American Journal of Epidemiology Copyright 2003 by the Johns Hopkins Bloomberg School of Public Health All rights reserved

Vol. 158, No. 4 Printed in U.S.A. DOI: 10.1093/aje/kwg144

PRACTICE OF EPIDEMIOLOGY Case-Crossover and Case-Time-Control Designs in Birth Defects Epidemiology

Sonia Hernndez-Daz1, Miguel A. Hernn2, Katie Meyer2, Martha M. Werler1, and Allen A. Mitchell1
1 2

Slone Epidemiology Center, Boston University School of Public Health, Boston, MA. Department of Epidemiology, Harvard School of Public Health, Boston, MA.

Received for publication October 29, 2002; accepted for publication February 6, 2003.
Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

The case-crossover and the case-time-control designs can be used to evaluate the effect of intermittent exposures on the risk of acute events. To explore how birth defects epidemiology could benefit from these approaches, the authors compared them with a traditional case-control study design that evaluated the association between use of folic acid antagonists during the second and third pregnancy months and the risk of cardiovascular defects. Among 3,870 cases and 8,387 control infants in the Slone Epidemiology Center Birth Defects Study (19761998), the odds ratio was 2.3 (95% confidence interval (CI): 1.4, 3.9). The case-crossover approach compared folic acid antagonist use between the 2-month embryologically sensitive period (case window) and the 2 months preceding the last menstrual period (control window) among mothers of case infants (odds ratio = 1.0, 95% CI: 0.5, 2.0). Although it controls between-person confounding and avoids issues of control selection, this design may be biased by time trends of exposure prevalence during pregnancy. The case-timecontrol design, which adjusts for exposure time trends under certain assumptions, yielded an odds ratio of 2.9 (95% CI: 1.2, 7.2), but it requires controls. In the presence of gestational time trends of exposure, the new designs do not offer clear advantages over the case-control design. abnormalities; case-control studies; pregnancy

Abbreviation: CI, confidence interval.

The causes of birth defects remain largely unknown. To evaluate potential human teratogens, epidemiologists often use cohort or case-control studies. Because specific defects are quite rare, even large cohorts usually contain too few cases to allow meaningful analysis, so case-control methods have emerged as the preferred approach for studying teratogens. With infrequent exposures for most drugs, even casecontrol studies require large sample sizes to evaluate the effect of maternal use of drugs during pregnancy. For example, a recent case-control study conducted by our group found an association between the use of folic acid antagonists and the risk of cardiovascular defects (1). Since multivitamins containing folic acid might reduce the risk of cardiovascular defects (24), the findings are biologically

plausible. To estimate these associations, the study included 3,870 cases and 8,387 controls recruited over 22 years of surveillance. It would be highly desirable to identify alternative study designs that would lead to a faster, easier, and less costly assessment of the association between drug use and birth defects, while maintaining or improving validity. We therefore explored the strengths and limitations of two designs infrequently used in birth defects epidemiology: the casecrossover design (5) and the case-time-control design (6). To do so, we compared the validity, efficiency, and ease of conduct of these designs with our traditional case-control study of the association between folic acid antagonist use and cardiovascular defects.

Correspondence to Dr. Sonia Hernndez-Daz, Slone Epidemiology Center, Boston University, 1010 Commonwealth Avenue, Boston, MA 02215 (e-mail: shernan@bu.edu).

