CALCULATING THE EFFECT OF LENGTH
BIASED SAMPLING ON SCREENDETECTED
CASES IN RANDOMIZED CONTROLLED
SCREENING TRIALS
by
Jean McKenzie Ethredge
B.S.Ed., University of Georgia, 1972
B.S., Metropolitan State College, 1992
A thesis submitted to the
University of Colorado at Denver
in partial fulfillment
of the requirements for the degree of
Master of Science
Applied Mathematics
December 2000
This thesis for the Master of Science
degree by
Jean McKenzie Ethredge
has been approved
by
3o f\]ov
Date
Ethredge, Jean McKenzie (M.S., Applied Mathematics)
Calculating the Effect of Length Biased Sampling on Screendetected Cases
in Randomized Controlled Screening Trials
Thesis directed by Professor Karen Kafadar
ABSTRACT
Screening tests are used frequently for the early detection of diseases
such as cancer. The benefit gained by these early detection tests can be assessed
by comparing the mortality rates or survival rates between a study arm and
a control arm in a randomized controlled screening trial. The comparisoh of
the two arms is affected by two important biases: lead time bias and length
bias. Lead time is the amount of time by which a preclinical diagnosis is
advanced over a clinical diagnosis as a result of the screening test. If survival
fs measured from the time of diagnosis, then the comparison between study
and control arms is biased by the lead time. Trial results affected by a lead
time bias can indicate an increase in survival time of a study arm over a
control arm even when there is no' benefit at all from early screening. This bias
can be eliminated by the use of randomized controlled screening trials wljere
survival is measured since the entry into the study. Lengthbiased sampling,
Caused by periodic screening, leads to an overrepresentation in the study arm
of cases having slower growing disease with better prognosis than in the general
m
population. This phenomenon can also lead to higher survival rates in the
study arm versus the control arm, even in the absence of any screening benefit;
but, unlike lead time bias, can not be eliminated by the study design as in a
randomized screening trial.
In this thesis, I calculate the mean and variance of the increase in
survival times that arise because of length biased sampling, when the sojourn
times are gamma distributed. I further show that by ignoring the length bias,
screening will appear more beneficial than it actually is. This bias must be
considered to avoid overoptimistic conclusions about the benefit of screeiiing
programs.
This abstract accurately represents the content of the candidates thesis. I
fecommend its publication.
Signed
Karen Kafadar
IV
DEDICATION
To Jack and to our four daughters: Kristen, Alicia, Elizabeth and Jacquelyn.
ACKNOWLEDGMENTS
It has been an honor and privilege to have done the calculations and
research to complete this thesis under the direction of Karen Kafadar, my
graduate adviser. I would like to thank her for her encouragement, patience,
guidance and generosity.
I would also like to thank William Briggs for his meticulous reading
of my thesis and his suggestions for improvement. Further thanks go to Cfaig
Johns for serving on my committee and suggesting that I work with Kareh in
the first place.
Throughout the years of my mathematical studies, I have received
Valuable advice and assistance from Thomas Kelley, an undergraduate professor
and from my friend Dave Brown. I would like to extend my appreciation to
them both. Additional thanks go to Cary Miller and Randy Chase for teaching
me enough about UNIX to make this thesis possible. To these and others, I
extend my thanks.
Finally, I thank Jack who helped maintain our family and home in so
many ways so Mom could pursue math.
I thank you all.
CONTENTS
Figures................................................................ ix
Tables................................................................. x
1. Introduction........................................................ 1
1.1 Natural History of Cancer......................................... 2
1.2 The Screening Process............................................. 3
1.3 Lead Time Bias.................................................... 3
f.4 Length Bias....................................................... 4
1.5 The Accuracy of Screening Tests .................................. 7
2. The Assessment of Screening Programs................................ 9
2.1 The HIP Breast Cancer
Random Control Study............................................. 10
2.2 Detection of Breast Cancers of the HIP Study..................... 12
2.3 Mortality in the HIP Trial....................................... 12
2.4 Conclusion of the HIP Trial...................................... 15
2.5 The CNBSS2 Breast Cancer Random Control Study................... 16
2.6 Detection of Breast Cancers by CNBSS............................. 17
2.7 Mortality in the CNBSS2 Trial................................... 19
2.8 Conclusion of CNBSS2 Trial...................................... 20
3. The Exponential Distribution of Sojourn Times....................... 22
4. The Gamma Distribution of Sojourn Times............................. 38
vii
4.1 The Gamma Distribution for Sojourn Times when r = 2............ 40
4.2 The Gamma Distribution for Sojourn Times for r = 4 ............ 49
5. Conclusion......................................................... 52
Notation Index........................................................ 53
Appendix
A..................................................................... 56
References............................................................ 63
vm
FIGURES
1.1 Natural History of Cancer Progression.................... 2
1.2 Lead Time Bias........................................... 4
1.3 LengthBiased Sampling or Length Bias.................... 6
IX
TABLES
2.1 HIP Compliance of the Study Arm by Number of Screening
Exams [9, p. 19].......................................... 12
2.2 HIP: Cumulative Detection of Breast Cancers by Year and
Arm [9, p. 50]............................................ 13
2.3 HIP: Cumulative Cases of Mortality Due to Breast Cancer . 14
2.4 HIP: Survival Rates [9, p. 71]................. 15
2.5 CNBSS: Detection of Invasive Cancers [5, p. 1492]......... 18
2.6 CNBSS: Mortality Cases by Cause to the End of 1993 by
Study Arm [5, p. 1494] ................................... 19
2.7 CCBSS: Cumulative Deaths due to Breast Cancer............. 20
3.1 Exponential Distribution: Means for j = 1,2,3,4 ......... 29
3.2 E(Y*)\ The Mean Preclinical Duration of Cases Eligible for
Detection by Screening....................................... 30
3.3 Exponential Distribution Ratio Comparison: E(Y*)/E(Y) . 31
3.4 Relative Increase in the Mean: (E(Y*)/fi 1)T00%......... 31
3.5 Var(Y*)\ Variances of Sojourn Times of Cases Eligible for
Detection by Screening....................................... 34
3.6 SD(Y*): Standard Deviations of Sojourn Times of Cases El
igible for Detection by Screening ....................... 34
3.7 Exponential Distribution Ratio (SD (Y*))/ji................ 35
x
3.8 Relative Error in the Standard Deviations ...................... 36
3.9 Exponential Distribution Summary of Y* when r = 1 .... 37
4.1 Gamma Distribution r = 2 Means for j 1, 2, 3,4........... 42
4.2 Gamma r = 2,E(Y*): The Mean Preclinical Duration of
Cases Eligible for Detection by Screening with /3 = 0.95 ... 42
4.3 Gamma r = 2,E(Y*): The Mean Preclinical Duration of
Cases Eligible for Detection by Screening with /3 0.95 ... 43
4.4 Values of E(Y*)//j, when fy* (y) is a Gamma Distribution r =
2............................................................... 43
4.5 Gamma Distribution Ratio when r = 2 E(Y*)//j,.............. 44
4.6 Gamma Distribution Relative Increase in the Means.......... 44
4.7 Gamma Formulas for the Expected Value of Yf* when r = 2 45
4.8 Gamma Formulas for the Expected Value Squared when r =
2 .............................................................. 46
4.9 Gamma Distribution Variances for r 2, j 1, 2, 3,4 . . . . 47
4.10 Gamma Distribution r = 2, f3 = 0.95 Var(Y*) ................. 47
4.11 Gamma Distribution r = 2, Standard Deviations of Y* ... 48
4.12 Gamma Distribution r = 2 Ratios of Standard Deviations . 48
4.13 Gamma Distribution r 2; Relative Error ( %) in Standard
Deviations of Y* ............................................... 49
4.14 Summary of Y* when r = 2: The Means, Relative Increase,
Variances, Standard Deviations, and Relative Errors........... 50
4.15 Summary of Y* when r = 4: The Means, Relative Increase,
Variances, Standard Deviations, and Relative Errors........... 51
xi
A.l Exponential Distribution Ratios E(Y*)/fj,..................... 57
A.2 Exponential Distribution Variances: Var(YJ*) ................. 57
A.3 Exponential Distribution Standard Deviations for Y*........... 58
A.4 Exponential Distribution (Standard Deviations (Y/))/// 58
A.5 Gamma Distribution Means r = 4 for j = 1,2,3,4................ 59
A.6 Gamma Distribution Ratio: when r = 4 60
A.7 Gamma Distribution r = 4 AIDA................................. 60
A.8 r = 4 Variances for E(Y?) for j = 1, 2, 3,4 61
A.9 r = 4 Ratio of Sampled Standard Deviations/^ ................. 62
A. 10 Gamma Distribution r = 4 Ratio Standard Deviations of
Y*/a........................................................ 62
xii
1. Introduction
The expression a stitch in time, saves nine parallels a common
belief that therapy is more effective when administered at an earlier stage pf a
developing cancer than at a later one. That early detection of cancer leads to
treatment at an earlier stage of disease development and therefore to a decline
in cancer mortality is a hypothesis, not a fact. Screening tests are developed
I
for the early detection and diagnosis of disease. The value and effectiveness
of screening programs that utilize the screening tests must be shown to either
reduce the mortality rate or increase the survival time for the disease targeted
by the screening. The benefits gained by early screen detection of a disease
can be assessed through the use of randomized screening trials. The actual
increase in survival time of participants offered screening over the survival time
among participants not offered screening is one useful measure of screening
effectiveness. Another measure is the reduction in mortality; but, the focus of
this thesis is the former.
