American JounaJ of Epidemiology Vol. 135, Mo.

9
Copyright © 1992 by The Johns Hopkins University School of Hygiene and Public Health Printed in U.S.A
All rights reserved

Selection of Controls in Case-Control Studies

III. Design Options

Sholom Wachoider,1 Debra T. Silverman,1 Joseph K. McLaughlin,1 and Jack S. Mandel :

Several design options available in the planning stage of case-control studies are

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

examined. Topics covered include matching, control/case ratio, choice of nested case-
control or case-cohort design, two-stage sampling, and other methods that can be used
for control selection. The effect of potential problems in obtaining comparable accuracy
of exposure is also examined. A discussion of the difficulty in meeting the principles of
study base, deconfounding, and comparable accuracy (S. Wacholder et al. Am J
Epidemiol 1992;135:1019-28) in a single study completes this series of papers. Am J
Epidemiol 1992:135:1042-50.

bias (epidemiology); epidemiotogic methods; retrospective studies

In our previous papers, we presented basic MATCHING

principles of control selection (1) and dis- Random sampling from the study base,
cussed different kinds of control groups (2). where controls are chosen independently of
This paper addresses some of the other de- characteristics of the cases, is the simplest
cisions involved in control selection, of strategy for control selection. Matching is an
which the major themes are issues of strati- option that sometimes can improve effi-
fication and efficiency and the effects of ciency in the estimation of the effect of
time. The principles of deconfounding and exposure by protecting against the situation
efficiency are the main concerns in consid- where the distributions of a confounder are
ering some special sampling techniques, in- substantially different in cases and controls
cluding matching, cluster sampling, and (3). However, the improvement is typically
two-stage sampling. Efficiency is paramount small (3), except for strong confounders.
in our discussion of the control/case ratio, There are several other reasons to match.
replacement of controls, and using a single Control of unmeasured confounders.
control group for multiple case series. Con- Identifiable but not quantifiable variables
sideration of the time of membership in the with many categories, such as neighborhood
study base is crucial in the discussion of or telephone exchange, can serve as proxies
nested case-control, case-cohort, and case- for environmental or socioeconomic con-
base designs. founding factors that are difficult to measure
(4). Matching on such a variable may bal-
ance cases and controls with respect to un-
Received for publication May 8, 1991, and in final form known confounders. Use of the cases' iden-
February 11, 1992.
1
tical twins as controls is an extreme form of
Biostatistics Branch, National Cancer Institute, Be- this kind of matching.
thesda, MD.
'Department of Environmental and Occupational Power. Matching can ensure that there are
Health, School of Public Health, University of Minnesota, sufficient controls to estimate an effect in a
Minneapolis, MN. particular subgroup or to identify an inter-
Reprint requests to Dr. Sholom Wacholder, Bostatistics
Branch, National Cancer Institute, 6130 Executive B)vd., action (5). For example, matching on smok-
EPN 403, Rockville, MD 20892. ing instead of choosing controls indepen-
1042
 Design Options for Case-Control Studies 1043

dently of smoking could make it easier to Reduced flexibility in analysis. Stratified

find an interaction between the exposure or matched analyses can be considered, even
and smoking if smoking had a large effect when there was no matching or stratification
on risk and was rare in the population, by in the design. But matching at the design
ensuring a sufficient number of smoking stage reduces the investigator's flexibility
controls. The two-stage design, a generali- during the analysis. For example, over-
zation of matching discussed below, can also matching that occurs in the design cannot
be used to achieve this goal. be corrected in the analysis. Furthermore,
Time comparability. In unmatched stud- matching usually precludes the ability to
ies, it can be difficult to achieve time com- directly estimate or test the effect of the
parability between cases and controls for matching variable as a risk factor and the