385

Am J Epidemiol 2003;158:385391

386 Hernndez-Daz et al. MATERIALS AND METHODS Study population

The present study is based on infants ascertained by the Slone Epidemiology Center Birth Defects Study in North America between 1976 and 1998 (7). Study subjects included infants with any of a wide variety of defects, who were identified through review of admissions and discharges at major referral hospitals and clinics and through regular contact with newborn nurseries in community hospitals. Physicians of study subjects confirmed the diagnoses. Institutional review board approval was obtained from all participating institutions, and women provided informed consent before the interview.
Outcome definition

Cases consisted of 3,870 infants with one or more anomalies of the cardiovascular system. Infants with coexisting neural tube defects (known to be associated with folic acid), as well as those with defects associated with a syndrome (chromosomal or Mendelian-inherited anomalies, amniotic bands, caudal regression, or twin disruption), were excluded (1).
Exposure assessment and definition

Trained study nurses who were unaware of the hypothesis interviewed the mothers of study infants within 6 months of delivery. The interview included questions on demographic characteristics, the mothers medical and obstetric history, the parents habits and occupations, and a detailed history of the use of medications (prescription and over the counter, including vitamins and minerals) from 2 months before conception through the entire pregnancy. Because most cardiovascular anomalies develop during the second and third months after the last menstrual period (8), mothers were considered exposed if they reported having used folic acid antagonists any time during these 2 months. We assumed no carry-over effect when the drugs were used before this period. We considered as folic acid antagonists drugs known to inhibit the enzyme dihydrofolate reductase (e.g., methotrexate, sulfasalazine, pyrimethamine, triamterene, and trimethoprim) and those that might affect other enzymes in folate metabolism, impair folate absorption, and/or increase folate destruction (e.g., carbamazepine, phenytoin, primidone, and phenobarbital).
Study designs

Ideally, a cohort study would follow a population of pregnant women from conception, ascertain their drug exposure during pregnancy, and then compare the frequency of cardiovascular defects between the offspring of exposed and unexposed mothers. The comparison could be carried out through the computation of the risk ratio of cardiovascular defects. (We will assume for now that the outcome is ascertained in all conceptuses (9).) However, it is important to distinguish the time during which the fetus is at risk of cardiovascular defects (around the second and third lunar

months) from the time one has to follow the woman in order to identify the defects (frequently through the end of pregnancy). In theory, the risks in the numerator and denominator of the risk ratio should correspond to the at risk period only, because earlier or subsequent person-time is immortal (10) (i.e., no structural birth defect develops before or after organogenesis). In practice, the timing of exposure is usually measured with a precision of only a few weeks. Because of this imprecision and the uncertainty about the exact onset date of the outcome, every mother/conceptus pair will be considered either exposed or unexposed during the whole at risk period (i.e., the 2 person-months of follow-up they contribute to the study). Thus, the estimated risk ratio and rate ratio would be identical, and we will not distinguish between them in the following discussion. Case-control design. In the case-control design, a random sample of women who had an infant with a cardiovascular defect, the case subjects, provides an estimate of the exposure distribution in all the cases in the cohort, and a random sample of women who had an infant without a cardiovascular defect, the control subjects, provides an estimate of the frequency of exposure in all the noncases in the cohort. Both cases and controls provide an estimate of the exposure distribution during the same window of time (second and third lunar months in our study); that is, the case-control design is based on time-matched sampling when applied to birth defects research. Exposure frequencies from cases and controls are then compared through an odds ratio (figure 1). In the absence of selection and recall biases, the odds ratio from the case-control design would be equal to the odds ratio from the underlying cohort, except for sampling variability. For rare outcomes such as birth defects, the odds ratio is approximately equal to the risk ratio (11). In our study, because mothers of malformed infants and mothers of nonmalformed infants may differ in their recall of antenatal drug exposure, the controls consisted of infants with malformations other than cardiovascular defects (12). Controls were selected to include a wide variety of defects not believed to be related to folic acid antagonists, so that a previously undocumented effect of folic acid antagonists on some of them would have had little impact on our findings (1). The designs presented below are slightly different strategies to obtain a control group that will provide an estimate of the expected frequency of exposure for the cases under the null hypothesis of no effect of the exposure. In all these designs, study cases are a sample of the cases in the cohort. Case-crossover design. This design includes only cases. No control subjects are required. Each case contributes one case window and one or more control windows. The case window is defined as the at risk period preceding the event (the second and third lunar months in our example). The control windows are periods of the same length as, and not overlapping with, the case window that provide an estimate of the expected frequency of exposure for each case. The case window and the control windows derive from the same person at different times; that is, the case-crossover design is based on subject-matched sampling (5). Exposure frequencies from case windows and control windows are then
Am J Epidemiol 2003;158:385391

Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

Study Designs for Birth Defects Epidemiology 387

FIGURE 1.