1
1.1 Natural History of Cancer
The idealized disease process is defined by using a threestate progres
sive disease model whereby an individual in a screened population is assumed
to be in one of three states: the disease free state, S0, the preclinical state, Sp,
or the clinical state 5C[8]. Individuals in the disease free state, S0 are either
free of the disease or have disease characteristics which are undetectable by a
screening test. Participants in the preclinical state, Sp, do not have any ap
parent clinical symptoms, but have asymptomatic cancer that is detectable by
a screening test. These individuals are unaware of their illness. The transition
from the disease free into the preclinical state is assumed to take place at the
first point in time at which a disease is detectable by the screening test. In the
clinical stage, Sc, the disease is characterized by overt signs or symptoms lead
ing to diagnosis. This state follows the preclinical state and is marked by the
point of clinical diagnosis. The duration of the preclinical state is also called
the sojourn time. The transition from S0 > Sp > Sc is the basic structure of
the cancer screening models [4, p. 601].
Figure 1.1. Natural History of Cancer Progression
Cancer free state Preclinical state Clinical state
or Undetectable cancer Cancer screendetected Overt symptoms
Asymptomatic
2
1.2 The Screening Process
The screening process used for the detection of a disease involves
periodic screenings of participants performed at regular intervals of time. The
initial screen is designated as j = 0 and marks the entry into a study. At
the initial screening, all participants in a study are disease free. The time
interval between screenings is generally assumed to be a constant. (In this
thesis, the constant is denoted by 5.) The term screendetected, refers to those
individuals whose disease is detected in the preclinical stage by the screening
process. Preclinical individuals collectively have a distribution of sojourn times.
The screening process therefore samples from that distribution of the nathral
disease process; that is, the screendetected individuals form a sampling of the
preclinical durations. This sampling gives rise to several biases; two of which
I
are: lead time bias which is a consequence of the sampling and length bias
directly caused by sampling.
1.3 Lead Time Bias
Lead time is the time span by which a disease is diagnosed earlier
as a result of screening than it would have been in the absence of screening.
t is defined as the length of time by which the diagnosis is advanced over
clinical detection by virtue of the screening procedure. The participant has the
3
opportunity to begin earlier treatment during the lead time interval because
the diagnosis was made prior to the clinical phase of the disease. Survival
time is automatically lengthened for cases detected by screening, even if there
is no increased therapeutic benefit, when survival is measured from time of
diagnosis. This phenomenon is known as a lead time bias.
figure 1.2. Lead Time Bias
Study Arm
Screened
Cancer Detected
in Preclinical Phase
Lead time
Survival time since diagnosis
Control Arm
Unscreened
Cancer Detected in
Clinical Phase
Survival time since diagnosis

1.4 Length Bias
Length bias or lengthbiased sampling is another important, but sub
tle screening bias. This bias is a major factor in determining which preclinical
(disease cases will be detected early; that is, which cases will become part of
the sample representing the distribution of preclinical durations. The cases of
4
cancer for all participants of a study in the preclinical phase of the natural dis
ease process are not all equally likely to be detected by periodic screening. The
sojourn time probability density function (pdf) for those screendetected indi
viduals is different from the sojourn time pdf for the population. The screening
process does not detect people at random, but rather favors those with longer
sojourn times; that is, the probability of being in the sample is a function of
the sojourn time itself, an hence the sampled sojourn times do not represeht a
random sample from the individuals in the study population [2, p. 604].
The length of the preclinical duration, or sojourn time, varies from
person to person due to the different growth rates of cancer. It is commonly
believed that cases with long preclinical phases indicate a slowly advancing
cancer, while cases with short preclinical phases indicate a rapidly spreading
disease. With periodic screening, those cases having longer preclinical dura
tions are more likely to be detected; that is, they have a higher probability
of being detected than do cases of shorter preclinical duratons. As a result,
the longer duration preclinical cases are overrepresented among the screen
detected cases. It is not unreasonable to assume that the clinical course of the
disease is positively correlated with the preclinical course. Therefore the di
agnostic screen will automatically select those individuals who are more lijrely
(;o have longer survival times regardless of whether or not screening offers a
5
survival benefit due to early detection and treatment [2, p. 604]. If cancer ad
vances slowly in the preclinical disease and then progresses to a slowgrowing
clinical disease, the cancer will tend to have characteristics of a good progno
sis. Such cases would have more favorable outcomes even in the absence of
screening; thus, studies may show increased survival rates or a decrease in the
jnortality rate because of this bias. Periodic sampling due to screening will af
fect the survival distribution by making the survival times appear longer than
they would be in the general population, even in the absence of a screening
benefit. In Figure 1.3 the horizontal lines represent sojourn times. Detection
Of cancer in the preclinical phase corresponds to a horizontal line intersecting
a vertical screening line.
figure 1.3. LengthBiased Sampling or Length Bias
Initial Screen Screening Screening Screening
j = 0 j = l j = 2 j = 3
Periodic screening occurs at regularly scheduled intervals. In cancer studies, the
shorter sojourn time, those with poorer prognosis, may be underrepresented.
6
1.5 The Accuracy of Screening Tests
Cancer screening is the testing of apparently healthy volunteers from
the general population for the purpose of separating them into groups with
high and low probabilities of having a given disorder [8, p. 225].
A screening program involves the designation and recruitment of par
ticipants, the performance of the screening test at certain ages or frequencies,
and the provision for followup of suspicious and positive screening results, es
pecially for the diagnosis and treatment of those testing positive for the disease.
A good screening test should possess properties of high sensitivity and
high specificity. Sensitivity is the proportion of individuals designated positive
by the screening test among all individuals who have the disease. Specificity is
the proportion designated negative by the test among all those who do not have
the disease. ( Low specificity leads to a large number of false positive cases and
unnecessary treatment.) A screening test should have not only high specificity
to avoid needless treatment of negative cases, but also high sensitivity to deliver
the appropriate care to positive cases. The predictive value of a positive test
is the proportion of individuals with a positive test who have the disease. The
probability of correctly identifying those who have the disease should be high;
jn this thesis, test sensitivity will be denoted by ft.
Once an effective screening test is developed it becomes part of a
7
screening program involving therapy and followup which then must be evalu
ated in terms of disease outcome.
8
2. The Assessment of Screening Programs
Breast cancer is the most frequent type of cancer diagnosed ampng
women in the United States. Despite changes between 1940 and 1980 in eco
nomic and social conditions, nutritional status, health care, and fertility pat
terns, the mortality rate due to breast cancer had remained relatively constant.
Approximately one woman in ten develops clinically detectable breast cancer
In her lifetime [9, p. 8]. According to the National Center for Health Statistics,
the death rate due to breast cancer is 23 deaths per 100,000 females [9, p. 2].
jVlammography, clinical breast examinations, and breast selfexaminations are
three tests currently used in screening programs to detect breast cancer.
An objective method for evaluating the effectiveness of a screening
program is a randomized control trial or RCT. There have been several long
term studies measuring the reduction in breast cancer mortality. Two of the
most prominent studies, one conducted by Health Insurance Plan of New York
(HIP) during the 1960s and another conducted in the 1990s by the National
Cancer Institute of Canada, have led to some very interesting findings.
9
The HIP study compared the use of mammography plus clinical breast
examination versus usual medical care and found a 30% reduction in breast
cancer mortality over a tenyear period [9, p. 2]. The Canadian study compared
mammography, clinical breast examination, and breast selfexamination versus
clinical and breast selfexamination, but found no significant difference in the
mortality rate of the two arms when mammography was added to the screening
program [5].
2.1 The HIP Breast Cancer
Random Control Study
The HIP project was initiated by the National Cancer Institute in
the United States to determine whether periodic breast cancer screening With
mammography and clinical examination of the breast would lead to lower mor
tality. It was the first longterm randomized control trial of its kind with a
followup ending eighteen years from the date of entry into the trial.
The participants for this project were randomly selected from 80,300
Women, ages 40 64, covered by the HIP group insurance plan. The selected
participants entered the project between December 1963 and June 1966. Ev
ery woman from the insurance plan who was eligible to be in the study arm
received an initial mammogram and clinical breast examination followed by
10
annual reexaminations for three years. Of the 30,131 women comprising this
arm, approximately one third (10,800 women) refused to participate; their re
sults were nonetheless maintained as part of the study arm (intentiontotreat
analysis).
The counterpart to this arm, the control arm, included 30,565 women
who received an initial clinical examination of the breasts and then followed
their own usual practices in obtaining medical care.
Sample sizes for the trial needed to be large enough to detect a 20%
or greater reduction in breast cancer mortality at an alpha level of 0.05 [9].
The two arms were highly comparable; however, within the study arm, the
characteristics of the refusers differed from the women in compliance. Be
cause the arms were comparable, shifting this group of refusers to the control
arm would have biased the trial. Although they did not participate, data were
still maintained on them through insurance and other records.
Of the 30,131 women in the mammography and clinical examination
group, 20,200 (66.8%) appeared for the initial screening [9, p. 19]. Half of all
study arm participants had three or four screenings; see Table 2.1. Even in the
face of such low compliance, the results of the HIP trial nonetheless showed
that screening was beneficial.
11
Table 2.1. HIP Compliance of the Study Arm by Number of Screening Exams
[9, p. 19]
Number of Exams Number of Participants Total (%) Study Arm (%) Study Arm Participants
One or more 20,128 66.8 100.0
Two or more 17,476 58.0 86.8
Three or more 15,096 50.1 75.1
Four or more 11,932 39.6 59.3
%.2 Detection of Breast Cancers of the HIP Study
The followup of all those detected with breast cancer from both arms
of the trial lasted eighteen years. By the fifth, sixth, and tenth year from entry
into the trial, breast cancer cases in the study and control arms had equalized
[9, p.60] (Table 2.2). It is at these points of equalization that the mortality
and survival of the two arms can be compared.
