Using simulation methods to assess the suitability of the case crossover design in environmental epidemiological studies of preterm birth

Abstract

Background: Case crossover (CCO) studies are widely used in environmental epidemiology. CCO studies are case-only within-person comparisons, and therefore not confounded by time invariant characteristics. Proper inference, then, with CCO studies relies on appropriate selection of control time periods, and the time-stratified design serves as a robust method to account for long- and short-term-time trends. However, CCO studies assume stable baseline outcome risks. Since the baseline risk of birth increases secularly over gestation, the CCO should be evaluated for preterm birth. Our study utilizes simulations of extreme ambient temperature exposures to assess the appropriateness of CCO for preterm birth.Methods: We conducted simulations using 2018 data for New York State. Data were acquired from National Weather Service records for LaGuardia Airport (temperature) and the Centers for Disease Control’s epidemiologic database (birth data). Baseline birth rates per gestational age served as the basis for baseline risk (β0). Baseline risks were then combined with exposure data and simulated effects to create expected counts per day. Relative risks ranged from 0.9 to 1.25 per 10°F increase. We used bootstrapped Poisson random number generation to create 1000 datasets per simulated effect. Counts were disaggregated into individual records for CCO analyses, and estimated via conditional logistic regressions with 2-week and month stratified control period selection.Results: Preliminary results demonstrate upward bias in point estimates of all models. Bias was markedly smaller for 2-week stratified (ranging between 0.18-0.29%) compared to month-stratified (1.18-1.55%) models. Coverage of 95% confidence intervals was higher for 2-week stratified results; between 92.5% and 95.4% of all intervals included the simulated effect. Coverage ranged between 4.9% and 82.6% for month-stratified results.Conclusions: Future analyses will include pooled logistic regression of simulated cohorts and age-varying differences in baseline risk. Characterizing the performance of the CCO under various conditions can improve methodological rigor and innovation.