Case-control (A), case-crossover (B), and case-time-control (C) designs.

Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

compared through a matched (on subject) odds ratio (figure 1). We used the 2 months preceding the last menstrual period as the primary control window, thus leaving a gap (washout period) of 1 month between the case window and the control window. To explore the sensitivity of the results to the choice of control window, we conducted separate analyses using two additional periods, of 2 months each, as secondary control windows: a period of 2 months preceding the case window (1 month before and 1 month after the last menstrual period) and the fourth and fifth months after the last menstrual period. It should be noted that all control windows lie outside the period during which a subject in the underlying cohort was considered to be at risk (i.e., second and third gestational months). In the absence of changes in exposure frequency, control windows may be sampled outside the period during which the subjects would be considered at risk in the underlying cohort (i.e., proxy controls) (13). This is similar to selecting nonpregnant women as controls for the casecontrol design. Exposed case-crossover. The case-crossover odds ratio can be estimated by the ratio of the number of cases exposed only during the case window to the number of cases exposed only during the control window (i.e., ratio of discordant pairs). Because only discordant pairs contribute to the estimation of the odds ratio in matched analyses, we would obtain the same estimates by including only the cases exposed at least once during the study. We refer to this design as exposed case-crossover. To emulate this design, we used only mothers of case infants in our case-control surveillance study who had used the drugs of interest anytime from 2 months before the last menstrual period to the end of pregnancy. We then compared case windows and control windows of exposure within case subjects as in the case-crossover design. Case-time-control design. This design includes both cases and controls (as defined for the case-control design). Each of them is considered twice, once for the case window
Am J Epidemiol 2003;158:385391

and once for the control window (as defined for the casecrossover design) (6). Within each case subject, exposure frequencies from the case window and the control windows are then compared through a matched odds ratio, as in the case-crossover design. Among control subjects, the frequencies of exposure during the case window and the control windows are an estimate of the exposure distribution in the underlying cohort during two different periods of time, and therefore the matched odds ratio between windows measures the time trend in exposure (on the odds ratio scale) (figure 1) (6). The case-time-control design, developed to remove the bias in the case-crossover design due to time trends in exposure, is based on two main assumptions: 1) the case-crossover odds ratio is the product of an odds ratio due to the causal effect of the exposure on the outcome and an odds ratio due to the time trend in exposure prevalence, and 2) the latter is the same among cases and controls. Thus, the casetime-control odds ratio is the case-crossover odds ratio (from the cases) divided by the time trend odds ratio (from the controls).
Statistical analyses

For the case-control design, we used unconditional logistic regression to estimate odds ratios and 95 percent confidence intervals of cardiovascular defects in relation to folic acid antagonist exposure. To adjust for confounding, we included the interview year, geographic region, maternal diabetes mellitus, age, multivitamin supplementation, and urinary tract infections in the model. Because we were interested in the acute effects of exposure during a particular period, we also adjusted for previous use of folic acid antagonists. For the case-crossover and case-time-control designs, we used conditional logistic regression to estimate matched odds ratios and 95 percent confidence intervals (5, 6). To adjust for characteristics that change over gestational time (time-dependent confounders), we included multivitamin supplementation and urinary tract infections in the model.

388 Hernndez-Daz et al.

Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

FIGURE 2. Folic acid antagonist exposure according to lunar months, in infants with cardiovascular defects and in malformed controls, Slone Epidemiology Center Birth Defects Study, 19761998. LMP, last menstrual period.

