy & Example Practices Denver: CDE. Retrieved from: https://www.cde.state.co.us/postsecondary/ce_poilcypractice Cook T.D., Shadish W.R., & Wong V.C. (2008). Three conditio ns under which experiments and observational studies produce compara ble causal estimates: new findings from within study comparisons. Journal of Policy Analysis and Management 27 :724 750 Conley, D. 2005. College Knowledge San Francisco: Jossey Bass. Conley, D. 2007. Redefining College Readiness, Volume 3 Eugene, OR: Educational Policy I mprovement Center. Conley, D. 2010. College and Career Ready. San Francisco: Jossey Bass. Conley, D. (2013). Rethinking the Notion of 'Noncognitive'. Education Week 32 (18), 20 21. Cowan, J., & Goldhaber, D. (2014). provide students? (CEDR Working Paper 2014 7). Seattle: University of Washington. Level Lear ning for High School Students: A Comprehensive National Policy Study and University of New York at Buffalo. Cyert, R.M. & March, J. G. (1973). A Behavioral theory of the firm. Englewood Cliffs, NJ.: Prentice Hall. Darling Hammond, L. (2010). The Flat World and Educatio n: How America's Commitment to Equity Will Determine Our Futures New York: Teachers College Press. Dehejia, R. H., & Wahba, S. (1999). Causal Effects in Nonexpe rimental Studies: Reevaluating the Evaluation of Training Programs. Journal of the American Stati stical Association 94 (448), 1053 1062. doi:10.1080/01621459.1999.10473858 Dehejia, R., & Wahba, S. (2002). Propensity score matching meth ods for nonexperimental causal studies. Review of Economics and Statistics 84 (1), 151 161. Dickie, L. O., and J. E. F arrell. (1991). The Transition f rom High School to College: An Impedance Mismatch? The Physics Teacher 29 (7): 440 445. Dreeben, R. (1987). Closing the divide: What teachers and admin istrators can do to help Black students reach their reading potential, Am erican Educator, 11 (4): 28 35. Eckstrom, R. & Villegas, A. M. (1991). Ability grouping in middle grade mathematics: Process and consequences. Research in Middle Level Education, 15 (1): 1 20.

PAGE 156

145 Education Commission of the States [ECS]. (2016). ECS State Po licy Database. Denver, CO. Retrieved February, 11, 201 7 Educational Policy Improvement Center. (2013). Metacognitive Learning Skills Retreived from https://www.epiconline.org/Issues/metacognition/ Finley, M. K. (1984). Teachers and tracking in a comp rehensive high school. Sociology of Education 57 : 233 243. Fry, R. (2011). Hispanic college enrollment spikes, narrowing gaps with other groups Washington, D.C.: Pew Hispanic Center. Funkhouser, E. (2009). The effect of kindergarten classroom size red uction on second grade student achievement: Evidence from California. Economics of Education Review, 28 (3), 403 414. Garcia, E. (1993). Language, culture, and education. In L. Darling Hammond (Ed.), Review of research in education (Vol. 19, pp. 51 98). Wa shington, DC: American Educational Research Association. Garrido, M. M., Kelley, A. S., Paris, J., Roza, K., Meier, D. E., Morrison, R S., & Aldridge, M. D. (2014). Methods for constructing and assessing propensity scores. Health Services Research 49 (5) 1701 20. doi:10.1111/1475 6773.12182 Giani, M., Alexander, C., & Reyes, P. (2014). Exploring variation in the impact of dual credit coursework on postsecondary outcomes: A quasi experimen tal analysis of Texas students. The High School Journal 97 :200 218 doi:10.1353/hsj.2014.0007 Gilardi, F., Fglister, K., & Luyet, S. (2009). Learning from Others: The Diffusion of Hospital Financing Reforms in OECD Countries. Comparative Political Studies 42 (4): 549 73. Good, T. L. & Brophy, J. (1987). Looking in cla ssrooms New York: Harper and Row. Guo, S. & Fraser, M.W. (2010). Propensity Score Ana lysis: Statistical Methods and Applications (Advanced Quantitative Techniques in the Social Sciences Series) Sage Publications, Inc. Los Angeles, California. Hanushek, E. A. (1979). Conceptual and Empirical Issues in the Estimation of Educational Production Functions. The Journal of Human Resources, 14 (3), 351 388. investigation of the Tennessee STAR experiment and from other investigations of class size effects. Education Evaluation and Policy Analysis, 21 (2), 143 163. Hanushek, E. A. (2003) The Failure of In put based Schooling Policies. The Economic Journal, 113 64 98. Hausman, J., & Taylor, W. (1981). Panel data and unobservable individual effects. Econometrica: Journal of the Econometric Society, 49 (6), 1377 1398. Retrieved from http://www.jstor.org/stab le/10.2307/1911406

