Healthy Live Births Should be Considered as Competing Events in Studies of Medication Use During Pregnancy

Background: In studies of medication use during pregnancy, there is growing recognition that pregnancy loss is an important competing event. Although counterintuitive, a healthy live birth can also function as a competing ‘risk’, making the pregnancy/fetus immune from many adverse pregnancy outcomes. Conceptualizing pregnancy as having a multistate outcome is one approach to addressing this problem.

Objectives: To use simulation to illustrate that conceptualizing pregnancy with multiple outcome states, including “good” late pregnancy outcomes, enables estimation of unbiased treatment effects.

Methods: We simulated 3 scenarios of 1,500,000 pregnancies beginning at conception. Ongoing pregnancies at gestational week 6 entered the trial and were randomly assigned to initiate or not initiate treatment. We simulated data so that treatment decreased the risk of preterm birth, preeclampsia, and small for gestational age (SGA). The effect of treatment on fetal death varied: decreased for scenario 1 (s1), increased for scenario 2 (s2), and null for scenario 3 (s3). We analyzed two outcomes: (1) composite fetal death and pre-delivery preeclampsia and (2) SGA in liveborn infants. We quantified risk and risk differences (RDs), using two estimators: (1) the Kaplan-Meier estimator (KM), censoring at competing events (i.e., the risk if we prevented competing events) and (2) the Aalen-Johansen estimator (AJ) with a multistate outcome (i.e., the risk allowing for competing events). We used a 2-level multistate outcome for the composite outcome (composite outcome and healthy live birth). For SGA, we used a 3-level multistate outcome (fetal death, SGA, and non-SGA live birth) and a 2-level outcome (fetal death and SGA) censoring non-SGA live births. We compared these estimates to the true risks estimated using the full data without censoring.

Results: Each scenario included 1,097,747 pregnancies. When treatment decreased (s1) or did not affect (s3) the risk of fetal death, the KM overestimated the treatment effect for the composite outcome (s3: RD = -9.29 [95% CI: -11.75, -6.83] vs true RD = -3.06), while the AJ did not (s3: RD = -3.07 [95% CI: -3.26, -2.87]). The KM over-estimated the RD for SGA in all scenarios (e.g., s1: RD=-8.21 [95% CI: -11.23, -5.19] versus true RD = -1.29) while AJ estimator with the 3-level outcome did not (e.g., s1: RD=-1.31 [95% CI: -1.43, -1.19]). The RD estimates for SGA from the AJ with the 2-level outcome were biased compared to the truth, though much less precise (s1: RD=1.94 [95% CI: -11.97, 15.84] versus true RD = -1.29).

Conclusions: In addition to accounting for early pregnancy loss as a competing risk, studies of prenatal exposures on pregnancy outcomes must also consider births free of the outcome of interest as competing risks to estimate unbiased treatment effects.