We used the standard error of the odds ratio to compare the efficiency of these designs. For comparison purposes, in a separate analysis, we randomly restricted the number of control subjects to one per case subject for the case-control and case-time-control designs.
RESULTS Case-control design

period (table 1). Further adjustment for other factors did not materially influence the estimates.
Case-crossover design

The distribution of folic acid antagonist use for mothers of cases and controls throughout pregnancy is shown in figure 2. Use of folic acid antagonists during the second and third lunar months was associated with an increased risk of cardiovascular defects (odds ratio = 2.3, 95 percent confidence interval (CI): 1.4, 3.9). The odds ratio was closer to the null for use of folic acid antagonists before or after this

The odds ratio for cardiovascular defects and folic acid antagonist exposure was 1.0 (95 percent CI: 0.5, 2.0) when we used the primary control window (table 2). The odds ratio was 0.7 (95 percent CI: 0.3, 1.5) when the control window was defined as the 2 months before the case window, and the odds ratio was 2.8 (95 percent CI: 1.2, 6.2) when the control window comprised the 2 months after the case window (data not shown). The odds ratio estimates did not change materially after controlling for time-varying exposures (infections or folic acid supplementation). By definition, the results using the exposed case-crossover design were identical.

TABLE 1. Case-control design showing the risk of cardiovascular defects according to folic acid antagonist use, Slone Epidemiology Center Birth Defects Study, 19761998
Controls (n = 8,387) Period of folic acid antagonist use No. % Cardiovascular defects (n = 3,870) No. %

OR*,

95% CI*

OR,

95% CI

21 month before last menstrual period 23 months after last menstrual period 45 months after last menstrual period

98 68 62

1.17 0.81 0.74

63 63 49

1.63 1.63 1.27

1.4 2.0 1.7

1.0, 2.0 1.4, 2.9 1.2, 2.5

1.4 2.3 1.0

1.0, 2.0 1.4, 3.9 0.6, 1.8

* OR, odds ratio; CI, confidence interval. Reference category is no use during each of the periods. Adjusted for interview year, geographic region, maternal age, diabetes mellitus, multivitamin supplementation, urinary tract infections or other infections during the first trimester of pregnancy, and use of folic acid antagonists in previous periods.

Am J Epidemiol 2003;158:385391

Study Designs for Birth Defects Epidemiology 389

TABLE 2. Case-crossover and case-time-control designs showing the risk of cardiovascular defects according to folic acid antagonist use, Slone Epidemiology Center Birth Defects Study, 19761998 Case window Cardiovascular defects cases Exposed Control window Exposed Nonexposed OR* (95% CI*) 48 15 1.0 (0.5, 2.0) Nonexposed 15 3,792 52 46 0.3 (0.2, 0.6) 1.0/0.3 = 2.9 (1.2, 7.2) * OR, odds ratio; CI, confidence interval. Case-crossover design and exposed case-crossover design. Case-time-control design. Controls Exposed Nonexposed 16 8,273

Case-time-control design

DISCUSSION

Among control subjects, the frequency of folic acid antagonist use was lower during the case window than during the primary control window (odds ratio = 0.3, 95 percent CI: 0.2, 0.6) (table 2). When we divided the case-crossover odds ratio by the corresponding odds ratio for time trends, we obtained the case-time-control odds ratio of 2.9 (95 percent CI: 1.2, 7.2). For the secondary windows, the odds ratios were 2.7 (95 percent CI: 1.0, 7.7) when the control window was defined as the 2 months before the case window and 2.0 (95 percent CI: 0.7, 5.6) for a control window that comprised the 2 months after the case window (data not shown). The odds ratio estimates did not change materially after controlling for time-varying covariates. A comparison of the efficiency of these designs is presented in table 3. Despite including fewer subjects, the case-crossover and exposed case-crossover designs had standard errors similar to that of the case-control design with one control per case. Standard errors were larger for the casetime-control design.