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

exposures that vary over time. Matching on fitting of a nonmultiplicative risk model in-
time-related variables provides a simple ref- volving the matching variable and the ex-
erence point for variables based on these posure (5). However, in the multiplicative
exposures. model, it is possible to fit interactions with
Feasibility. Matching may be the most the matching variables. Below, we discuss a
feasible method of obtaining controls. For method that can be used to estimate the
example, when a case is defined as a peri- effect of matching variables and to fit non-
natal death, the next live birth can be chosen multiplicative models, as long as the values
as a control (6). Use of a random sample of of the matching variables are retained for
live births from a computer tape of births subjects who are identified for the study but
would delay the interview of controls, lead- excluded because of the matching criteria.
ing to possible violation of the comparable
accuracy principle as well as possibly in- Variables on which to match
creasing nonresponse. Matching reduces the possibility of severe
Completeness of control for confounding. loss of efficiency due to a major discrepancy
Perfect matching, followed by a matched in the empiric distributions of a strong risk
analysis, results in complete control for a factor between cases and controls. Matching
continuous confounder under a multiplica- should be considered only for risk factors
tive model of the joint effects. Alternative whose confounding effects need to be con-
strategies, such as regression adjustment for trolled for but that are not of scientific in-
the confounder, can result in bias if its effect terest as independent risk factors in the
is misspecified, e.g., if linearity is wrongly study. Matching on variables that are unre-
assumed. Categorization may leave some lated to risk of disease is pointless; it can
residual confounding, but this is of little only reduce a study's efficiency (4). Age, sex,
importance unless there is a substantial gra- and race are often used as matching variables
dient in risk within strata (7). because they are usually strong confounders
On the other hand, matching has several and because their effects are usually well-
disadvantages. known from descriptive epidemiology (10).
Cost. Matching can add cost and com-
plexity to a sampling scheme by requiring
Forms of matching and stratification
extra effort to recruit controls (5).
Exclusion of cases. Matching can result One form of matching is individual
in the exclusion of cases when no matched matching where a selected control must have
control can be found (8, 9), particularly exactly or approximately the same value of
when matching on several variables. the matching factor as the corresponding
Longer study duration. Matching will slow case. Frequency matching or quota match-
down a study when control selection must ing results in equal distributions of the
wait for cases to be identified and for com- matching factors in the cases and the se-
plex matching variables to be obtained for lected controls. For these forms of matching,
cases and potential controls. the control cannot be recruited until the case
1044 Wacholder et al.

is identified. Approximate frequency match- The other main form of overmatching can
ing can begin immediately; it uses the antic- reduce the efficiency of a study by restricting
ipated, rather than the actual, case distribu- the variability of an exposure that is corre-
tion and thereby allows the control selection lated with the exposure under study (16).
process to operate independently of the case This form of "overmatching" can occur even
selection process. However, if some of the when matching per se was not used in the
matching strata are extremely small, approx- selection of controls, as when an overly ho-
imate frequency matching can be wasteful mogeneous base is used for the study.
(II), since the control/case ratio will vary. Miettinen and Cook (17) note that the use
Probability matching (12) defines strata of a variable indicating the presence of yel-
based on the matching variables. A random low fingers, presumed to be related to smok-