2.3 Mortality in the HIP Trial
When the year of diagnosis was between one and five years after entry,
the number of breast cancer deaths within the first five years in the study arm
and control arms had accumulated to 39 and 63, respectively, implying a 38%
Reduction in mortality due to breast cancer. (When designing the sample size
for their study, the Canadian researchers [5] used this approximately 40% figure
as a target for achieving a desired power level of 0.80.) Within ten years from
entry, the reduction in mortality was 28.6% or approximately 30%, a frequently
12
Table 2.2. HIP: Cumulative Detection of Breast Cancers by Year and Arm
[9, p. 50]
Year from Entry Study Arm Cases Control Arm Cases Cumulative Study Arm Cases Cumulative Control Arm Cases
1 79 58 79 58
2 59 66 138 124
3 49 41 187 165
4 62 54 249 219
5 55 76 304 295
6 63 69 367 364
7 59 75 426 439
8 71 51 497 490
9 61 75 558 565
10 59 52 617 617
11 80 63 697 680
12 70 60 767 740
13 59 59 826 799
14 63 75 889 874
15 57 53 946 927
13
quoted percentage. By year 18, the reduction in mortality had dropped to 22.6
% [9, p. 65], (Table 2.3).
Table 2.3. HIP: Cumulative Cases of Mortality Due to Breast Cancer
Cumulative Deaths by Trial Arms
Mortality Ratios, Percent Reduction in Mortality
with 95% Confidence Intervals [9, p. 63]
Years from Entry Cumulative Study Arm Deaths Cumulative Control Arm Deaths Cum. Study Cum. Control Reduction in Mortality 95% Confidence Intervals
1 5 39 63 0.619 38.1 (8.9 59.5)
1 10 95 133 0.714 28.6 (7.4 45.5)
1 18 126 163 0.773 22.6 (2.739.0)
Most studies usually focus on mortality rates among patients with
breast cancer rather than survival rates because of the biases due to lead time
and length bias. The HIP project provided the first opportunity to develop and
apply models for estimating lead time that utilized a randomized control group.
The estimates were subject to large sampling errors and differed depending on
the model used [9, p. 36]. The lead time estimates for this trial varied by age
group. The estimated lead time for women between 40 and 49 was 5.2 months;
j:or 50 59, 21.9 months and for women 60 64, there was no clear evidence
of any lead time [9, p.36]. Survival rates calculated in the HIP project used an
adjustment of one year for lead time [9, p.71]. Table 2.3 shows that survival
among the study arm cases exceeded that in the control arm with or without
14
the lead time adjustment. No estimate of lengthbiased sampling was available
to adjust the survival rate. This length bias is the focus of this thesis.
Table 2.4. HIP: Survival Rates [9, p. 71]
Survival Rate Among Confirmed Breast Cancer Cases
Diagnosed During the First Five Years after Entry By Arms
Trial Number of Five Years Ten Years
Arm Cases from Entry from Entry
Study Arm 304 74 54.9
Control Arm 295 59.7 46.4
Study Arm Adj* 304 71.6 53.8
*The lead time adjustment was used to calculate the survival rate.
Â£.4 Conclusion of the HIP Trial
In conclusion, by ten years from entry into the trial, there were about
30% fewer breast cancer deaths in the study arm than in the control arm;
however, over the long term of eighteen years the reduction in the mortality
rate was approximately 23 %. The HIP project provided strong evidence that
periodic screening was beneficial, but the reduction in mortality as a result of
mammography alone was not demonstrated.
15
2.5 The CNBSS2 Breast Cancer Random Control Study
The Canadian National Breast Screening Study2 (CNBSS2) was de
signed to compare the incremental effect on the mortality rate of adding mam
mography to the screening program. So the study arm was offered annual
screening using mammography, physical breast examination plus breast self
examination; the control arm was offered annual physical breast examination
and breast selfexamination only. The sample size of 40,000 participants was
determined on the basis of an estimated power of 0.80 to detect a 40% reduc
tion in breast cancer mortality, similar to the reduction observed in the HIP
trial at five years from entry [5, p. 1491].
Participants for the study were recruited by general publicity, by per
sonal invitation, group mailings, and through physicians. From January 1980
through March 1985, 39,405 women, 50 59 years of age, were randomly and
individually assigned to one of the trials arms. The study arm, denoted by
MPlus, was comprised of 19,711 women who received annual mammograms
and clinical breast examinations. For the control arm (denoted BE), 19,694
women received annual clinical breast examination without mammography,
feoth groups were taught breast selfexamination and were encouraged to use
this technique regularly.
Five annual examinations were offered to the first 62% of the women
16
entering the trial and four examinations were offered to the remainder. Com
pliance in the study arm (Mplus) varied from 100% at the first screening to
86.7% by the fifth; 1.8% to 3.2% refused mammography at various screenings
and 6.1% or 1196 participants had interval mammograms.
The compliance of the control arm, BE, after the first screen varied
between 89.1% and 85.4% by the fifth screening; 16.9% (3300) of the partici
pants of this group had one or more interval mammographs; 8% of the 3300 liad
mammograms between screens four and five. These statistics reflect a much
better compliance than was observed in the HIP study.
2.6 Detection of Breast Cancers by CNBSS
The last screens were conducted in 1988 (all clinics closed), but the
annual followups for all women known to have breast cancer continued until
June 30,1996. The mean followup from entry into the trial was 13 years. Cases
of breast cancer detection were classified as either insitu or invasive cancers.
In the MPlus arm, there were 71 insitu cases in comparison to 16 in the BE
arm. Invasive cancers were then classified into one of three categories: screen
detected, interval (those cases occurring within twelve months of a negative
screening), and incident cancers (those cases occurring twelve or more months
after the previous CNBSS screening examination).
17
Table 2.5. CNBSS: Detection of Invasive Cancers [5, p. 1492]
MPlus BE
Screen Interval Incident Screen Interval Incident
Detected Detected
Year 1 118 114 0 64 16 0
Years 25 149 36 32 84 72 47
Years 69 0 175 217
Total 267 50 207 148 88 264
The invasive cancers totaled 524 and 500 for the MPlus and BE participants
respectively (see Table 2.5). By December 1993, the totals rose to 622 and
610 in the respective groups. The number of invasive cancers approximately
equalized by the end of the thirteenth year of the study [5, p. 1493]. The de
tection rates were higher in the study arm than in the control arm throughout
screening, resulting in lead time for the MPlus participants during which they
received earlier diagnosis and treatment. Despite this earlier treatment, mor
tality rates between the two arms were almost identical (see section 2.2.2). The
average lead time for MPlus participants has been estimated to be 3.6 years
(95.% Cl: 2.7 5.5) in contrast to only 1.5 years for the BE. (Cl: 2.7 55).
The lead time gained by the study arm over the control arm was 2.1 years on
average [5, p. 1492]. As a result of mammography, more diagnostic procedures
were recommended and performed on the MPlus arm than on the participants
of the BE group. Interestingly, the cancers detected by mammography alone
18
were less likely to be lymph node positive than those detected by physical ex
amination. Small tumors were less likely to be lymph node positive than large
Rumors. The MPlus arm had a higher biopsy rate for benign lesions anc) an
excess of mastectomies due to uncertainty over the appropriate treatment [5,
p. 1496].
2.7 Mortality in the CNBSS2 Trial
The total number of deaths from various causes were similar in the
study and control groups (734 and 690 respectively) illustrating the compari
blity of the two arms (see Table 2.6) [5, p. 1494],
Table 2.6. CNBSS: Mortality Cases by Cause to the End of 1993 by Study
Arm [5, p. 1494]
Causes MPlus BE
Breast cancer 88 90
Other cancers 376 313
Noncancer 270 287
Total 734 690
Among those diagnosed with breast cancer during the first five years
after entry into the study, the mortality ratio, M~BUS, is 1.09. Similar ratios
resulted as deaths accumulated [5, p. 1495] ( see Table 2.7).
19
Table 2.7. CCBSS: Cumulative Deaths due to Breast Cancer
Number of deaths from breast cancer to June 30, 1996 by study arm and year
of detection
Year MPlus Deaths BE only Death Rate MPlus/BE 95% Confidence Interval
Year 5 74 68 1.09 (0.78 1.51)
Year 6 84 76 1.10 (0.81 1.51)
Year 7 93 83 1.12 (0.83 1.50)
Year 8 99 89 1.11 (0.48 1.48)
Year 9 104 97 1.07 (0.81 1.41)
Beyond Year 9 107 105 1.02 (0.78 1.33)
Â£.8 Conclusion of CNBSS2 Trial
The CNBSS2 is the only trial that has evaluated the effect of mam
mography over and above physical breast examination and breast selfexamination
in women aged 50 59 at entry into the trial. The Canadian trial concluded
that screening women in this age category with yearly mammography in ad
dition to physical examination detected more lymph node negative and small
breast cancers than screening with physical breast examination alone but had
no impact on mortality from breast cancer [5, 1496]. This is the first study to
show lead, time without benefit for breast cancer screening [5, 1497].
In the HIP trial, 87.5 % of women diagnosed with impalpable canbers
were still alive at ten years. This is in contrast to the Canadian trial showing
89.9% of the cancers detected in the BE arm experienced survival at ten years
[5, 1497]. According to the Canadian trial, the survival in the HIP trial is
20
almost certainly due to lead time and length bias....at least 70% of the benefit
time may have come from the physical examinations of the breasts rather
than the mammography [5, 1497].
21
3. The Exponential Distribution of Sojourn Times
The beginning of a preclinical phase can not be determined unless
screenings are conducted continuously rather than periodically. So the question
is how can the mean length of the preclinical phase or the mean sojourn time
be estimated given that the preclinical stage is not totally observable in the
disease process? The clinical phase can be observed because symptoms are
overt. (Many believe that the clinical and preclinical phases are positively
correlated, but the correlation is unknown.) The survival time is also observable
if the measurement is taken from the time of entry into the trial, given there
are no detectable cases in the healthy population originally. This survival
measurement would not be affected by the lead time bias because survival
would no longer be measured from time of diagnosis.
Consider the lengthbiased density of sojourn times when only one
screening is conducted. Let Y* be the random variable that denotes the pre
clinical duration of a screen detected case and Y be the random variable that
denotes the preclinical duration for a case in the general population.