We compared the findings from three different study designscase-control, case-crossover, and case-timecontrolapplied to the same data. In the case-control and case-time-control designs, women who used folic acid antagonists during the embryologically relevant time period had more than a twofold increased risk of having an infant with a cardiovascular defect. However, the case-crossover design showed no association. These findings might relate to the different design-specific assumptions. The main assumptions, limitations, and relative advantages and disadvantages of these designs are discussed below. The case-control design permits study of a wide range of exposures and outcomes, including chronic exposures and gradual outcomes. The case-crossover and case-time-control designs are appropriate only to evaluate transient exposures with immediate and transient effects in relation to abrupt outcomes (abrupt outcomes being defined as outcomes with a short sensitive time period) (13). Although some drugs are given for acute illnesses (e.g., a 7-day course of trimethoprim for urinary tract infection), others are given for chronic

Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

TABLE 3. Efficiency of different study designs to estimate the risk of cardiovascular defects according to use of folic acid antagonists, Slone Epidemiology Center Birth Defects Study, 19761998
No. of subjects Study designs Cases Controls Odds ratio 95% confidence interval Log odds ratio standard error

Case-control All controls One control per case Case-crossover Exposed case-crossover Case-time-control All controls One control per case 3,870 3,870 8,387 3,870 2.9 2.6 1.2, 7.2 0.9, 7.8 0.4664 0.5531 3,870 3,870 3,870 104 8,387 3,870 0 0 2.3 2.1 1.0 1.0 1.4, 3.9* 1.0, 4.2 0.5, 2.0 0.5, 2.0 0.2980 0.3591 0.3652 0.3652

* Adjusted for interview year, geographic region, maternal age, diabetes mellitus, multivitamin supplementation, and urinary tract infections or other infections during the first trimester of pregnancy, as well as use of folic acid antagonist in previous periods.

Am J Epidemiol 2003;158:385391

390 Hernndez-Daz et al.

conditions (e.g., carbamazepine for epilepsy). The latter cannot be evaluated by these designs because of lack of variations in exposure between case windows and control windows. Birth defects can be considered abrupt outcomes by defining the case window as the period of potential excess risk around the embryologically sensitive period for the specific birth defect. Although the short period of vulnerability for a birth defect may qualify as abrupt, these designs would be difficult to apply to conditions such as low birth weight, where the sensitive period may last through most of gestation. The case-control design can deal with time trends of exposure. An inherent assumption in the case-crossover design (and in the exposed case-crossover design) is a constant subject-specific exposure distribution. This condition can be particularly troublesome for drug-safety studies during pregnancy, given the gestational time trends in drug use. The use of many drugs decreases after conception, which may bias estimates toward a negative association when the control windows precede the case window. In our study, the casecrossover odds ratio for folic acid antagonist use was most likely an underestimate, because folic acid antagonist use decreased between the primary control window and the case window. Moreover, the three control windows considered in our case-crossover analysis led to different estimates (odds ratios of 1.0, 0.7, and 2.8); changes in time trends, together with chance, might explain the differences. The bias would go in the opposite direction for drugs with pregnancy-related indications (e.g., nausea or vomiting) or drugs thought to be safe for the fetus (e.g., amoxicillin), where use in pregnancy might increase relative to prepregnancy. A bidirectional sampling of control windows (i.e., before and after the case window) would eliminate part of the bias from exposure gestational time trends were the trends monotonic (13, 14). However, if the outcome (e.g., prenatal diagnosis of a birth defect) affects subsequent exposure, then control time periods should precede the case window. The case-timecontrol design was developed to remove the bias due to time trends, and the case-time-control estimate (odds ratio = 2.9) might have removed the effects of time trends from the casecrossover estimate. This design assumes that the reduction in folic acid antagonist use is equal among cases and controls (in the odds ratio scale) under the null hypothesis. However, this assumption might not always hold in practice (15, 16). While the case-crossover design avoids the difficulties of selecting a valid group of control subjects (11), one still must define a control window, which may be troublesome because there may be induction periods and carry-over effects. Leaving a washout gap between the control window and the case window might avoid potential carry-over effects and will decrease the number of cases whose exposure bridged periods (and therefore do not contribute information). However, the time trends of use and differential recall between windows might pose problems if the control window were substantially farther back in time than the case window. All the designs might be affected by nondifferential misclassification of the exposure. In fact, under certain circumstances, matched studies can be more sensitive to misclassification than unmatched ones (17). However,

