Statistical Issues in Modeling Pregnancy Outcome Data

Child Health And Human Development

Investigators

Linked publications & trials

Abstract

To address how best to model a sequence of pregnancy outcomes, we utilized data from the U.S. Collaborative Perinatal Project to identify determinants of infant birth weight and small-for-gestational age (SGA). The CPP Study enrolled approximately 48,197 pregnant women at one of 12 clinical centers in the United States between 1959-1964 (Niswander and Gordon, 1972). For study purposes, we restricted our sample to 2,211 mothers with 2+ consecutively born infants with complete information on study covariates (i.e., clinical site, maternal age, race, pre-pregnancy weight, cigarette smoking, family income, infant sex). [We intentionally constructed this sample to represent the ideal world with complete pregnancy history and covariate data to examine the issue how best to model prior history.] Modeling strategies included: 1) generalized estimating equations (GEE) with working independence; 2) ignoring prior history; 3) treating history as a confounder, and 4) mixed models with a variety of correlation structures. The various approaches resulted in differences in estimated effect size for birth weight (slopes) and robustly estimated standard errors, and odds ratios and 95% confidence intervals for SGA for known biologic determinants of fetal growth. For example, smoking >1 ppd reduced birth weight from 220-259 grams depending upon modeling strategy with standard errors ranging from 2-32. Risk of SGA ranged from an odds ratio of 2.19 to 2.89. The direction of point estimates for some covariates varied by model selected for analysis.