The following notation will be used:
22
n = the number of observations.
Yi, Y2, ...Yn are the sojourn times of the n cases.
/y(') = the probability density function (pdf) of preclinical durations.
/y*(y) = the probability density function of the sampled preclinical
durations.
Py = mean of (Yi, Y2, ...Yn) i.e., p = /0 y}v{y)dy
ny = the number of preclinical durations, Yi, with lengths y
Then, the density of the sampled preclinical durations, /y. (y), can be
derived from the pdf of all preclinical durations as follows: [4, p. 4]
fy (V) =
= lim
lim (proportion of ^ Y due to intervals of length y)
i1
yny
nkx> yv Yi
M), ELi r<
lim
71>00
fy(y)
7
n
y
pY
Thus, the density of Y* is related to the density of Y [4, p. 4]. In this
thesis, I compare the mean of the sampled preclinical durations with the mean
of the population preclinical durations. If there were no effect from length
biased sampling, the mean from the sampled preclinical durations would equal
23
the population mean of preclinical durations. Thus, this ratio gives an estimate
of the increase due to lengthbiased sampling and can be expressed in terms of
the coefficient of variation as follows:
E(Y*) a
J = (1 + CV3) where the coefficient of variation CVy = .
E(Y) //
Now, consider screening occuring at regular or periodic intervals. The
following notation is used to derive the cumulative distribution for preclinical
durations at the first screening.
n = the number of observed preclinical durations.
5 denotes the interval of time between screenings.
X represents the beginning of a preclinical phase. Assume that X is
uniformly distributed over the interval [0,5] because the disease ex
hibits no particular preference for any time between the initial and
subsequential screenings. [4],
Y denotes the continuous random variable of the preclinical duration
of a case in the general population.
2/i> 2/2j Vn denote the observed preclinical durations of a population.
fy(y) is the probability density function of all preclinical durations.
Yj* = is the continuous random variable of a screendetected preclinical
duration observed at the jth screen, j = 1,2,3,4.
Fyj (y) is the cumulative distribution function (cdf) for the preclinical
24
durations of cases detected at the jth screening, j = 1,2,....
Cases that should be detected by the first screening will be those that
meet the conditions: 0 < X < 5 and X + Y > S. (If X + Y <5, the case is an
interval case which has progressed to a clinical case at the time of screening.)
Thus, the cdf of the sojourn times for these cases is derived as follows using a
conditional probability that includes the conditions above.
Fyi (y)
P{Y < y\X + Y > 8 and 0 < X < 8
lmax(0,sy) p[Y ^ V and X + Y > 8 and 0 < X < 5 and X = x}fx(x)dx
fg P{X + Y > 6 and 0 < x < 6 and X = x}fx(x)dx
SL*Ma)P{Sx
SSPlY>Sx}$dx
yFyjy) F(u)du
$ ~~ Jo F(u)du
SFY(y) J^F(u)du
5f*F(u)d
where y <
where y >
u
8
5.
Thus, the pdf of Y* and the mean, E(Y*), for the screenable sojourn
times at the first screening can be calculated as follows: [4]
x f ^ min(y, 5) fy(y)
frM ~ TSjF
E(Yi) = Svfy(y)dy
1 r5 roo
E(Yl) = , { y2fY{y)dy + 5 yfY{y)dy\
S ~ fo F(y)dy Jo Js
While the preclinical durations are not directly observable, studies
25
have suggested that the preclinical durations from the HIP study have a prob
ability distribution that is approximately exponentially distributed [2]. This
distribution is used frequently as a model for the distribution of times between
jhe occurrence of successive events.
The exponential pdf, is the special case of the general gamma pdf with
r = 1. Let the preclinical durations of a population, Y, have an exponential
distribution.
where A > 0.
Xe Xy for y > 0
<
0 otherwise
For some fixed value y, the probability that the observed value of Y
Will be at most y can be determined by employing the cumulative distribution
junction.
F(y,\) = P{Y
It follows that
F(y,X) = {
1 e~Xy
0
for y > 0
y < 0
Xe Xudu for y > 0.
Using the exponential distribution for Y (unsampled) in the case of
periodic screenings, the probability density function of the sojourn times de
tected at the jth screening, /yy(y; A), is derived in two intervals as follows:
26
[y U l)5]Ae
N/
fyivA)
s ~ [/o'5!1 e~Xy)dy /o 1)5(1 e~Xy)dy\
for (j 1)6
6Xe Xy
5 ~ [/o'5!1 e~xv)dy $ 1)5(1 e~Xy)dy\
for y > j6
\
E(Yj) represents the mean preclinical duration among cases that
started before 5 and extended to at least j5, the time of the jth screen. It
js given by:
E(Y>)=
[y (j ~ l)5]Xe
\y
s [/
8\e~Xy
my
+
Xy)dy]
)dy
5 [/o'5!1 e~Xy)dy $ 1)5(1 e Xy)dy]
)dy.
using the notation Djs for the denominator, we have,
5\ r
Djs = 6 [/0j5(l e~Xy)dy /o01^!1 e~Xy)dy\ =
Now, E(Yj) can be simplified as:
ris
E(Yj) = [\(y
j a
X.
(v
[y (j l)S]Xe Xy
+
01)5
00 SXe~Xy
D
js
)dy
)dy.
j& Djg
2 j5 A + esx (2 8X + j5X)
X(e5x 1)
27
Letting the parameter A = ^, it follows that lim^o E(Yj) = yu. That is, if
screening occurred continuously instead of periodically (5 > 0), then in the
limit E(Y*) > E(Y) = i = p and Var(Yj) = A = /r2.
Using Mathematica and different values for fi, the values of E(Yj)
were calculated: (see Table 3.1).
Suppose the time interval between screenings is 5 = 2 years. From Table 3.1,
the mean length of the preclinical phase through the third screening, (j 3)
given jj, 0.5, is 4.96 years. Furthermore, with that table, E(Y*), the mean
preclinical duration of all cases that are eligible for detection by screening given
that four screenings have occurred can be calculated as follows:
E(Y*) = '^'jLoE^Y^P{missed on (j 1) previous screens,
detected on the ^screen}
= pew) + (i mm) + a  + a 0?mY()
where /3 is the test sensitivity [4, p. 7].
The results of applying this equation to the matrices E(YJ*) for j =
1,2, 3,4 and setting /? equal to 0.95 follow in Table 3.2.
When the true mean preclinical duration is 1 year long, the sampled duration
is 1.47 years long, almost a half of a year longer. Thus, even in the absence
of any screening benefit, the survival time since diagnosis would be almost one
half of a year longer. One might erroneously conclude an increase benefit from
28
Table 3.1. Exponential Distribution: Means for j = 1, 2,3,4
E(YJ*): The mean length of the sampled sojourn times at the jth screening.
Values of 5
E(Y?) V 1 2 3 4 5
E(Y{) 0.5 0.84 0.96 0.99 1.00 1.00
1.0 1.42 1.69 1.84 1.93 1.97
2.0 2.46 2.84 3.14 3.37 3.55
5.0 5.48 5.93 6.35 6.73 7.09
10.0 10.49 10.97 11.42 11.87 12.29
e(y2) 0.5 1.84 2.96 3.99 5.00 6.00
1.0 2.42 3.69 4.84 5.93 6.97
2.0 3.46 4.84 6.14 7.37 8.55
5.0 6.48 7.93 9.35 10.74 12.09
10.0 11.49 12.97 14.43 15.87 17.29
E(Yi) 0.5 2.84 4.96 6.99 9.00 11.00
1.0 3.42 5.69 7.84 9.93 11.97
2.0 4.46 6.84 9.14 11.37 13.55
5.0 7.48 9.93 12.35 14.74 17.09
10.0 12.49 14.97 17.43 19.87 22.29
E(Yt) 0.5 3.84 6.96 9.99 13.00 16.00
1.0 4.42 7.69 10.84 13.93 16.97
2.0 5.46 8.84 12.14 15.37 18.55
5.0 8.48 11.93 15.35 18.74 22.09
10.0 13.49 16.97 20.43 23.87 27.29
29
Table 3.2. E(Y*): The Mean Preclinical Duration of Cases Eligible for De
tection by Screening
Values of 5
A4 1 2 3 4 5
0.5 0.90 1.07 1.15 1.210 1.26
1.0 1.47 1.79 2.00 2.14 2.23
2.0 2.51 2.94 3.30 3.58 3.82
5.0 5.54 6.04 6.51 6.95 7.35
10.0 10.54 11.07 11.58 12.08 12.56
screening. The table of ratios of E(Y)*) divided by /j, for j = 1,2,3,4 is given
in the appendix (see table A.11). Using the four matrices within that table,
the ratios:
i
are given in Table 3.3.
Using the ratios in Table 3.3, the relative increase in the estimation of the
sampled mean can be calculated by subtracting one from each value and mul
tiplying by 100% to obtain percentages. The results are shown in Table 3.4.
30
Table 3.3. Exponential Distribution Ratio Comparison: E(Y*)/E(Y)
Values of 5
A* 1 2 3 4 5
0.5 1.79 2.14 2.30 2.42 2.53
1.0 1.47 1.79 2.00 2.14 2.23
2.0 1.26 1.47 1.65 1.79 1.91
5.0 1.11 1.21 1.30 1.39 1.47
10.0 1.05 1.11 1.16 1.21 1.26
Table 3.4. Relative Increase in the Mean: (E(Y*)/fj, 1)100%
Values of 8
1 2 3 4 5
0.5 79 113 130 142 153
1.0 47 79 100 114 123
2.0 26 47 65 79 91
5.0 11 21 30 39 47
10.0 5 11 16 21 26
Prom Table 3.4, when screening occurs every two years and the true
Mean sojourn time of the population is five years, the average sojourn time
among the screendetected cases is 21% larger than the overall mean. As the
screening interval increases, the estimations of the preclinical durations get even
larger. (Intuitively, diseases with shorter durations would need more frequent
31
screenings.)