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

mechanism is used to select eligible subjects, ing but unrelated to risk of lung cancer (after
with the probabilities of inclusion for each controlling for smoking), would be an ex-
subject determined by the investigators ample of overmatching. There is much less
based on the odds ratios for disease associ- variability in smoking, conditional on pres-
ated with the subject's stratum. This ap- ence of yellow finger, than unconditionally;
proach does not require knowledge of the since having yellow fingers is not a risk
exposure distribution in the cases and allows factor, it does not affect the point estimate
for a more informative analysis (13), as dis- but does reduce efficiency.
cussed below.
RATIO OF CONTROLS TO CASES
Determination of the number of controls
Overmatching to be selected is another important design
decision. It is useful to consider the ratio of
We use the term "overmatching" to refer
controls to cases. There is usually little mar-
to matching that is counterproductive, by
ginal increase in precision from increasing
either causing bias or reducing efficiency
the ratio of controls to cases beyond four
(14). Matching on an intermediate variable (18), except when the effect of exposure is
in a causal pathway between exposure and large (19). In general, the best way to in-
disease can bias a point estimate downward crease precision in a case-control study is to
(7), since the exposure's effect on disease, increase the number of cases by widening
adjusting for (conditional on) intermediate the base geographically or temporally rather
variable, is less than the unadjusted effect. than by increasing the number of controls,
For example, matching on presence of en- because the marginal increase in precision
dometrial hyperplasia in a study of the re- from an additional case is greater than from
lation between estrogen and endometrial an additional control (assuming there are
cancer is overmatching leading to bias (14, already more controls than cases in the
15). Why? The parameter we seek to esti- study). In matched and stratified studies, the
mate is a measure of the impact on disease most efficient allocation of a fixed number
risk of a change in the level of exposure. of controls into strata is usually one that sets
Matching on endometrial hyperplasia effec- the ratio of controls to cases to be approxi-
tively restricts the comparisons of exposure mately equal (4).
to subjects concordant on presence of hy-
perplasia; this does not allow the full impact
REPLACEMENT OF CONTROLS
of estrogen on cancer risk to be assessed,
since presence of hyperplasia itself is strongly Controls who refuse to participate in a
influenced by estrogen use. Matching on a study should sometimes be replaced on ef-
factor that is a surrogate for or a conse- ficiency grounds, as when replacement can
quence of disease or matching on a correlate prevent wasting a case who otherwise would
of an imperfectly measured exposure (15) have no matched control. However, subjects
can also lead to overmatching and bias. who refuse to participate in case-control
 Design Options for Case-Control Studies 1045

studies may have a different exposure distri- case-control study or case-control within a
bution from those who do participate. Re- cohort study (19, 25, 26). This design para-
placing refusers will not increase the validity digmatically satisfies the study base principle
of a study, since refusers will still be ex- since the base is the cohort as it moves
cluded. through time. Typically, for each case, a set
The situation is different when informa- of controls is selected from subjects at risk
tion on the primary exposure, perhaps ob- at the time of disease occurrence of the case.
tained from medical records, is available, The matching in the design allows for tight
but some information on a confounder (per- control of the confounding effects of time in
haps obtained by interview) is not. Then, the analysis. Thus, this design is useful when
the impact of excluding the control is prob- close matching on time is required, as in

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

ably more serious than that of the missing studies of the incidence of a rare disease,
information regarding the confounder, and such as cancer.
the control should not be replaced. When Just as in the calculation of person-years
controls are replaced, reported response in a cohort study, membership in the base
rates should reflect the actual percentage of depends on time (1, 4). A subject is in the
eligible subjects who refuse to participate. base while under follow-up, i.e., when the
subject would be enrolled as a case in the
ONE CONTROL GROUP FOR SEVERAL study upon development of disease. So a
DISEASES
subject cannot be selected before entry to
Use of a single control group for more the cohort, after loss to follow-up, after
than one case series can lead to savings of death, or after becoming a case (unless sub-
money and effort (20-22). Systematic errors sequent occurrences of disease would make
in assembling the control series would pre- the subject a case again, as, perhaps, in a
sumably affect each individual series case-control study of ear infection in young
equally, but the availability of a larger num- children). In nested case-control studies with
ber of controls would increase the precision age matching, a subject chosen to be a con-
of point estimates. While the use of the same trol for a case at a given age should not be
controls for different diseases induces some excluded from the set of controls because of
dependence in the estimates of effect for subsequent development of disease. Thus, a
different diseases (23), no special analysis is control who subsequently develops disease
required, except when comparing the risk can also serve as a case (27).
factors for the various diseases (24). In fact, If the times when a subject is actually in
this strategy can have another advantage; the base are available with the roster, selec-
i.e., it can help to calibrate the control series tion of controls from the base is simply a
by identifying exposures having stronger (or matter of sampling. Sampling can be strati-
weaker) than expected associations with sev- fied according to factors available for all
eral diseases, resulting from special charac- members of the roster at the time of entry
teristics of the control group. The fact that to the cohort. Control selections at the var-
the same control series was used for several ious times of diagnosis of the cases should
diseases should be discussed in the reports be mutually independent and should not be
from the studies, so that readers can judge influenced by future disease status of the
whether findings resulted from the unique subject or by use as a control for another
characteristics of the control series. case (27-29); thus, the same individual can
serve as a control for more than one case
NESTED CASE-CONTROL AND CASE-
(27, 28). These rules mirror the approach in
COHORT STUDIES
the analysis of the proportional hazards
Controls in nested case-control studies model with time-to-event data, where vir-
A particular form of case-control study tually all cases served previously as "com
that, in fact, does have a roster of subjects trols" and virtually all controls are used
available for control selection is the nested more than once (30).
1046 Wacholder et al.