differential recall may vary with each design. The casecrossover design, unlike the case-control and the case-timecontrol designs, avoids differential recall between mothers of cases and mothers of controls (5), but within person recall of exposures may be different across study periods; for example, mothers may be particularly aware of exposures during the first months of pregnancy. On the other hand, the case-time-control design would remove the bias due to trends in recall between the case window and the control windows within subjects only if the relative accuracy between windows (in a multiplicative odds ratio scale) were equal for cases and controls (15). Designs that use each case as her or his own control (i.e., case-crossover and case-time-control) eliminate betweenperson confounding (due to factors that vary among persons) (5). In pregnancy-related studies, characteristics such as socioeconomic status, maternal age at conception, prior medical or family history, and unknown and/or unmeasured factors that stay constant through pregnancy will also be controlled by this design. In contrast, within-person confounding (due to transient factors correlated in time with the exposure that vary over time within an individual) would not be automatically controlled by these designs. In our casecontrol study, adjusting for measured potential confounders (both constant and transient) did not materially affect the odds ratio estimates. Self-matched designs would have greater value in situations where unmeasured betweenperson confounding is a major concern. Compared with a case-control design, the case-crossover design may reduce the complexity and cost of conducting the study because no control subjects are required (13). Furthermore, using more than one control period per case window or, even better, using the entire case history to estimate the usual frequency would increase the precision of the estimates (at no cost if data have already been collected) (18). However, the self-matching used in this approach can decrease precision if the number of concordant pairs is high (i.e., the exposure is long lasting and bridges periods) (11). The case-time-control design uses the same subjects as would a conventional case-control study, so cost and complexity reductions are not advantages of this design and neither is increase in precision. In our example, estimates from this design had wider 95 percent confidence intervals than those from the case-control design. The exposed case-crossover design is subject to the same limitations and advantages as the case-crossover design but, in certain settings, it may offer a faster answer at less cost. For example, when only series of exposed cases are available (e.g., Food and Drug Administration case reports, clinical series, or cohorts of exposed women identified by teratology information services), one could study the effect of the exposure, but only under the critical assumption that the timing of exposure did not affect reporting. In conclusion, our example suggests that only in certain limited settings would innovative designs be advantageous over the case-control design in birth defects research. In the absence of time trends of exposure during pregnancy, the case-crossover design might reduce biases while decreasing the costs and difficulties for patients and investigators. Under the critical assumptions concerning the time trends of
Am J Epidemiol 2003;158:385391

Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

Study Designs for Birth Defects Epidemiology 391

exposure among cases and controls, the case-time-control design might reduce between-person confounding but it is unlikely to be more efficient than a case-control design. When one cannot be confident of whether the underlying assumptions about exposure can be met, the conventional case-control design may be the safest design.

