Selection Biases in Perinatal Research: A Comparison of Inverse Probability Weighting, Instrumental Variable and Sibling-Comparison Design.

Abstract

Background: Longitudinal perinatal studies that study the effects of preconception or prenatal treatments on pregnancy outcomes can have inherent forms of selection bias. For example, these studies often restrict analyses to those who had a livebirth, those with a specified gestation duration or those with complete follow-up. These selection factors are often associated with the treatment and have shared causes with the outcome, which may induce bias in estimating causal effects. Though such selection bias can affect all causal inference approaches, what is unknown is how this bias compares in direction and magnitude across different approaches.

Objectives: We conducted a simulation study to assess and compare the direction and magnitude of bias due to censoring across three common analytic approaches: inverse probability weighting (IPW), instrumental variable (IV) and sibling-comparison design.

Methods: We simulated data for various scenarios under two censoring mechanisms (loss to follow-up; and competing events) with a null true causal treatment effect. The simulated scenarios varied in the probability of the censoring mechanism or its strength of association with treatment or outcome. For each scenario, we generated 500 datasets (sample size = 10,000) and calculated the mean bias in risk difference estimates obtained from the three analytic approaches.

Results: Across all approaches, the proportion of censoring had no specific effect on mean bias. However, increasing the association of censoring with treatment or outcome increased the mean bias. The mean bias in all approaches was generally away from the null in the same direction and often to a similar extent (e.g., 0.5 percentage points away from the null in simulated scenarios with moderate association between treatment and censoring). However, in simulated scenarios with strong association between treatment and censoring, IV analyses were meaningfully more biased than IPW and sibling-comparison design analyses, with mean bias reaching two percentage points.

Conclusions: Across the simulated scenarios, the mean bias in all three approaches was generally away from the null in the same direction and often to a similar extent. Thus, triangulating effect estimates from different analytic approaches in perinatal studies is challenging and may lead to invalid interpretations in the presence of selection processes.