Controls in case-cohort studies sification if information about cases is ob-

tained before that about controls (24, 34).
The case-cohort design is an alternative to Other practical considerations have recently
the nested case-control design with a simpler been discussed (24). The statistical efficiency
sampling scheme but a more complex analy- of the nested case-control design is often
sis (24, 31, 32). In its simplest form, a sub- slightly higher than that for the case-cohort
cohort or random sample from all members design when a single disease is being studied
of the cohort is selected to be the source of (34), except when there are small amounts
all controls. Adjustment for the confounding of censoring and late entry (31). More re-
effects of time is achieved in the analysis, by fined approximations of the efficiency of
comparing the exposures of each case to nested case-control, case-cohort, and related

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

those of a set of controls consisting of all designs can be obtained when the cohort has
members of the subcohort who were at risk been assembled, i.e., when the cases and
at the time of diagnosis of the case. We think their event times, as well as the interval at
of a case-cohort study as a variant of a nested risk for all subjects, are known, but exposure
case-control study where controls are se- information is not yet available (29).
lected without matching on time (24, 32).
The following several advantages of the case-
cohort study are due to the use of "un- CASE-BASE STUDIES
matched" controls. Controls are sampled for the case-base
External comparisons. It is easy to obtain design (20, 35) in the same way as the sub-
estimates comparing the risk in the cohort cohort is selected in the case-cohort design;
with that of an external population (32). For it differs from the standard case-control de-
example, the case-cohort design has been sign only in that control subjects are sampled
used to compare the risk of breast cancer in from the base, regardless of their disease
a cohort of women who received noncontra- status. The case-base design can be thought
ceptive estrogen treatment with the risk for of as a variant of the case-control design that
other women living in the same region (33). allows estimation of the risk ratio, because
Ease of selection. Sampling of controls the exposure odds in the base (not just in
can begin before the roster of subjects and the cases and noncases as in the case-control
the list of cases have been completely iden- study) can be estimated.
tified. As soon as each member of the roster
is identified, randomization can be used im- ARE CASES ELIGIBLE TO BE
mediately to determine inclusion in the sub- CONTROLS?
cohort (24).
Multiple diseases. Since there is no time We have noted several situations, includ-
matching of cases to controls, a single sub- ing the nested case-control, case-cohort, and
cohort can serve as a source of controls for case-base designs, where subjects who qual-
multiple disease types (20, 24, 31). ify as cases can be included as controls. In
Primary time scale. Unlike the nested general, a future case is in the base until his
case-control design, the case-cohort design or her disease is diagnosed and, therefore,
does not require a decision about the pri- should not be excluded from the sampling
mary time scale until the analysis stage (24). of controls for a case diagnosed at an earlier
Thus, for example, all analyses of a nested time.
case-control design with age matching con-
trol for age in a study of a treatment-related TWO-STAGE SAMPLING
second cancer, while controlling for age
alone or time since first cancer alone is Recently, some two-stage sampling de-
possible in a case-cohort design (24). signs have been proposed as economical al-
It is worth noting that for either design, ternatives to standard case-control studies
there is a possibility of differential misclas- (13, 36-38). The savings accrue from not
 Design Options for Case-Control Studies 1047

