Perinatal & Pediatric
Understanding the impact of selection biases inherent in pregnancy research across different causal inference approaches: a simulation study Basma Dib* Basma Dib Ellen Caniglia Sean Brummel Roger Shapiro Sonja Swanson
Randomized trials and observational studies that study the effect of pre-/during-pregnancy treatments on maternal and neonatal outcomes often have inherent forms of selection or collider-stratification bias. For example, these studies often restrict analyses to those who had a livebirth, those with a specified gestation duration, and/or those with complete follow-up. These selection factors and the outcome of interest frequently have unmeasured or even unknown shared causes which may induce bias in estimating the effect of treatment. Though such selection biases can affect all causal inference approaches, what is unknown is the extent to which the biases meaningfully impact different analytic approaches in pregnancy studies. We conducted a simulation study to assess and compare the magnitude and direction of selection biases in a hypothetical study of treatment effect on pregnancy outcomes across three different causal inference methods: inverse probability weighting (IPW), instrumental variable (IV), and sibling comparison design (SCD) analysis. We generated simulated data for various scenarios under two main conditions: (1) presence of loss to follow-up and (2) presence of a competing event. For each scenario, we generated 500 samples of data, each with a sample size of 10,000, and estimated an average causal effect. In presence of loss to follow-up, the mean bias in the risk difference estimates increased with a stronger association of loss to follow-up with treatment and outcome, with estimates obtained from IPW analysis and SCD analysis biased to nearly the same extent and in the same direction. The estimates obtained from IV analysis were consistently more biased by about one percentage point in the same direction. The mean bias varied across which causal estimand was targeted in light of the competing event. Our simulation study provides insight on the extent to which different analytic methods are impacted by selection bias in pregnancy research.