PAGE 157

146 Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrica, 47 (1): 153 161. Heckman, J., Ichimura, H., and Todd, P. (1997). Matchi ng as an econometric evaluation estimator: Evidence from evaluating a job trainin g programme. Review of Economic Studes 64 (4): 605 654. Heller, D. (1997). Student price response in higher educ ation: An update to Leslie and Brinkman. Journal of Higher Education 68 (6), pp. 624 659. Heller, D. E. (2006). State Support of Higher Educatio n: Past, Present, and Future In D. M. Priest, & E. P. St. John (Eds.), Privatization and Public Universities (pp. 1 7). Bloomington: Indiana University Press. Hernandez, D.J. (2011). How Third Grade Reading Skills and Poverty Influence High School Gradua tion Baltimore, MD: Annie E. Casey Foundation. Hofmann, E. (2012). Why dual enrollment? In Hofmann, E. & Voloch, D., eds., New Directions for Higher Education no. 158, San Francisco, CA: Jossey Bass. Hoffman, N. (2005). Add and subtract: Dual enrollmen t as a state strategy t o increase postsecondary success for underrepresented students. Boston, MA: Jobs for the Future Hoffman, N., Vargas, J., & Santos, J. (2008 a ). On ramp to college: dual enrollment Boston, MA: Jobs for the Future. Hoffman, N., Vargas, J. and Santos, J. (2008 b ), Blending high sch ool and college: Rethinking the transition. New Directions for Higher Education, 2008 : 15 25. doi: 10.1002/he.322 Hout, M. (2012). Social and economic returns to college educat ion in the United States. Annual Review of Sociology 38 (1), 379 400. doi:10.1146/annurev.soc.012809. 102503 Imbens, G. (2004) Nonparametric estimation of average trea tment effects under exogeneity: a review. Review of Economics and Statistics 86 (1): 4 29 Hoyman, M., & Weinberg, M. (2006). The Process of Policy Innovation : Prison Sitings in Rural North Carolina. Policy Studies Journal 34 (1), 95 112. doi:10.1111/j.1541 0072.2006.00147.x Jobs for the Future. 2006. Dual Enrollment in Rhode Island: Opportun ities for State Policy. Rep ort to the Statewide PK 16 Council Boston: Jobs for the Future. Ka, S. & Teske, P. (2002). Ideology and Professionalism: Electricity Regulation and Deregulation over American Politics Research 30 : 323 343. Kahlenberg, R. D. (Ed.) (2004) America's untapped resource: Low income students in higher education New York: Century Foundation Press. Kane, T., & Rouse, C. (1995). Labor market returns to two and four year college. The American Economic Re view, 85 (3), 600 614. Retrieved from http://www.jstor.org/stable/10.2307/2118190

PAGE 158

147 Karch, A ( 2006 ) National Intervention and the Di ffusion of Policy Innovations. American Politics Quarterly 34 (4 ): 403 26. readiness strategy. New Directions for Higher Education 158, 21 28. doi: 10.1002/he.20011 Karp, M. M., Bailey, T. R., Hughes, K. L., & Fermin, B. J. (2005). Update to state dual enrollment policies: Addressing access and quality New York: Columbia University, Teachers College, Community College Research Center. Karp, M. M., Calcagno, J. C., Hughes, K. L., Jeong, D. W., & Bailey, T. R. (2007). The pos tsecondary achievement of participants in dual enrollment: An analysis of student outcomes in two states St. Paul, MN: University of Minnesota. Kingston, P. W., Hubbard R., Lapp B., Schroeder P., Wilson J. (2003). Why education matters. Sociology of Educ ation 76 :53 70. Kirst, M.W., & Venezia, A. (2004). From high school to college: Improving opportunities for success in postsecondary education. San Francisco: Jossey Bass, An Imprint of Wiley. Lange F. & Topel R. (2006). The social value of education an d human capital. Handbook of the Economics of Education 1 :459 509. Le, H., Casillas, A., Robbins, S., & Langley, R. (2005). Motivationa l and skills, social, and self management predictors of college outcomes: Constructing the Student Readiness Inventory. Educational and Psychological Measurement, 65 : 482 508. Lee, J. (2014). The mechanism of state charter school policy adoption: The case of American states. International Journal of Public Administration 37 :1, 20 34. Lerner, J. B. & Brand, B. (2006). The college ladder: Linki ng secondary and postsecondary education for success for all students Washington, DC: American Youth Policy Forum. Leslie, L. L., & Brinkman, P. T. (1987). Student price response in higher education. Journal of Higher Education, 58 181 204. Lopez, J. (2011). 20 Education Coordinating Council: 2007 2010 Denv er: P 20 Education Coordinating Council. Lubell, M., Schneider, M., Sholz, J. T., & Mete, M. (2002). Watershed Partnerships and the Emergence of Collective Action Institutions. American Journal of Political Science 46 (1), 148 163. Makse, T. & Volden C ( 2011 ) The Role o f Policy Attributes in the Diff usion of Innovations. Journal of Politics 73 (1): 108 24. Margolis, H., & McCabe, P. P. (2004). Self efficacy a key to improving the motivation of struggling learners. Clearing House, 77 (6), 241.