The formula for the variance or the mean squared distance of the
sojourn times from the mean is
4 = Var(Y)
POO
= (y VY)2fY{y)dy for y > 0.
Jo
POO
= (y E(Y))2fY(y)dy for y > 0.
Jo
The variance for the exponential distribution at the jth screening becomes:
Var(Y*)
j y2fYj{y)dy (J yfyj(y)dy)2
E(Y2*) [E{Y*)}2
pjd roo
/ (v2(y ~ (j ~ l)S)\e^)/Djs + (y2S\e^)/DjS
J{j1)5 JjS
pjd poo
( / , (y{y (3 l)S)\e~*)IDjs + (yd Ae~*)/DiS )2
J(j1)5 Jj5
2 esx{4 2e5X + 62A2)
A2(eaA l)2
rjS r(jl)6
where Djs = 5 [I (1 e~Xy)dy I (1 e~ y)dy\ 
e5x 1
Ae5jA
2e5A(42 e5X + S2\2)
A2(e5A l)2
S2e5/ + h2(2 4ea^ + 2e26
(e5^ l)2
where A = 1/y,.
32
Note that the variance of sojourn times is independent of j so Yj* =
y; = YÂ£ = y4* and is therefore equal to Y*. In this thesis, the calculation for
Y* is made anyway using equation 3.3 verifying this fact (Table 3.5). Var(Y*)
is calculated in the same manner as E(Y*), again letting j3 be the test sensi
tivity, we have [4, p. 7]
Var(Y*) = '^2 Var(Yj)P{ missed on (j 1) previous screens, (3.1)
3=0
detected on jth screen} (3.2)
= 0Var(Y{) + (1 f3)!3Var(Y2*) + (1 /3)2/3Var(Y3*) (3.3)
+ (l/3fl3Var(Y:)
?
The variances of the sampled preclinical durations, Y*, are shown in Table 35;
they are indeed the same as the variances for each Y* for the various values
of ft. From the variances, the standard deviations are calculated by taking the
square root. (SD(Y*) = [Var(y*)]s) (Table 3.6)
Table 3.5. Var(Y*)\ Variances of Sojourn Times of Cases Eligible for Detec
tion by Screening
Values of 5
n 1 2 3 4 5
0.5 0.32 0.42 0.48 0.49 0.50
1.0 1.08 1.28 1.50 1.70 1.83
2.0 4.08 4.32 4.67 5.10 5.56
5.0 25.08 25.33 25.74 26.29 26.98
10.0 100.08 100.33 100.75 101.32 102.06
Table 3.6. SD(Y*): Standard Deviations of Sojourn Times of Cases Eligible
for Detection by Screening
Values of 5
1 2 3 4 5
0.5 0.56 0.65 0.69 0.70 0.71
1.0 1.04 1.13 1.23 1.30 1.35
2.0 2.02 2.08 2.16 2.26 2.36
5.0 5.01 5.03 5.07 5.13 5.19
10.0 10.00 10.01 10.04 10.07 10.10
34
Table 3.7. Exponential Distribution Ratio (SD (Y*))/ji
Values of 5
1 2 3 4 5
0.5 1.13 1.30 1.38 1.41 1.41
1.0 1.04 1.13 1.23 1.30 1.35
2.0 1.01 1.04 1.08 1.13 1.18
5.0 1.00 1.01 1.01 1.03 1.04
10.0 1.00 1.00 1.00 1.01 1.01
Subtracting one from the values in Table 3.7 and multiplying by 100 will pro
duce the relative error, in percentages, between the standard deviation of the
sampling distribution and the population distribution of sojourn times. These
values are shown in Table 3.8. Note that the relative error in the standard de
viations is smaller than that in the means but nonetheless shows an increase.
As with the means (Table 3.4), the relative error increases as <5 increases for a
given value of ji and decreases as /i increases for a given value of 5. Thai is,
the relative error, in both the mean and standard deviation increases as the
screening interval increases and as the mean preclinical durations decrease.
35
Table 3.8. Relative Error in the Standard Deviations
Values of 5
1 2 3 4 5
0.5 13 30 38 41 41
1.0 4 13 23 30 35
2.0 1 4 8 13 18
5.0 0 1 1 3 4
10.0 0 0 0 1 1
Table 3.9 is a summary of the findings for the exponential distribution. As
the interval between screenings gets larger, the estimation of the sample means
also gets larger. The relative increase also gets larger. For example: When
5 = 1 and [x = 0.5, the mean of Y* is estimated to be 0.90 or 79% larger than
jhe overall mean of Y. The relative error in standard deviations gets larger for
each value of fx as 5 increases but decreases as jx gets larger.
36
Table 3.9. Exponential Distribution Summary of Y* when r = 1
The Means, Relative Increase, Variances, Standard Deviations, and Relative
Errors
The Preclinical Durations of Cases Eligible for Detection by Screen
ing
Values of 5
E(Y*) V 1 2 3 4 5
0.5 0.90 1.07 1.15 1.21 1.26
Sampled Means 1.0 1.47 1.79 2.00 2.14 2.23
2.0 2.51 2.94 3.30 3.58 3.82
5.0 5.54 6.04 6.51 6.95 7.35
10.0 10.54 11.07 11.58 12.08 12.56
Relative Increase V 1 2 3 4 Ej
0.5 79 113 130 142 153
o o tH 1 1 h =* 1.0 47 79 100 114 123
2.0 26 47 65 79 91
5.0 11 21 30 39 47
10.0 5 11 16 21 26
Var{Y*) 1 2 3 4 5
0.5 0.32 0.42 0.48 0.49 0.50
Variances of 1.0 1.08 1.28 1.50 1.70 1.83
sampled sojourn 2.0 4.08 4.32 4.67 5.10 5.56
times 5.0 25.08 25.33 25.74 26.29 26.98
E(Y*2) (E(Y*))2 10.0 100.08 100.33 100.75 101.32 102.06,
sd (V*) 1 2 3 4 5
0.5 0.56 0.65 0.69 0.70 0.71
Standard Deviations 1.0 1.04 1.13 1.23 1.30 1.35
of sampled sojourn 2.0 2.02 2.08 2.16 2.26 2.36
times 5.0 5.01 5.03 5.07 5.13 5.19
10.0 10.00 10.01 10.04 10.07 10.10
Relative Error 1 2 3 4 3
0.5 13 30 38 41 41
(sd{Y*)/iM 1)100 1.0 4 13 23 30 35
2.0 1 4 8 13 18
5.0 0 1 1 3 4
10.0 0 0 0 1 1
37
4. The Gamma Distribution of Sojourn Times
The gamma distribution is used to model waiting times or lifetipies
such as human longevity after age twenty or so [7, p. 189]. It is appropriate to
use this distribution as a model for preclinical durations. The exponential pdf
is a special case of the general gamma probability density function for r = 1;
therefore, this chapter will consider the cases r = 2 and r = 4 Once again,
iet the preclinical durations of a population be represented by the continuous
random variable Y. This random variable is said to have a gamma distribution
if the pdf of Y is
f(y,r, A) =
' Aryr~1eXy
r(r)
0 otherwise
for y < 0
where r > 0 and A > 0. The gamma function is defined by T(r) =
/oyT~le~vdy but for any positive integer r, F(r) = (r 1)!; thus, for r =
2,T(2) = 1 and T(4) = 6. The frequencies of preclinical durations that are y
units of time long are given by the gamma distribution.
The probability that the observed value of Y will be at most y for some
fixed value y, can be determined by use of the cumulative distribution function
38
(cdf) which is obtained by integrating the gamma distribution as shown:
F(,V, r, A) = P{Y
Jo Jo i (r)
for y > 0.
F{v,r, A)
' i 1 + XV
e\y
for y > 0 and r = 2
0 y < 0
When the distribution of unsampled preclinical durations (V) is gamma dis
tributed and screenings are periodic, the pdf of the sojourn times detected at
j;he jth screening, fyj(y,rX), is calculated over two intervals,
p' 1)5
________________[y ~ {j ~ l)S]Xryr~1e~Xy/T(r)____________
5 [/o<5(1 + eXy A y)/exydy /o5_1^(1 + e.Xy A y)/exydy\
fyj (y; r, A)
for (j 1)6
5Xr yr~le~Xy/r(r)
s ~ + eXy) ~ Ay)/dy  1 + eXy ~ \y)/exvdy]
for y > jÂ§.
The denominator of the above density function is denoted by Dj$
e5X(2 5X + 5jX) 5Xj 2
where Djs =
X e5Jx
Using this notation the density
function becomes
[y (j l)5]A'y1e*7r(r)
(y; r, A) = <
DjS
5XTyr~1e~Xy/T{r)
Di,
JS
for (j 1)6
for y > jS
39
4.1 The Gamma Distribution for Sojourn Times when r = 2
The main purpose of this section is to generate the mean sojourn
time of screendetected cases and compare them to the overall sojourn times,
as well as calculate the percentage increase between the two. In this section,
all calculations use the gamma distribution with r = 2. In the next section,
father than repeat the exact same steps, only the summary results are given
for r = 4. Tables from which the summary is derived for r = 4 are located in
the appendix.
The first goal is to calculate the sampled mean of sojourn tiifies,
E(Y*). To do so, each E(Yj) for j = 1,2,3, and 4 are needed, where E(Yj)
represents the mean preclinical duration among cases that started before 5 but
no later than j5. The expected value or mean of the distribution of sojdurn
times, fyfYj{y)dy, by the jth screening, is given by
which becomes for any j > 0,
6 + 5A(4 + 4j + 8X 2SXj + j252X2)
X(25X + 5jX)
25 + 52X{2 4j + 5jX 5j2X)
{28X + 8jA)(2 5jX + e5X{2 5X + 5jA))
40
Screenings performed continuously are found by letting 5 > 0. We find that
lim^o E(Yj) = (j, when A = jj. The table of means (Table 4.1) was generated
using Mathematica with different values for n and setting A = ^ in the equation
for E(Yj) above.