ACKNOWLEDGMENTS

Partially funded by the National Center for Environmental Health, Centers for Disease Control and Prevention, through a grant to the Massachusetts Center for Birth Defects Research and Prevention, Massachusetts Department of Public Health (National Institute of Child Health and Human Development grant HD27697), and by National Heart, Lung, and Blood Institute grant HL50763. Additional support for the Slone Epidemiology Center Birth Defects Study was provided by Rhone Poulenc Rorer (Collegeville, Pennsylvania), Hoechst Marion Roussel, Inc. (Kansas City, Missouri), Pfizer, Inc. (New York, New York), and the Glaxo-Wellcome Company (Research Triangle Park, North Carolina). Dr. Miguel Hernn was partially supported by National Institutes of Health grant K08 AI49392. The authors thank Dr. Malcolm Maclure for his inspiring talk at the 17th International Conference on Pharmacoepidemiology in 2001 and for his encouraging words in private conversation. They also thank Dr. Murrey A. Mittleman for his comments and Dawn Jacobs, Rachel Wilson, Fiona Rice, Rita Krolak, Sally Perkins, Mary Krieger, Kathleen Sheehan, Karen Bennett Mark, Deborah Kasindorf, Clare Coughlin, Joan Shander, Diane Gallagher, Valerie Hillis, Thomas Kelley, and Nastia Dynkin for their assistance. Finally, the authors are indebted to the medical and nursing staffs at each participating hospital; a list of these hospitals is available from the authors.

REFERENCES 1. Hernndez-Daz S, Werler MM, Walker AM, et al. Folic acid antagonists during pregnancy and risk of specific birth defects. N Engl J Med 2000;343:160814.

2. Shaw GM, OMalley CD, Wasserman CR, et al. Maternal periconceptional use of multivitamins and reduced risk for conotruncal heart defects and limb deficiencies among offspring. Am J Med Genet 1995;59:53645. 3. Czeizel AE. Reduction of urinary tract and cardiovascular defects by periconceptional multivitamin supplementation. Am J Med Genet 1996;62:17983. 4. Botto LD, Khoury MJ, Mulinare J, et al. Periconceptional multivitamin use and occurrence of conotruncal heart defects: results from a population-based, case-control study. Pediatrics 1996;98:91117. 5. Maclure M. The case-crossover design: a method for studying transient effects on the risk of acute events. Am J Epidemiol 1991;133:14453. 6. Suissa S. The case-time-control design. Epidemiology 1995;6: 24853. 7. Mitchell AA, Rosenberg L, Shapiro S, et al. Birth defects related to Bendectin use in pregnancy. I. Oral clefts and cardiac defects. JAMA 1981;245:231114. 8. Moore K, Persaud T. The developing human: clinically oriented embryology. Philadelphia, PA: WB Saunders Company, 1998. 9. Hernn M, Hernndez-Daz S, Werler MM, et al. Causal knowledge as a prerequisite for confounding evaluation: an application to birth defects epidemiology. Am J Epidemiol 2002;155:17684. 10. Walker AM. Observation and inference. An introduction to the methods of epidemiology. Newton Lower Falls, MA: Epidemiology Resources, Inc, 1991. 11. Rothman KJ, Greenland S. Modern epidemiology. 2nd ed. Philadelphia, PA: Lippincott-Raven Publishers, 1998. 12. Mitchell AA. Special considerations in studies of drug-induced birth defects. In: Strom BL, ed. Pharmacoepidemiology. New York, NY: John Wiley & Sons, Inc, 2000:595608. 13. Maclure M, Mittleman MA. Should we use a case-crossover design? Annu Rev Public Health 2000;21:193221. 14. Navidi W. Bidirectional case-crossover designs for exposures with time trends. Biometrics 1998;54:596605. 15. Greenland S. Confounding and exposure trends in case-crossover and case-time-control designs. Epidemiology 1996;7:231 9. 16. Suissa S. The case-time-control design: further assumptions and conditions. Epidemiology 1998;9:4415. 17. Greenland S. The effect of misclassification in matched-pair case-control studies. Am J Epidemiol 1982;116:4026. 18. Mittleman MA, Maclure M, Robins JM. Control sampling strategies for case-crossover studies: an assessment of relative efficiency. Am J Epidemiol 1995;142:918.

Downloaded from http://aje.oxfordjournals.org/ by guest on April 22, 2013

Am J Epidemiol 2003;158:385391