requiring all exposure and confounder mea- subject's level of the first-stage variable (12,
surements for all subjects; sometimes a sub- 13).
stantial reduction in expense can be A two-stage approach can be more effi-
achieved with little loss in statistical effi- cient than matching for estimating the main
ciency compared with the study with infor- effects of exposures and interactions (37)
mation on all subjects. In these designs, first- and allows for estimation of the effects of
stage variables, i.e., exposure or confounder first-stage variables, in contrast to standard
measurements that are relatively easy to ob- matched studies (13). However, any
tain, are gathered for all subjects. The re- matched study can be viewed and analyzed
maining, second-stage variables are obtained as a special case of the two-stage study, if
on only a subset of subjects, with the sam- information on the matching factor is re-

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

pling fractions depending on disease status tained for all eligible subjects, including
and the first-stage variables. The first-stage those excluded because they did not satisfy
variable might be an exposure that would be the matching criteria. This allows estimation
obtained from a record, while the second- of the main effect of the matching variable
stage variable might be a confounder, such as well as the fitting of nonmultiplicative
as smoking, which required a personal inter- models.
view. Alternatively, the first-stage variable
might be a confounder and the second-stage
CLUSTER SAMPLING
variable might be the exposure (13). This
approach could be helpful, for example, in In cluster sampling, controls are selected
a study of the effect of residential radon in groups, to reduce expense, rather than
exposure on lung cancer risk (12, 13). First- independently (39). Choosing several con-
stage variables might include age and trols who live in the same household or near
smoking. Nonsmoking cases and smoking one another can be economical when less
controls will have higher probabilities of se- effort is needed to include an additional
lection for the part of the study requiring member of the cluster than to include an
expensive fieldwork for residential radon independent control. Thus, cluster sampling
measurements. The power for assessing a might be appropriate for population control
radon-smoking interaction will be en- groups when blood samples are needed. The
hanced, compared with a matched or an clusters themselves must be selected so that
unmatched design, by forcing the propor- each member of the base population has an
tions of cases and controls in the smoking equal chance of being selected (1), and an
and nonsmoking strata to be near 0.5 (13, analysis taking clustering into account must
37). A two-stage design can also be consid- be used (39).
ered where the second-stage variable is a
more refined version of the first-stage vari-
AVOIDING INFORMATION BIAS
able. For example, the probability of obtain-
ing a detailed occupational history can be A widespread concern about interview-
allowed to vary, depending on the subject's based case-control studies is that cases recall
current job title. previous exposures differently than do con-
The two-stage design proposed originally trols. Cases may spend time thinking about
uses random samples of the subjects in each possible reasons for their illness, may search
cell of the cross-classification of the first- their memories for past exposure or even
stage variable and disease to determine exaggerate or fabricate exposure, or may try
which subjects to include for second-stage to deny any responsibility for the disease.
measurements (36-38). An alternative ran- Therefore, some suggest using control
domized recruitment approach uses ran- groups of diseased subjects in the name of
domization, with the probabilities, which are equal accuracy (40). While accuracy of in-
dependent on the approximate odds ratios formation and how that accuracy differs be-
(determined a priori) associated with the tween cases and controls are considerations
1048 Wacholder et al.

in the choice of control group, one must also prevalence would be less likely to cause bias
be concerned about choosing controls with (49, 50). Similarly, matching on calendar
conditions possibly related to exposure (2, time should be considered, if it can ensure
41,42). that exposure measurements are comparable
Unfortunately, the literature on the with respect to time, as in case-control stud-
question of differential recall for cases and ies performed in an occupational setting,
controls is sparse, and the interpretation of where industrial hygiene data from different
the published studies is difficult (43). Most years may be affected by changes in the
recent empiric research suggests that differ- quality of the measuring instruments.
ential accuracy does not cause serious dis-
tortion (40, 44-48). Empiric work on the
DISCUSSION

