Accounting for Twins and Other Multiple Births in Perinatal Pharmacoepidemiological Studies Conducted Using Healthcare Administration Data

Background: In perinatal studies, different approaches are available to handle multiple gestation births. However, choice is often limited when using healthcare administration data, as it can be difficult to distinguish individual infant outcomes in a multiple birth.

Objectives: To compare three commonly used approaches to account for multiple births: restriction to singletons, pregnancy-level analysis (counting the presence of ≥1 outcome among multiples), and infant-level analysis including multiples using generalized estimating equations (GEE).

Methods: To compare approaches we estimated relative risks (RRs) using Poisson models in both simulated data and an applied example using US insurance claims data (MarketScan Commercial Claims and Encounters [CCAE]). In Monte Carlo simulations we generated a maternal binary treatment, an indicator for multiple gestation, and a binary infant outcome. In the applied example, in a cohort of pregnancies ending in live birth, we estimated RRs for neonatal intensive care unit (NICU) admission and major congenital malformation with assisted reproductive technology (ART) relative to both intrauterine insemination (IUI) and non-use of ART or IUI. In MarketScan CCAE, infant claims from a multiple birth are often assigned to a single identifier. Therefore, to estimate an infant-level effect in this cohort we excluded multiples linked to only one infant identifier and used inverse probability of selection weighted GEE.

Results: In simulations, RR estimates based on restriction to singletons differed from infant-level analysis when there was mediation by multiple gestation. The differences between RRs from pregnancy-level and infant-level analysis depended on outcome correlation within multiples and was greatest when the association between treatment and multiples was strong. In the applied example, comparing ART to non-use, adjusted RRs for NICU admission were 1.27 (95% CI 1.22-1.32) when restricting to singletons, 1.57 (1.52-1.61) at the pregnancy-level, and 1.69 (1.63-1.76) at the infant-level. Corresponding estimates for malformations were 1.17 (1.10-1.23), 1.29 (1.23-1.35) and 1.19 (1.12-1.26). Smaller differences were observed between approaches relative to IUI with corresponding RRs for NICU of 1.18 (1.09-1.27), 1.25 (1.18-1.33), and 1.23 (1.14-1.33), and RRs for malformations of 1.05 (0.94-1.18), 1.11 (1.00-1.22), and 1.11 (0.99-1.25).

Conclusions: Estimates from these approaches frequently differed, though only to a considerable extent when treatment was strongly associated with multiples. When choosing between approaches important considerations are the estimand of interest, the data available, and the relationship between multiples and both treatment and outcome.