|Friday, February 22|
|PS2 Poster Session 2 (with refreshments)||
Fri, Feb 22, 4:45 PM - 6:15 PM
An evaluation of programs in statistical software packages for fitting hierarchical multilevel logistic regression models (302550)
Keywords: multilevel modeling, logistic regression, infant mortality
Hierarchical multilevel modeling is an indispensable method in the analysis of clustered data because it incorporates both individual- and cluster-level predictors and allows for modeling of cluster variance. Various statistical software packages offer programs for fitting mixed models that are often used to estimate multilevel models due to their wide availability and built-in data management capabilities. There are overlaps and differences in functionality of these software solutions. Because they are not designed specifically for fitting multilevel models, they differ in how multilevel models can be formulated, algorithm used for estimating model parameters and how they trade off precision versus computational burden.
Using the modeling of mortality data among premature infants from a multicenter study as an example, we show how hierarchical multilevel logistic regression models can be specified in SAS (PROC NLMIXED and PROC GLIMMIX), STATA (xtmelogit) and R (lme4), and compare the estimation results. Issues such as centering of individual-level predictors, inclusion of center-level predictors, and their effects on center variance will be discussed.