Repeating the process that was used for the exponential distribution,
the expected preclinical durations for the cases eligible for detection by screen
ing was calculated using Table 4.1 to produce Table 4.3.
E(Y*) = E(Y?yP{missed on (j 1) previous screens,
detected on the jthscreen}
= PEfXi) + (1 mE(Yl) + (1 mEK) + (1 PYPE(Yi)
where is the test sensitivity [4, p. 7] (See Table 4.3) .
From Table 4.3, the ratios of E(Y*)///, are calculated producing Table
4.5. These ratios being greater than one indicate that the means of the sampled
durations are larger than the overall means, thus illustrating that the sample
is lengthbiased.
The relative increase in the means for the gamma density is obtained
by subtracting one from each value above and multiplying by 100 to obtain
percentages. The percentage of increase in the sojourn times for r 2 over
;he actual expected preclinical durations are found in Table 4.6.
The variance of preclinical durations at the jth screening follows using the
41
Table 4.1. Gamma Distribution r = 2 Means for j = 1, 2,3,4
Values of 5
E(Y') 1 2 3 4 5
07) 0.5 0.70 0.75 0.75 0.75 0.75
1.0 1.22 1.40 1.47 1.49 1.50
2.0 2.18 2.44 2.66 2.81 2.90
5.0 5.10 5.32 5.58 5.85 6.11
10.0 10.06 10.20 10.40 10.64 10.90
E(Y{) 0.5 1.55 2.55 3.54 4.53 5.52
1.0 2.02 3.11 4.11 5.10 6.08
2.0 2.86 4.03 5.15 6.21 7.23
5.0 5.56 6.59 7.74 8.92 10.08
10.0 10.35 11.13 12.10 13.19 14.32
EW) 0.5 2.52 4.53 6.52 8.51 10.51
1.0 2.96 5.05 7.06 9.05 11.05
2.0 3.73 5.91 8.03 10.10 12.12
5.0 6.24 8.26 10.42 12.61 14.78
10.0 10.84 12.48 14.43 16.52 18.67
E(Y.,*) 0.5 3.51 6.52 9.51 12.51 15.51
1.0 3.93 7.02 10.04 13.04 16.03
2.0 4.66 7.85 10.98 14.05 17.07
5.0 7.03 10.07 13.25 16.45 19.63
10.0 11.44 14.06 17.03 20.14 23.31
Table 4.2. Gamma r = 2,E(Y*): The Mean Preclinical Duration of Cases
feligible for Detection by Screening with (5 0.95
Values of 5
1 2 3 4 5
0.5 0.75 0.84 0.90 0.95 1.00
1.0 1.26 1.49 1.61 1.68 1.74
2.0 2.22 2.53 2.79 2.99 3.13
5.0 5.13 5.39 5.70 6.01 6.32
10.0 10.07 10.25 10.49 10.78 11.08
42
Table 4.3. Gamma r 2,E(Y*): The Mean Preclinical Duration of Cases
Eligible for Detection by Screening with /? = 0.95
Values of 8
t* 1 2 3 4 5
0.5 0.75 0.84 0.90 0.95 1.00
1.0 1.26 1.49 1.61 1.68 1.74
2.0 2.22 2.53 2.79 2.99 3.13
5.0 5.13 5.39 5.70 6.01 6.32
10.0 10.07 10.25 10.49 10.78 11.08
Table 4.4. Values of E(Yj)/fj, when fy{y) is a Gamma Distribution r 2.
Values o:
5
Screen P 1 2 3 4 5
0.5 1.40 1.49 1.50 1.50 1.50
3 = 1 1.0 1.22 1.40 1.47 1.49 1.50
2.0 1.09 1.22 1.33 1.40 1.45
5.0 1.02 1.06 1.17 1.17 1.22
10.0 1.01 1.02 1.04 1.06 1.09
0.5 3.11 5.10 7.07 9.06 11.05
3= 2 1.0 2.02 3.11 4.11 5.10 6.08
2.0 1.43 2.02 2.58 3.11 3.61
5.0 1.11 1.32 1.55 1.78 2.02
10.0 1.04 1.11 1.21 1.32 1.43
0.5 5.05 9.05 13.04 17.03 21.02
3 = 3 1.0 2.96 5.05 7.06 9.05 11.15
2.0 1.87 2.96 4.02 5.05 6.06
5.0 1.25 1.65 2.08 2.52 2.96
10.0 1.08 1.25 1.44 1.65 1.87
0.5 7.02 13.04 19.03 25.02 31.02
3 = 4 1.0 3.93 7.02 10.04 10.04 16.03
2.0 2.33 3.93 5.49 7.02 8.54
5.0 1.41 2.01 2.65 3.29 3.93
10.0 1.14 1.41 1.70 2.01 2.33
43
Table 4.5. Gamma Distribution Ratio when r = 2 E(Y*)/fi
Values of 5
A* 1 2 3 4 5
0.5 1.49 1.68 1.79 1.90 2.00
1.0 1.26 1.49 1.61 1.68 1.74
2.0 1.11 1.26 1.40 1.49 1.56
5.0 1.03 1.08 1.14 1.20 1.26
10.0 1.01 1.03 1.05 1.08 1.11
Table 4.6. Gamma Distribution Relative Increase in the Means
Values of 5
V 1 2 3 4 5
0.5 49 68 79 90 100
1.0 26 49 61 68 74
2.0 11 26 40 49 56
5.0 3 8 14 20 26
10.0 1 3 5 8 11
44
Table 4.7. Gamma Formulas for the Expected Value of Y2* when r = 2
E{Yf*) Equations
eiy;1') iff) iff) Â£(ff) 24 24edA + 185A + 65'2X2 + 63A3
A2 (2 2e5x + 5A) 24 + e5X(24 + 185A + 652A2 + 53A3) 45A(9 + 65X + 252X2)
A2(2 25X + esx(2 + 5A)) 24 27
X2[235X + 2e5X(l + 5X)] 8(3 95X 12 52A2 1853A3) + 3e"[8 + 95A(2 + 25X + 52A2)]
A2f2 45A + e5A(2 + 35A)j
gamma distribution:
VariYj) =
J y2fYj(y)dy (Jyfyj(y)dy)
= E(Y2)[E(Y)]2
t
 c
jS y2{y (j l)^)Aryr~1e~A/r(r)  f y26\ryr1exv)/T{r)
jS y(y ~ (j ~ l)5)YyrleXy/T(r)
(jl)5 Djs
+
/
Jj5
r
JjS
) jS Djs
r y5Xryr~io~
D
e^)/r(r))2
36
where Djs
edX(26X + 6j\) SXj2
Xe5ix
Unlike the exponential, the variances of sojourn times using this pdf for r 2,
are not independent of j. The formulas in Table 4.7 and Table 4.8 can be used
the calculate E(Y2*) [.E(YJ*)]2, the variances for Y*, F^*, V3* and V4*.
The lim<5_^o Var{Y?) = ^ when A = ^. To calculate the variances,
let A = in the formula tables 4.6 and 4.7.
This is the same value that was used when the mean sojourn times
45
Table 4.8. Gamma Formulas for the Expected Value Squared when r = 2
Equations
(E(Y)f (J5(n*))2 W))! 66eAX + 45X + 52X2 2 ^ A(2 2esx + 5X) 2[3 + 25X(2 + <5A)1 + e5X[6 + 5X(4 + 6A)1.2 ( A[2(l+)+e(2 + )] > 6 12dA 952A2 + 2e(3 + 4SX + 2<52A2) 2 ^ A[235A + 2e*A(l +
were calculated. The variances using this substitution are found in Table 4.6.
The same equation for E(Y*) can now be applied to Var(Y*)\ that is,[4, p. 7]
Var(Y*) = T,
detected on the jthscreen}
= pVar(Y*) + (1 /3)pVar(Y2*) + (1 /3)2/3Var(Ys*)
+ (l/3)3/3Var(Y4*)
where is the test sensitivity [4, p. 7].
To obtain a comparison, that is a ratio between the sampled standard devia
tions, Y*, and the overall standard deviations of Y, each value in Table 4.9 is
divided by to obtain Table 4.12.
The relative differences in the sampled standard deviations and the
standard deviations of preclinical duration are very small since each ratio is
close to one. This completes the analysis for the gamma distribution for r = 2.