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

accuracy of recall for a broader range of
variables would help in the decision of what In the first paper of this series (1), we
is the appropriate source of controls in situ- presented four basic principles—study base,
ations when there is a suspicion of differen- deconfounding, comparable accuracy, and
tial accuracy. efficiency—that we believe provide a theo-
In many studies, information about time- retical framework for the evaluation of issues
dependent exposures (variables whose val- in control selection. Various practical prob-
ues can vary over time), such as consump- lems have been addressed, and possible so-
tion of food items or cigarette smoking, lutions have been examined using these
should be obtained in such a way that the principles.
entire history will be available. This is typi- It may be difficult or impossible to satisfy
cally not practical, and, instead, questions all principles in a study. Sometimes an at-
usually refer to a particular period of time. tempt that is feasible turns out to be harmful.
Unless the periods of time correspond for Just as unnecessary matching can reduce
cases and controls, the comparable accuracy efficiency and even cause bias, avoiding vi-
principle may be violated. If controls are olations of principles that are purely theo-
matched to cases on age, questions about retical and have no effect on inference is not
exposures should refer to exposures at the advisable. It is important to remember that
same age for cases and controls. In a diet the validity of a study can be undermined
and cancer study, if a case is asked about more by an equivocal violation of principle
usual diet 5 years prior to the diagnosis of than by a clear violation of principle that
cancer at age 60, the questions to a perfectly results in only minor bias. Since all biases
age-matched control should refer to usual are not created equal, quantification of the
diet at age 55, regardless of the control's age extent of bias is important (51, 52); other-
at interview. For frequency-matched or un- wise, an attempt to avoid a violation of one
matched studies, some sort of average might principle may induce a more serious viola-
be attempted, such as starting exposure tion of another. Therefore, the implications
questions for all subjects with, "Before 1985, of alternative approaches need to be consid-
. . . . " This should be the practice even if a ered carefully. For example, allowing con-
control, who was selected from among those cerns about the theoretical possibility of re-
free of disease at age 60, is not interviewed call bias to determine the type of controls to
until age 63. (Of course, one has to assume choose, when in fact little or no recall bias
that the respondents are answering the ques- may exist, could lead to a more biased study,
tions the way they are asked.) The time if controls were drawn from a diseased group
intervals between interviews of cases and of related to the study exposure.
controls should be similar (and as short as It is important to recognize that develop-
possible), so the elapsed times from the ment of a protocol that deals with the theo-
period to which the questions refer in the retical considerations discussed here is not
interview will be similar in cases and con- enough; careful fieldwork is needed to make
trols; also, any secular trends in exposure sure the study is properly executed. Thus, a
 Design Options for Case-Control Studies 1049

low response rate, particularly when nonre- 13. Weinberg CR, Wacholder S. The design and analy-
sponse might depend on exposure level, may sis of case-control studies with biased sampling.
Biometrics 1990;46:963-75.
violate the study-base principle and threaten 14. Cole P. Introduction. In: Breslow NE, Day NE,
a study's validity. eds. Statistical methods in cancer research. Vol 1.
The "ideal" (53) control group rarely ex- The analysis of case-control studies. Lyon: Inter-
national Agency for Research on Cancer, 1980:14—
ists in epidemiologic studies. Besides addi- 40. (IARC scientific publication no. 32).
tional theoretical work, empiric studies are 15. Breslow N. Design and analysis of case-control
needed to measure the impact of violations studies. Annu Rev Public Health 1982,3:29-54.
16. Day NE, Byar DP, Green SB. Overadjustment in
of the principles so intelligent trade-offs can case-control studies. Am J Epidemiol 1980; 112:
be made when planning a study. We believe, 696-706.
however, that although proper control selec- 17. Miettinen OS, Cook EF. Confounding: essence and

