PURPOSE: Administrative claim databases are increasingly being used to study the safety of medication exposures during pregnancy. These studies are restricted to live births due to a reliance on algorithms for estimating gestational age that are based on codes associated with live delivery. Conditioning on live birth may induce selection bias when studying the effect of a drug on a pregnancy complication if fetal death is a competing risk for the complication or is caused by the complication. METHODS: We simulated a population of 100,000 pregnancies and estimated the impact of selection bias on relative estimates for the effect of antidepressant exposure on the outcome of preeclampsia. We assumed that the exposure, outcome, and covariates increased the risk of fetal loss. RESULTS: A downward bias in the risk ratio was consistently observed when conditioning on live births. When an unmeasured covariate was assumed to be a common cause of fetal death, antidepressant use, and preeclampsia, the direction of bias varied depending on the strength of the confounding relationship coupled with the selection bias. Despite the very low prevalence of stillbirth, the strength of the relationship between antidepressant use and stillbirth had a substantial impact on bias. CONCLUSIONS: Conditioning on live birth can be problematic when studying pregnancy complications. Simple quantitative selection bias analysis in populations restricted to live births may not fully account for selection bias.
PURPOSE: Administrative claim databases are increasingly being used to study the safety of medication exposures during pregnancy. These studies are restricted to live births due to a reliance on algorithms for estimating gestational age that are based on codes associated with live delivery. Conditioning on live birth may induce selection bias when studying the effect of a drug on a pregnancy complication if fetal death is a competing risk for the complication or is caused by the complication. METHODS: We simulated a population of 100,000 pregnancies and estimated the impact of selection bias on relative estimates for the effect of antidepressant exposure on the outcome of preeclampsia. We assumed that the exposure, outcome, and covariates increased the risk of fetal loss. RESULTS: A downward bias in the risk ratio was consistently observed when conditioning on live births. When an unmeasured covariate was assumed to be a common cause of fetal death, antidepressant use, and preeclampsia, the direction of bias varied depending on the strength of the confounding relationship coupled with the selection bias. Despite the very low prevalence of stillbirth, the strength of the relationship between antidepressant use and stillbirth had a substantial impact on bias. CONCLUSIONS: Conditioning on live birth can be problematic when studying pregnancy complications. Simple quantitative selection bias analysis in populations restricted to live births may not fully account for selection bias.
Authors: Anne Sofie Dam Laursen; Benjamin Randeris Johannesen; Sydney K Willis; Elizabeth E Hatch; Lauren A Wise; Amelia K Wesselink; Kenneth J Rothman; Henrik Toft Sørensen; Ellen Margrethe Mikkelsen Journal: Eur J Nutr Date: 2022-04-24 Impact factor: 4.865
Authors: Dominique Heinke; Janet W Rich-Edwards; Paige L Williams; Sonia Hernandez-Diaz; Marlene Anderka; Sarah C Fisher; Tania A Desrosiers; Gary M Shaw; Paul A Romitti; Mark A Canfield; Mahsa M Yazdy Journal: Paediatr Perinat Epidemiol Date: 2020-04-06 Impact factor: 3.103
Authors: Mollie E Wood; Angela Lupattelli; Kristin Palmsten; Gretchen Bandoli; Caroline Hurault-Delarue; Christine Damase-Michel; Christina D Chambers; Hedvig M E Nordeng; Marleen M H J van Gelder Journal: Epidemiol Rev Date: 2022-01-14 Impact factor: 6.222