Perinatal & Pediatric
Bias induced by method of pregnancy identification in studies of prenatal exposures using administrative healthcare data Chase D. Latour* Chase Latour Jessie K. Edwards Elizabeth A. Suarez Kim Boggess Mollie E. Wood
Background: Identifying pregnancies in healthcare data has typically required observing pregnancy outcomes (eg, delivery), potentially inducing selection bias in effect estimates for prenatal exposures. Recent efforts use prenatal encounters to identify pregnancies, thus also identifying those with unobserved outcomes (eg, at-home miscarriage). However, including these pregnancies requires appropriate methods to address missing data, which depend on the source of missingness.
Objective: Evaluate bias under two pregnancy identification approaches (outcome-based, prenatal-based) across measured and unmeasured sources of missingness.
Methods: We simulated 5,000,000 pregnancies and estimated the total effect of initiating antihypertensives on the risk of preeclampsia. We generated data for 9 scenarios characterized by the effect of treatment on miscarriage and the cause of missingness: 1) measured hypertension severity, 2) unmeasured miscarriage, and 3) both severity and miscarriage. Treatment decreased the risk of preeclampsia, and 20% of pregnancies were missing outcomes. We then created 3 analytic samples as possible approaches to deal with missing outcomes: restricting to deliveries (excluding miscarriages), restricting to pregnancies with observed outcomes, and including all pregnancies. Treatment effects were estimated using non-parametric direct standardization.
Results: RDs were most biased when restricted to deliveries (range: -0.07, 0.02) and least among all pregnancies (range: -0.04, 0.03). Within the latter group only, bias decreased as the proportion of missingness due to miscarriage decreased and was zero when all missingness was due to severity (-0.04 to 0.03 when 0% versus 0.00 when 100% of missingness was due to severity).
Conclusions: Incorporating pregnancies with unobserved outcomes did not eliminate bias when missingness was due to unmeasured variables. However, analyzing these pregnancies affords additional analytic approaches to investigate the bias.