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

detection. Am J Epidemiol 1981;114:593-603.
tion will continue to be problematic, the 18. Ury HK. Efficiency of case-control studies with
most serious mistakes in control selection multiple controls per case: continuous or dichoto-
can be avoided by keeping a few basic prin- mousdata. Biometrics 1975;31:643-9.
19. Breslow NE, Lubin JH, Marek P, et al. Multiplic-
ciples in mind. ative models and cohort analysis. J Am Stat Assoc
1983;78:1-12.
20. Kupper LL, McMichael AJ, Spirtas R. A hybrid
epidemiologic study design useful in estimating
relative risk. J Am Stat Assoc 1975;70:524-8.
REFERENCES 21. Siemiatycki J, Wacholder S, Richardson L, et al.
Discovering carcinogens in the occupational envi-
1. Wacholder S, McLaughlin JK, Silverman DT, et ronment. Scand J Work Environ Health 1987;13:
al. Selection of controls in case-control studies. I. 486-92.
Principles. Am J Epidemiol 1991; 135:1019-28. 22. Greenland S. Control-initiated case-control studies.
2. Wacholder S, Silverman DT, McLaughlin JK, et Int J Epidemiol 1985; 14:130-4.
al. Selection of controls in case-control studies. II. 23. Thomas DC, Goldberg M, Dewar R, et al. Statis-
Types of controls. Am J Epidemiol 1991; 135: tical methods for relating several exposure factors
1029-41. to several diseases in case-heterogeneity studies.
3. Thomas DC, Greenland S. The relative efficiencies StatMed 1986;5:49-60.
of matched and independent sample designs for 24. Wacholder S. Practical considerations in choosing
case-control studies. J Chronic Dis 1983;36: between the case-cohort and nested case-control
685-97. design. Epidemiology 1991,2:155-8.
4. Miettinen OS. Theoretical epidemiology: principles 25. Mantel N. Synthetic retrospective studies and re-
of occurrence research in medicine. New York: lated topics. Biometrics 1973;29:479-86.
John Wiley & Sons, Inc, 1985. 26. Liddell FDK, McDonald JC, Thomas DC. Meth-
5. Thomas DC, Greenland S. The efficiency of match- ods of cohort analysis: appraisal by application to
ing in case-control studies of risk factor interac- asbestos mining (with discussion). J R Stat Soc A
tions. J Chronic Dis 1985;38:569-74. 1977; 140:469-91.
6. Clarke M, Clayton D. The design and interpreta- 27. Lubin JH, Gail MH. Biased selection of controls
tion of case-control studies of perinatal mortality. for case-control analysis of cohort studies. Biomet-
Am J Epidemiol 1981;113:636-45. rics 1984;40:63-75.
7. Breslow NE, Day NE, eds. Statistical methods in 28. Robins J, Gail MH, Lubin JH. More on biased
cancer research. Vol 1. The analysis of case-control selection of controls. Biometrics 1986;42:293-9.
studies. Lyon: International Agency for Research 29. Wacholder S, Gail MH, Pee D. Selecting an effi-
on Cancer, 1980. (IARC scientific publication no. cient design for assessing exposure-disease relation-
32). ships in an assembled cohort. Biometrics 1991 ;47:
8. McKinlay SM. Pair matching—a reappraisal of a 63-76.
popular technique. Biometrics 1977;33:725-35. 30. Cox DR. Regression models and life tables (with
9. Thompson WD, Kelsey JL, Walter SD. Cost and discussion). J R Stat Soc B 1972;34:187-220.
efficiency in the choice of matched and unmatched 31. Prentice RL. A case-cohort design for epidemio-
case-control study designs. Am J Epidemiol 1982; logic cohort studies and disease prevention trials.
116:840-51. Biometrika 1986;73:1-11.
10. MacMahon B, Pugh TF. Epidemiology: principles 32. Wacholder S, Boivin J-F. External comparisons
and methods. Boston: Little, Brown & Company, with the case-cohort design. Am J Epidemiol 1987;
1970. 126:1198-1209.
11. McTiernan A, Thomas DB, Whitehead A, et al. 33. Bergkvist L, Adami HO, Persson I, et al. The risk
Efficient selection of controls for multi-centered of breast cancer after estrogen and estrogen-proges-
collaborative studies of rare diseases. Am J Epide- tin replacement. N Engl J Med 1989;321:293-7.
miol 1986; 123:901-4. 34. Langholz B, Thomas DC. Nested case-control and
12. Weinberg CR, Sandier DP. Randomized recruit- case-cohort methods of sampling from a cohort: a
ment in case-control studies. Am J Epidemiol 1991; critical comparison. Am J Epidemiol 1990,131:
134:421-32. 169-76.
1050 Wacholder et al.

