Seasonal Confounding in Studies of Temperature and Preterm Birth: A Simulation Study

Abstract

BACKGROUND AND AIM: Seasonal patterns of conception may confound associations between seasonally-varying exposures, such as temperature and air pollution, and birth outcomes. Our aim was to evaluate three commonly-used study designs (time-stratified case-crossover, time-series, and case-control) and seasonality adjustment methods in this context. METHODS: Simulations were conducted under the null, i.e., there was no causal acute effect of 7-day average temperature on preterm birth. Pseudo birth data were generated from the empirical seasonal patterns of conception of births in the United States (based on publicly available data 1982-1988, with birth dates and reported gestational ages) and then analyzed using a case-crossover, matched case-control (matching on location), or time-series approach, with and without adjustment for seasonality. Conditional logistic regression was used for case-crossover and matched case-control designs, and Poisson regression was used for the time-series design. Seasonality control in the case-control models was adjustment for the study month of conception. In the time-series, we added the number of pregnancies-at-risk (gestational weeks 20-36 weeks) as an offset and logarithms of a weighted probability of birth to adjust for seasonality. RESULTS:In the presence of seasonality of conception, we observed a 1.8% positive bias in the regression coefficient of mean temperature per 10°C increase in the warm season for the case-crossover approach. For the time-series design, the seasonality of conception created a 4.5% positive bias for each 10°C increase in mean temperature, in the warm season analysis. However, the pregnancy-at-risk approach completely adjusted for this bias. No bias was observed in the matched case-control design with or without adjusting for study month. Variance of the coefficients increased in the adjusted models and was higher in the case-control design. CONCLUSIONS:This study suggests that there might be slight residual confounding by seasonality in the time-stratified case-crossover design. Both the adjusted model of time-series and case-control can provide unbiased estimates. KEYWORDS: Birth outcomes, Methodological study design, Environmental epidemiology