Three-step statistical analysis of risk factors for 1999-2001 infant mortality among White and Black singletons in the United States

A three-step statistical analysis procedure was used to evaluate the joint effects of social, demographic, and medical risk factors on infant mortality among US non-Hispanic White and non-Hispanic Black singletons using the 1999-2001 linked birth/infant death datasets. Step one involves a series of univariate analyses to select potential prognosis factors for infant mortality. The factors with p-values of 0.2 or less will be included in the second step, where joint effects of multiple factors are assessed in a single-level logistic regression model. The multi-level logistic regression analysis in the third step extends the single-level analyses to quantify county-level variations in infant mortality. County is modeled as a random effect and fixed effects are selected from the second step (p-value ≤ 0.05) using forward selection. SAS PROC NLMIXED is used to fit the model. The quasi-Newton algorithm is applied to maximize the likelihood over random effect with integration carried out via 10-point adaptive Gaussian quadrature. To accelerate the convergence of the computational process, the initial values of coefficients for the entered factors are those estimated from the previous selection step and the single-level logistic regression.

Analysis of 6,732,344 singletons identifies low birth weight, pre-term birth, pregnancy-associated hypertension and hydramnios/oligohydramnios during pregnancy as the most significant risk factors for infant mortality. County-level variations quantified in the multi-level nonlinear logistic regression were significant. Survival analyses among four race-gender subgroups show that the Black male singletons had the worst survival, while White male singletons appeared to have the best survival during the neonatal period.