PAGE 159

148 Martin, T. C. (2013). Cognitive and noncognitive college rea diness of participants in three concurrent enrollment programs. Community College Journal of Research and Pract ice, 37 (9), 704 718 Mas A. & Moretti, E. (2009). Peers at work. American Economic Review 99 :112 45. Maurin E. & McNally S. (2008). Vive la revolution! Long term educ ational returns of 1968 to the angry students. Journal of Labor Economics 26 :1 33 Mitch ell, M.N. and Chen X. 2005. Visualizing main effects and interactions for binary logit model. Stata Journal 5(1): 64 82. Mintrom, M. (1997). Policy Entrepreneurs and the Diffusion of Innovation. American Journal of Political Science 41 (3), 738 770. Min trom, M., & Vergari, S. (1998). Policy Networks and Innovati on Diffusion: The Case of State Education Reforms. The Journal of Politics 60 (1), 126 148. Mohr, L. (1969). Determinants of Innovation in Organizations. American Political Science Review 63 (1), 111 126. Mokher, C. G., & Mclen Event History Analysis of State Adoption of Dual Enrollment Policies. American Journal of Education 115 (2), 249 277. Morest, V.S., and Karp, M.M. (2006). Twice the Credit, Half the Tim e? The Growth of Dual Credit at Community Colleges and High Schools. In Bailey, T.R.& Morest V.S. (Eds.) Defending the Community College Equity Agenda Baltimore: Johns Hopkins University Press.. Moretti E. (2012). The New Geography of Jobs New York: Ho ughton Mifflin Harcourt. Morrison, M. C. (2008). The benefits of acceleration: Graduation advantages Mason City, IA: North Iowa Area Community College. Mosteller, F. (1995). The Tennessee study of class size in the early school grades. The Future of Ch ildren, 5 (2), 113 127. National Center for Higher Education Management Systems. (2013, February 1). Presentation to the Colorado Commission on Higher Education, February 1, 2013. National Concurrent Enrollment Partnership [NACEP]. (2011). NACEP Concurre nt Enrollment Partnership Standards Chapel Hill, NC. Retrieved from: http://nacep.org/docs/standards/NACEP Standards 2011.pdf National Concurrent Enrollment Partnership [NACEP]. (20 17). NACEP in State Policy Chapel Hill, NC. National Conference of State Legislatures. (2016). No Time to Lo se: How to Build a World Class Education System State by State. Denver, CO: NCSL.