35. Miettinen OS. Design options in epidemiologic al. A dental x-ray validation study: comparison of
research: an update. Scand J Work Environ Health information from patient interviews and dental
1982;8(suppl 1):7-14. charts. Am J Epidemiol 1985;121:43O-9.
36. White JE. A two stage design for the study of the 45. Mackenzie SG, Lippman A. An investigation of
relationship between a rare exposure and a rare report bias in a case-control study of pregnancy
disease. Am J Epidemiol 1982; 115:119-28. outcome. Am J Epidemiol 1989; 129:65-75.
37. Breslow NE, Cain K.C. Logistic regression for two- 46. Linet MS, Van Natta ML, Brookmeyer R, et al.
stage case-control data. Biometrika 1988,75: Familial cancer history and chronic lymphocytic
11-20. leukemia: a case-control study. Am J Epidemiol
38. Cain K.C, Breslow NE. Logistic regression analysis 1989; 130:655-64.
and efficient design for two-stage studies. Am J 47. Drews CD, Kraus JF, Greenland S. Recall bias in
Epidemiol 1988; 128:1198-1206. a case-control study of sudden infant death syn-
39. Graubard B, Fears TR, Gail MH. Effects of cluster drome. Int J Epidemiol 1990; 19:405-11.
sampling on epidemiologic analysis in population- 48. Friedenreich CM, Howe GR, Miller AB. An inves-

Downloaded from aje.oxfordjournals.org at McGill University Libraries on December 2, 2010

based case-control studies. Biometrics 1989;45: tigation of recall bias in the reporting of past food
1053-71. intake among breast cancer cases and controls. Ann
40. Werler MM, Pober BR, Nelson K, et al. Reporting Epidemiol 1991;l:439-53.
accuracy among mothers of malformed and non- 49. Greenland S, Thomas DC. On the need for the rare
malformed infants. Am J Epidemiol 1989;129: disease assumption in case-control studies. Am J
415-21. Epidemiol 1982;116:547-53.
41. Swan SH, Shaw GM. Re: "Reporting accuracy 50. Gail M, Lubin JH, Silverman DT. Elements of
among mothers of malformed and nonmalformed design in epidemiologic studies. In: DeLisi C,
infants." (Letter). Am J Epidemiol 1990; 131: Eisenfeld J, eds. Statistical methods in cancer epi-
935-6. demiology. Amsterdam: North Holland Publishing
42. Berg AT. Re: "Reporting accuracy among mothers Co, 1985:313-23.
of malformed and nonmalformed infants." (Let- 51. Ibrahim MA, Spitzer WO. The case-control study:
ter). Am J Epidemiol 1990; 131:936-7. the problem and the prospect. J Chronic Dis 1979;
43. Harlow SD, Linet MS. Agreement between ques- 32:139-44.
tionnaire data and medical records: the evidence 52. Rothman KJ. Modern epidemiology. Boston: Lit-
for accuracy of recall. Am J Epidemiol 1989,129: tle, Brown & Company, 1986.
233-48. 53. Greenberg BG. The future of epidemiology. J
44. Preston-Martin S, Bernstein L, Maldonado AA, et - Chronic Dis 1983;36:353-9.