Mediation analysis in sibling designs: an application to the effect of prenatal antidepressant exposure on toddler depression mediated by gestational age at birth

BackgroundMediation analysis requires strong assumptions of no unmeasured confounding. Sibling designs offer a method for controlling confounding shared within families, but no previous research has done mediation analysis using sibling models. MethodsWe demonstrate the validity of the sibling mediation approach using simulation, and show its application using the example of prenatal antidepressant exposure and toddler anxiety and depression, with gestational age at birth as a mediator. We used data from the Norwegian Mother and Child Cohort Study, a cohort comprising 41% of births in Norway between 1999 and 2008 to identify 91,333 pregnancies, of which 25,776 were part of sibling groups. ResultsIn simulations, sibling models were less biased than cohort models in cases where non-shared confounding was weaker than shared confounding, and when stronger non-shared confounding was controlled, but more biased otherwise. In the full cohort, the estimated mean difference in depression/anxiety scale z-scores for natural direct effects (NDE) were 0.31 (95% confidence interval 0.23 to 0.39) and 0.14 (95% CI 0.03 to 0.24), without and with adjustment for non-shared confounders, respectively. The natural indirect effect was 0.01 (95% CI 0.00 to 0.02) after adjustment. Adjustment for shared and non-shared confounding showed similar point estimates with wider confidence intervals (NDE 0.18, 95% CI -0.21 to 0.47; NIE -0.01, 95% CI -0.06 to 0.06). ConclusionsFindings suggest that the modest association between prenatal antidepressant exposure and anxiety/depression is not mediated by gestational age and is likely explained by both shared confounders and non-shared confounders, and chance.