46
Table 4.9. Gamma Distribution Variances for r 2, j = 1,2,3,4
Va:
ues of 5
Var(Y*) P 1 2 3 4 5
Var(Y*) 0.5 0.15 0.18 0.19 0.19 0.19
1.0 0.49 0.59 0.68 0.73 0.74
2.0 1.92 1.95 2.12 2.34 2.56
5.0 12.31 12.06 11.93 11.97 12.17
10.0 49.75 49.24 48.70 48.23 47.89
Var(Y2) 0.5 0.13 0.15 0.14 0.14 0.14
1.0 0.43 0.52 0.57 0.58 0.57
2.0 1.71 1.72 1.89 2.07 2.21
5.0 11.66 10.91 10.57 10.56 10.78
10.0 48.68 46.65 44.91 43.62 42.76
Var(Y3') 0.5 0.12 0.14 0.13 0.13 0.13
1.0 0.40 0.49 0.53 0.54 0.54
2.0 1.57 1.59 1.77 1.94 2.07
5.0 11.00 10.03 9.68 9.71 9.97
10.0 47.26 44.00 41.65 40.10 39.18
Var(YÂ£) 0.5 0.12 0.13 0.13 0.13 0.13
1.0 0.38 0.47 0.52 0.53 0.53
2.0 1.47 1.52 1.70 1.88 2.00
5.0 10.44 9.42 9.13 9.21 9.51
10.0 45.81 41.78 39.22 37.69 36.85
Table 4.10. Gamma Distribution r = 2, {3 = 0.95 Var(Y*)
Values of 5
1 2 3 4 5
0.5 0.15 0.18 0.18 0.19 0.18
1.0 0.48 0.58 0.68 0.72 0.74
2.0 1.91 1.94 2.10 2.33 2.54
5.0 12.28 12.00 11.86 11.90 12.10
10.0 49.70 49.11 48.50 48.00 47.63
47
Table 4.11. Gamma Distribution r 2, Standard Deviations of Y
Values of S
1 2 3 4 5
0.5 0.38 0.42 0.43 0.43 0.43
1.0 0.70 0.76 0.82 0.85 0.86
2.0 1.38 1.39 1.45 1.53 1.59
5.0 3.50 3.46 3.44 3.45 3.48
10.0 7.05 7.01 6.96 6.93 6.90
Table 4.12. Gamma Distribution r = 2 Ratios of Standard Deviations
(Standard Deviations of Y* divided by ^=)
_______________Values of 5 ____________,
1 2 3 4 5
0.5 1.08 1.20 1.22 1.22 1.22
1.0 0.98 1.08 1.16 1.20 1.21
2.0 0.98 0.98 1.03 1.08 1.13
5.0 0.99 0.98 0.97 0.98 0.98
10.0 1.00 0.99 0.98 0.98 0.98
48
Table 4.13. Gamma Distribution r = 2; Relative Error ( %) in Standard
Deviations of Y*
Values of 5
1 2 3 4 5
0.5 8 20 22 22 22
1.0 2 8 16 20 21
2.0 2 2 3 8 13
5.0 1 2 3 2 2
10.0 0 1 2 2 2
A summary of the findings is located in Table 4.14.
4.2 The Gamma Distribution for Sojourn Times for r = 4
The summary for the comparison of the screendetected sojourn tifries
with the overall sojourn times using gamma distribution for r 4 is given in
4.15. A comparison of the three summary tables for r = 1, r = 2, and r 4 is
given in the conclusion.
49
Table 4.14. Summary of Y* when r = 2: The Means, Relative Increase,
Variances, Standard Deviations, and Relative Errors
The Preclinical Durations of Cases Eligible for Detection by Screen
ing
Values of 5
E(Y*) A4 1 2 3 4 5
0.5 0.75 0.84 0.90 0.95 1.00
Sampled Means 1.0 1.26 1.49 1.61 1.68 1.74
2.0 2.22 2.53 2.79 2.99 3.13
5.0 5.13 5.39 5.70 6.01 6.32
10.0 10.07 10.25 10.49 10.78 11.08
Relative Increase 1 2 3 4 5
0.5 49 68 79 90 100
(iip 1)100 1.0 26 49 61 68 74
2.0 11 26 40 49 56
5.0 3 8 14 20 26
10.0 1 3 5 8 11
Var(Y*) A4 1 2 3 4 5
0.5 0.15 0.18 0.18 0.19 0.18
Variances of 1.0 0.48 0.58 0.68 0.72 0.74
sampled sojourn 2.0 1.91 1.94 2.10 2.33 2.54
times 5.0 12.28 12.00 11.86 11.90 12.10
E(Y*2) (E(Y*))2 10.0 49.70 49.11 48.50 48.00 47.63
sd(Y*) A4 1 2 3 4 5
0.5 0.38 0.42 0.43 0.43 0.43
Standard Deviations 1.0 0.70 0.76 0.82 0.85 0.86
of sampled sojourn 2.0 1.38 1.39 1.45 1.53 1.59
times 5.0 3.50 3.46 3.44 3.45 3.48
10.0 7.05 7.01 6.96 6.93 6.90
Relative Error A4 1 2 3 4 5
0.5 8 20 22 22 22
(si(Y)lfc 1)100 1.0 2 8 16 20 21
2.0 2 2 3 8 13
5.0 1 2 3 2 2
10.0 0 1 2 2 2
50
Table 4.15. Summary of Y* when r = 4: The Means, Relative Increase,
Variances, Standard Deviations, and Relative Errors
The Preclinical Durations of Cases Eligible for Detection by Screen
ing
Values of S
E(Y*) M 1 2 3 4 5
0.5 0.65 0.71 0.77 0.82 0.87
Sampled Means 1.0 1.15 1.31 1.37 1.43 1.48
2.0 2.08 2.30 2.49 2.61 2.69
5.0 5.02 5.11 5.29 5.51 5.74
10.0 10.00 10.03 10.11 10.23 10.39
Relative Increase V 1 2 3 4 5
0.5 31 43 53 64 74
(Sni 1)100 1.0 15 31 37 43 48
2.0 4 15 24 31 34
5.0 0 2 6 10 15
10.0 0 0 1 2 4
Var(Y*) 1 2 3 4 5
0.5 0.14 0.21 .028 0.36 0.43
Variances of 1.0 0.26 0.34 0.40 0.44 0.48
sampled sojourn 2.0 0.96 0.96 1.06 1.18 1.26
times 5.0 6.21 6.04 5.86 5.77 5.82
E(Y*2) (E(Y*))2 10.0 24.98 24.83 24.53 24.14 23.74
sd(Y*) P 1 2 3 4 5
0.5 0.26 0.28 0.28 0.28 0.27
Standard Deviations 1.0 0.48 0.53 0.55 0.55 0.55
of sampled sojourn 2.0 0.97 0.96 1.01 1.06 1.09
times 5.0 2.49 2.46 2.42 2.40 2.40
10.0 5.00 4.98 4.95 4.91 4.87
Relative Error 1 2 3 4 5
0.5 6 10 10 10 1
(sd(y*)/f 1)100 1.0 4 6 10 10 10
2.0 3 4 1 6 9
5.0 0 2 3 4 4
10.0 0 0 1 2 3
51
5. Conclusion
The effect of lengthbiased sampling on the survival among screen
detected cases is not trivial. The mean increase in survival time among cases
detected by screening can be as large as 79% when the screening interval (5)
is twice the mean sojourn time (/i) in the general population. This increase is
only 47% when the screening interval is equal to the mean sojourn time, and
decreases to 0 as 5 > 0 (continuous screening). The standard deviations are
less affected, but some increases, particularly with the exponential distribution,
are observable, on the order of 13% when 5 = 2/j, and 4% when S = /i.
When the sojourn time distribution is gamma with values of r > 2
the effect of lengthbiased sampling is reduced. Studies suggest that r is likely
(o be less than 2 (Zelen and Feinlieb), so these results should be taken into
consideration when evaluating the potential benefits of a screening program,
lengthbiased sampling can cause screening to appear more beneficial than it
actually is and therefore can result in overoptimistic conclusions concerning
the benefit of screening programs.
52
NOTATION INDEX
coefficient of variation: measures the amount of variability relative to the
value of the mean: CV = (j/fi
6: Delta, the length of time between screenings
E(Yj): the mean preclinical duration among cases that started before <5 but
no later than jS the time of the jth screening.
E(Y*): mean length of preclinical durations of the sampled cases
fy(y): the probability density function (pdf) of all preclinical durations.
Fy(y): the cumulative distribution function, (cdf), for the all preclinical
durations
HIP: Health Insurance Plan of Greater New York; a prepaid comprehen
sive group medical plan insuring approximately 700,000 city, state and
federal government employees and their families.
incident cases: those cases of disease occurring twelve of more months af
ter the previous CNBSS screening examination
53
interval cases: those cases of disease occurring less than twelve months
after a negative screening exam.
j: an integer representing the jth screening for a disease
lead time: forward recurrence time; length of time by which the diagnosis
is advanced over clinical detection by virtue of the screening procedure
length bias: biased sample caused by periodic sampling; observations of
longer duration are more likely to be detected than those of short du
ration.
power of an hypothesis test: The ability of a test to reject the null hy
pothesis when the alternative hypothesis is true is called the power of
the test.
Sc: the clinical phase of disease characterized by overt signs or symptoms
sensitivity: the proportion of individuals designated positive by the screen
ing test among all individuals who have the disease
significancelevel or. The probability of rejecting the null hypothesis when
the null hypothesis is true.
SD: diseasefree state of a disease characterized by being either free of the
disease or having disease characteristics of that are undetectable by a
screening test
54
sojourn times: preclinical durations
Sp: preclinical state of disease; Symptoms are asymptomatic.
specificity: the proportion designated negative by the test among all those
who do not have the disease
Y: the random variable that denotes the preclinical duration of a case of
disease in the general population
Y*: the random variable that denotes the the preclinical duration for a
screendetected case
55
A. APPENDIX
The values found in the following tables, A.l A.4, were used to calculate the
Summary table 3.9.
The values found in tables A.5 A.10 were used to calculate the Sumrriary
Table 4.15.
56
Table A.l. Exponential Distribution Ratios E(YJ*)/fj,
Mean Sojourn Time through the jth Screening / /j,
Values of 5
i 2 3 4 5
E(Yi)/ 0.5 1.69 1.93 1.99 2.00 2.00
Â£(*i)/1.0 1.42 1.69 1.84 1.93 1.97
E(Yi)/2.0 1.23 1.42 1.57 1.69 1.78
E{YX)15.0 1.10 1.19 1.27 1.35 1.42
^(yo/io.o 1.05 1.10 1.14 1.19 1.23
E(Y2)/0.5 3.69 5.93 7.99 10.00 12.00
e(y2)/i.o 2.42 3.69 4.84 5.93 9.97
E(Y2)/2.0 1.73 2.42 3.07 3.69 4.28
E(Y2)/5.0 1.30 1.59 1.87 2.15 2.42
E(Y2)/10.0 1.15 1.30 1.44 1.59 1.73
^(y3)/o.5 5.69 9.93 13.99 18.00 22.00
E(Y3)/1.0 3.42 5.69 7.84 9.93 11.97
E(Y3)/2.0 2.23 3.42 4.57 5.69 6.78
E{Y3)/5.0 1.50 1.99 2.47 2.95 3.42
E(Y3)/10.0 1.25 1.50 1.74 1.99 2.23
E(Y4)/0.5 7.69 13.93 19.99 26.00 32.00
E(Y4)/1.0 4.42 7.69 10.84 13.93 16.97
E(Y4)/2.0 2.73 4.42 6.07 7.69 9.28
Â£(y4)/5.o 1.70 2.39 3.07 3.75 4.42
E(y4)/io.o 1.35 1.70 2.04 2.39 2.73
Table A.2. Exponential Distribution Variances: VariY'*)
Values of 5
1 2 3 4 5
0.5 0.32 0.42 0.48 0.49 0.50
1.0 1.08 1.28 1.50 1.70 1.83
2.0 4.08 4.32 4.67 5.10 5.56
5.0 25.08 25.33 25.74 26.29 26.98
10.0 100.08 100.33 100.75 101.32 102.06
57
Table A.3. Exponential Distribution Standard Deviations for Y?
