Confounding variables for the adjusted models were selected in a 2-stage process. First, a series of statistical tests were performed to assess relationships of selected patient characteristics with ART use and each adverse pregnancy outcome. Tests included analysis of variance for continuous confounders and χ2 tests for categorical confounders. Confounders that were significantly (P < .05) associated with either ART use or the given adverse outcomes were considered in the next stage. The list of potential confounders was further reduced using multivariate logistic regression modeling and a backward elimination stepwise regression approach, keeping only those variables that were significant at P < .05.