Simulations and directed acyclic graphs explained why assortative mating biases the prenatal negative control design

Abstract

Objective: The negative control design can be used to provide evidence for whether a prenatal exposure-outcome association occurs by in utero mechanisms. Assortative mating has been suggested to influence results from negative control designs, although how and why has not yet been adequately explained. We aimed to explain why mutual adjustment of maternal and paternal exposure in regression models can account for assortative mating.

Study design and setting: We used directed acyclic graphs to show how bias can occur when modeling maternal and paternal effects separately. We empirically tested our claims using a simulation study. We investigated how increasing assortative mating influences the bias of effect estimates obtained from models that do and do not use a mutual adjustment strategy.

Results: In models without mutual adjustment, increasing assortative mating led to increased bias in effect estimates. The maternal and paternal effect estimates were biased by each other, making the difference between them smaller than the true difference. Mutually adjusted models did not suffer from such bias.

Conclusions: Mutual adjustment for maternal and paternal exposure prevents bias from assortative mating influencing the conclusions of a negative control design. We further discuss issues that mutual adjustment may not be able to resolve.

Keywords: Assortative mating; Bias; Causal inference; Directed acyclic graphs; Negative control; Prenatal; Simulation.