PAGE 160

149 National Student Clearinghouse. (2013). Clearinghouse Facts. Ret rieved from http://www.studentclearinghouse.org/about/clearinghouse_facts.php Oakes, J. (1993). Ability grouping, tracking, and within scho ol segregation in the San Jose Unified Shool District Los Angeles: UCLA. Oakes, J. (2005). Keeping track: How scho ol structure inequality (2 nd ed). New Haven, CT: Yale University Press Peikes, D. N., Moreno, L., & Orzol, S. M. (2008). Propensity score matching: A note of caution for evaluators of social programs. The American Statistician, 62 (3), 222 231. Pelavin, S. H. & Kane, M. (1990). Changing the odds: Factors increasing access to college New York: College Entrance Examination Board. Peterson, P. (1989). Remediation is no remedy. Educational Leadership, 46 (60), 24 25. Purkey, S. C., & Smith, M. S. (1983). Ef fective Schools: A Review. The Elementary School Journal, 83 (4), 426 452. Putnam, R. D. (1995). Tuning In, Tuning Out: The Strange Disappearance of Social Capital in America. Political Science and Politics 28 (4):664 83. Radunzel, J. & Noble, J. (2012). P redicting Long Term College Success through Degree Completion Using ACT Composite Score, ACT Benchmarks, and High School Grade Point Average Iowa City, IA: ACT, Inc. Reville, P. (2014). Stop the tinkering: We need a new K 12 engine Education Week Apri l 23, 28. Robbins, S., Allen, J., Casillas, A., Peterson, C., & Le, H. (2006 ). Unraveling the differential effects of motivational and skills, social, and self manage ment measures from traditional predictors of college outcomes. Journal of Educational Psy chology, 98 : 598 616. Robbins, S., Lauver, K., Le, H., Langley, R., Davis, D., & Carlstrom, A. (2004). Do psychosocial and study skill factors predict college out comes? A meta analysis. Psychological Bulletin, 130 : 261 288. Rogers, E.M. (2003). Diffusio n of Innovations (5 th Ed.). New York: The Free Press. Roksa, J. (2008). Structuring access to higher education: T he role of differentiation and privatization. Research in Social Stratification and Mobility 26 (1), 57 75. doi:10.1016/j.rssm.2007.11.002 Ro senbaum, P.R. (2002). Observational Studies (2 nd ed.). New York: Springer. Rosenbaum, P.R. & Rubin, D.R. (1985). Constructing a Co ntrol Group using Multivariate Matched Sampling Methods That Incorporate the Propensity Score. The American Statistician 39 (1 ): 33 8. Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66 :688 701.

PAGE 161

150 Rubin, D. B. (1997). Estimating causal effects from large data sets using propensity scores. Ann als of Internal Medicine, 127 (8), 757 763. Rutschow, E. Z., & Schneider, E. (2011). Unlocking the gate : What we know about improving developmental education New York: MDRC. Sawyer, R. (2010). Usefulness of high school average an d ACT scores in making col lege admission decisions. (ACT Research Report No. 2010 2). Iowa City, IA: ACT, Inc. Schneider, B., Carnoy, M., Kilpatrick, J., Schmidt, W. H., & Shavelson, R. J. (2007). Estimating Causal Effects Using Experimentaland Observational Designs Washington, D .C.: American Educational Research Association. Shadish W.R., Clark M.H., & Steiner P.M. (2008). Can n onrandomized experiments yield accurate answers? A randomized experiment comparing random to nonrandom assignment. Journal of the American Statistical As sociation 103 :1334 1343. Shadish, W. R., & Steiner, P. M. (2010). A Primer on Propensity Score Analysis. Newborn and Infant Nursing Reviews, 10 (1), 19 26. doi:10.1053/j.nainr.2009.12.010 Shah B.R., Laupacis A., Hux J.E., & Austin P.C. (2005). Propens ity score methods gave similar results to traditional regression modeling in observational studies: a systematic review. Journal of Clinical Epidemiology 58 :550 559. Shields, N. (2002). Anticipatory Socialization, Adjustment to University Life, and Perceived Stress: Generational and Sibling Effects. Social Psychology of Education 5 : 365 392. Shipan, C. R., & Volden, C. (2008). The Mechanisms of Policy Diffusion. American Journal of Political Science, 52 (4), 840 857. Shipan, C. R. & Volden, C. (2012). Policy Diffusion: Seven Lessons for Scholars and Practitioners. Public Administration Review 72 (6), 788 796. doi:10.111/j.1540 6210.2012.02610.x.788 Singleton, R. A. Jr., and Straits, B. C. (2010). Approaches to Social Research (5 th ed.). New York, NY: Oxford University Press. Smith, A. A. (2016). Community College Enrollments Drop Inside Higher Ed. Oct. 18. Retrieved from: www.insidehighered.com/news/2016/10/18/community colleges examining low and stagnant enrollments Smith, J. A., & Todd, P. E. (2005). D oes matching ove nonexperimental estimators? Journal of Econometrics 125 : 305 353). doi:10.1016/j.jeconom.2004.04.011 Soule, S. A. and Earl, J. (2001). The Enactment of State Leve l Hate Crime Law in the United States: Intrastate and Interstate Factors. Sociological Perspectives 44 281 305. Speroni, C. (2011). Role of Advanced Placement and Dual Enrollment Programs New York, NY: National Center for Postsecondary Research, Teachers College, Columbia University.

