Bias due to selection on live births: A simulation study

Abstract

Background: A recent case-control study found a strong protective association between prenatal exposure to air pollution and risk of autism spectrum disorder (ASD). This unexpected finding can be potentially explained by live-birth bias; that is, the non-exchangeability of exposure groups that arises due to the inevitable restriction of the analytic sample to fetuses that survive. This bias has been suggested to manifest from two separate but related selection mechanisms: 1) collider-stratification, where bias is induced by conditioning on a common consequence (i.e., fetal loss) of air pollution and an independent risk factor for ASD, and 2) the preferential depletion of susceptible fetuses for ASD, which can occur due to the misalignment of exposure assignment and eligibility (i.e., exposure is ascertained at birth, but eligibility begins at conception). Methods: We simulated the magnitude of bias under the null from these two hypothetical selection mechanisms, and when they both occur simultaneously. Simulation inputs were based on characteristics of the original case-control study and a range of priors for the prevalence of unmeasured factor U and the odds ratios (ORs) for the selection effects (i.e., the effects of air pollution and U on loss, and U on ASD). Each scenario was simulated 1000 times. Results: We found that the magnitude of bias was small for collider-stratification, slightly stronger for depletion of susceptibles, and was strongest when both mechanisms work together. For example, when the OR for each of the selection effects was 3 and the prevalence of U was 0.75, the observed ORs for ASD and air pollution were 0.95, 0.87 and 0.80 for the three mechanisms, respectively. Conclusion: Our simulations illustrate that live-birth bias may lead to an underestimation of the effects in studies of prenatal exposures and postnatal outcomes, with the extent of the bias depending on the fetal selection process.