Values of 5
1 2 3 4 5
0.5 0.56 0.65 0.69 0.70 0.71
1.0 1.04 1.13 1.23 1.30 1.35
2.0 2.02 2.08 2.16 2.26 2.36
5.0 5.01 5.03 5.07 5.13 5.19
10.0 10.00 10.02 10.04 10.07 10.10
Table A.4. Exponential Distribution (Standard Deviations (Y*))/fj,
Values of S
/* 1 2 3 4 5
0.5 1.13 1.30 1.38 1.41 1.411
1.0 1.04 1.13 1.23 1.30 1.35
2.0 1.01 1.04 1.08 1.13 1.18
5.0 1.00 1.01 1.01 1.03 1.04
10.0 1.00 1.00 1.00 1.01 1.01
58
Table A.5. Gamma Distribution Means r = 4 for j = 1, 2, 3,4
Values of 5
E(Y?) V 1 2 3 4 5
E(Yi) 0.5 0.61 0.62 0.63 0.63 0.63
1.0 1.12 1.23 1.25 1.25 1.25
2.0 2.05 2.23 2.38 2.46 2.49
5.0 5.01 5.08 5.21 5.39 5.58
10.0 10.00 10.02 10.07 10.15 10.27
EK) 0.5 1.34 2.30 3.28 4.27 5.27
1.0 1.73 2.67 3.62 4.59 5.57
2.0 2.49 3.47 4.41 5.34 6.28
5.0 5.14 5.79 6.70 7.69 8.66
10.0 10.03 10.29 10.83 11.58 12.46
E(Yi) 0.5 2.29 4.27 6.27 8.26 10.26
1.0 2.63 4.59 6.56 8.55 10.62
2.0 3.25 5.25 7.22 9.18 11.15
5.0 5.53 7.17 9.13 11.14 13.13
10.0 10.16 11.06 12.55 14.34 16.27
EM) 0.5 3.28 6.27 9.26 12.26 15.26
1.0 3.58 6.56 9.54 12.53 15.53
2.0 4.14 7.16 10.15 13.12 16.10
5.0 6.12 8.85 11.86 14.90 17.91
10.0 10.42 12.23 14.81 17.69 20.69
59
Table A.6. Gamma Distribution Ratio:
E(Yf)
when r = 4
Va
ues of <5
E(Y?)/p V 1 2 3 4 5
EQKVv 0.5 1.23 1.25 1.25 1.25 1.25
1.0 1.12 1.23 1.25 1.25 1.25
2.0 1.03 1.12 1.19 1.23 1.24
5.0 1.00 1.02 1.04 1.08 1.12
10.0 1.00 1.00 1.01 1.02 1.03
Â£(*?)// 0.5 2.67 4.59 6.56 8.55 10.54
1.0 1.73 2.67 3.62 4.59 5.57
2.0 1.25 1.73 2.21 2.67 3.14
5.0 1.03 1.16 1.34 1.54 1.73
10.0 1.00 1.03 1.08 1.16 1.25
Â£)//* 0.5 4.59 8.55 12.53 16.52 20.52
1.0 2.63 4.59 6.56 8.55 10.62
2.0 1.63 2.63 3.61 4.59 5.57
5.0 l.ir 1.43 1.83 2.23 2.63
10.0 1.02 1.11 1.25 1.43 1.63
E(Y:)/ 0.5 6.56 12.53 18.52 24.52 30.52
1.0 3.58 6.56 9.54 12.53 15.53
2.0 2.07 3.58 5.07 6.56 8.05
5.0 1.22 1.77 2.37 2.98 3.58
10.0 1.04 1.22 1.48 1.77 2.07
Table A.7. Gamma Distribution r = 4 E^Y ^
Values of 5
1 2 3 4 5
0.5 1.31 1.43 1.53 1.64 1.74
1.0 1.15 1.31 1.37 1.43 1.48
2.0 1.04 1.15 1.24 1.31 1.34
5.0 1.00 1.02 1.06 1.10 1.15
10.0 1.00 1.00 1.01 1.02 1.04
60
Table A.8. r = 4 Variances for E(Y'?) for j = 1,2,3,4
Values of 5
Var(Y*) /J 1 2 3 4 5
Var(Y*) 0.5 0.07 0.08 0.08 0.08 0.08
1.0 0.23 0.28 0.31 0.31 0.31
2.0 0.96 0.93 1.02 1.13 1.20
5.0 6.22 6.08 5.90 5.81 5.84
10.0 24.99 24.90 24.67 24.33 23.95
Var (I?) 0.5 0.05 0.04 0.04 0.04 0.04
1.0 0.19 0.20 0.18 0.17 0.16
2.0 0.79 0.76 0.81 0.80 0.77
5.0 5.92 5.17 4.75 4.68 4.78
10.0 24.80 23.70 22.11 20.67 19.63
yar(y3*) 0.5 0.04 0.04 0.04 0.03 0.03
1.0 0.16 0.17 0.16 0.15 0.14
2.0 0.64 0.64 0.67 0.66 0.64
5.0 5.35 4.26 3.90 3.89 4.01
10.0 24.18 21.38 18.78 17.05 16.07
Var (5^*) 0.5 556.04 1153.12 1752.08 2351.56 2951.24
1.0 258.19 556.04. 854.14 1153.12 1452.50
2.0 106.86 258.19 407.40 556.04 704.95
5.0 22.33 76.93 137.19 197.90 258.19
10.0 4.21 22.33 48.13 76.93 106.86
61
Table A.9. r = 4 Ratio of Sampled Standard Deviations/^
Values of 5
e(y;) V 1 2 3 4 5
E(Y1*) 0.5 1.06 1.12 1.12 1.12 1.12
1.0 0.97 1.06 1.11 1.12 1.12
2.0 0.98 0.97 1.01 1.06 1.10
5.0 1.00 0.99 0.97 0.96 0.97
10.0 1.00 1.00 0.99 0.99 0.98
TO 0.5 0.90 0.82 0.79 0.77 0.76
1.0 0.87 0.90 0.86 0.82 0.80
2.0 0.89 0.87 0.90 0.90 0.88
5.0 0.97 0.91 0.87 0.87 0.87
10.0 1.00 0.97 0.94 0.91 0.89
E*) 0.5 0.82 0.77 0.75 0.74 0.73
1.0 0.80 0.82 0.79 0.77 0.76
2.0 0.80 0.80 0.82 0.82 0.80
5.0 0.92 0.83 0.79 0.79 0.80
10.0 0.98 0.92 0.87 0.83 0.80
Â£(K*) 0.5 0.78 0.75 0.74 0.73 0.72
1.0 0.76 0.78 0.76 0.75 0.74
2.0 0.75 0.76 0.78 0.78 0.77
5.0 0.87 0.77 0.74 0.75 0.76
10.0 0.96 0.87 0.81 0.77 0.75
Table A.10. Gamma Distribution r = 4 Ratio Standard Deviations of Y*/a
Values of 5
V 1 2 3 4 5
0.5 1.06 1.10 1.10 1.10 1.10
1.0 0.96 1.06 1.10 1.10 1.10
2.0 0.97 0.96 1.01 1.06 1.09
5.0 1.00 0.98 0.97 0.96 0.96
10.0 1.00 1.00 0.99 0.98 0.97
62
REFERENCES
[1] Jay Devore. Probability and Statistics for Engineering and the Sciences.
Brooks/Cole Publishing Company, Pacific Grove,California, 1991.
[2] M Feinlieb and M. Zelen. On the theory of screening for chronic diseases.
Biometrika, 56(3):601613, 1969.
[3] M. C. Jones. Kernal density estimation for length biased data. Biometrika,
78(3) :511519, 1991.
[4] Karen Kafadar and Philip C. Prorok. Effect of length biased sampled
sojourn times on the survival distribution from screendetected diseases.
Work performed for the National Cancer Institue, Biometry Branch, Divi
sion of Cancer Prevention and Control. This work is still in progress, May
1999.
[5] Anthony B. Miller, Teresa To, Cornelia J. Baines, and Claus Wall. Cana
dian national breast screening study2: 13year results of a randomized
trial in women aged 5059 years. Journal of the National Cancer Institute,
92(18) :14901499, 2000.
[6] Alan Morrison. Sequential pathogenic components. American Journal of
Epidemiology, 109(6):709718, 1979. A formula for the mortality rate is
derived. The components of increased survival time can be separated that
is the benefit, time, lead time and length bias can be separated.
[7] Joseph Petruccelli, Balgobin Nandram, and Minghui Chen. Applied Statis
tics for Engineers and Scientists. Prentice Hall, Upper Saddle River, New
Jersey, 1999.
[8] Philip C. Prorok and Robert J. Conner. Screening for the early detection
of cancer. BioStatistics Cancer Investigation, 4(3):225238, 1986.
[9] Sam Shapiro, Wanda Vent, Phillip Strax, and Loius Venet. Periodic Screen
ing for Breast Cancer the Health Insurance Plan Project and Its Sequelae.
John Hopkins University Press, Baltimore, Maryland, 1988.
63