PAGE 162

151 St. John, E. P. (1990). Price response in enrollment decisions: An analysis of the high school and beyond sophomore cohort. Research in Higher Education, 31 (2), 161 176. Stream, C. (1999). Health reform in the states: A model of stat e small group health insurance market reform. Political Research Quarterly 53 499 525. Strong American Schools. (2008). Diploma to Nowhere Wa shington D.C.: Strong American Schools. Swanson, J. (2008). An analysis of the impact of high school dual enrol lment course participation on post secondary academic success, persistence and degree completion Paper presented at the meeting of the National Association for Gifted Children, Tampa, FL and the National Alliance of Concurrent Enrollment Partnerships, Kan sas City, MO Talbert, J. E. (1990). Teacher tracking: Exacerbating inequalities in the high school Stanford, CA: Center for Research on the Context of Secondary Teaching, Stanford University. Terenzini, P. T., Cabrera, A. F., & Bernal, E. M. (2001). Sw imming Against American Higher Education New York City: College Entrance Examination Board. Titus, M. A. (2007). Detecting selection bias, using propensity score matching, and estimating treatment effects: An application to the private returns t Research in Higher Education 48 (4), 487 521. doi:10.1007/sl U.S. Department of Education, Institute of Education Sciences, What Works Clearinghouse. (2017). Transition to College intervention report: Dual Enrollment Programs Retrie ved from https://whatworks.ed.gov U.S. Department of Education, National Center for Education St atistics. (2006) Second Follow up, 2006. Education Longitudinal Study of 2002 (ELS:2002). Venezia, A., Kirst, M., & Antonio, A. (2003). Betraying the college dr eam: How disconnected K 12 and post secondary education systems undermine student aspirations Palo Alto, CA: The Bridge Project, Stanford Institute for Higher Education Research. Volden, C. (2006). States as Policy Laboratories: Emulating Success in th Program. American Journal of Political Science 50 (2), 294 312. Volden, C., Ting, M. M., & Carpenter, D.P. (2008). A Formal Model of Learning and Policy Diffusion. American Political Science Review 102 (3): 319 32. Walberg, H. (1982) What makes schooling effective. Contemporary Education: A Journal of Review, 1 22 34. Walker, J. L. (1969). The Diffusion of Innovations among the American States. The American Political Science Review 63 (3), 880 899. Ward, D. S. & Vargas, J (2012). Using Dual Enrollment Policy to Improve College & Career Readiness. A Web Tool for Decision Makers Boston, MA: Jobs For the Future.

PAGE 163

152 Welch, S. & Thompson, K. ( 1980 ) The Impact of Federal Incentiv es on State Policy Innovation. American Journal o f Political Science 24 (4): 715 26. Wenglinsky, H. (2002). How schools matter: The link between t eacher classroom practices and student academic performance. Educational Policy Analysis Archives, 10 (2). Williams, R. (2012). Using the margins command to es timate and interpret adjusted predictions and marginal effects. Stata Journal 12 : 308 331. Winship, C., & Morgan, S. L. (1999). The estimation of causal effects from observational data. Annual Review of Sociology, 25 659 707. Wong, K. K., & Shen, F. X. (2002). Politics of State Led Reform in Education: Market Competition and Electoral Dynamics Educational Policy (Vol. 16, pp. 161 192). doi:10.1177/0895904802016001009 Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data 2nd ed Cambridge, MA: MIT Press. Zinth, J. D. (2014a). Dual enrollment: A strategy to improve college going and college completion among rural students. Denver, CO: Education Commission of the States. Zinth, J. D. (2014b). Increasing Student Access and Succe ss in Dual Enrollment Programs: 13 Model State Level Policy Components. Denver, CO: Education Commission of the States. Zinth, J. D. (2015a). Dual Enrollment Course Content and Instructor Quality Denver, CO: Education Commission of the States. Zinth, J D. (2015b). State Approaches to Funding Dual Enrollment Denver, CO: Education Commission of the States. Ziomek, R. L., & Svec, J. C. (1995). High school grades and achievemen t: Evidence of grade Inflation (Research Rep. No. 95 3). Iowa City, IA: ACT.