Multivariate Decomposition for Nonlinear Models

Daniel A. Powers, University of Texas at Austin
Thomas W. Pullum, University of Texas at Austin

This paper illustrates a decomposition method that can be applied in a wide variety of settings, and which is especially useful for nonlinear models, such as logit, probit, and hazard rate models. We draw on extensions of the well-known Oaxaca-Blinder multivariate decomposition to nonlinear regression settings. The proposed methodology has wide applicability in demographic research but has not received a great deal of exposure. Related methods that have been applied in recent work suffer from several shortcomings that are remedied with the proposed methodology. For example, detailed decompositions are free from problems of path dependency, and the technique is valid for predicted probabilities in the tails of a distribution (i.e., rare events). We apply this method to investigate compositional and coefficient components of the differential in infant mortality due to respiratory distress syndrome between 1991 and 1998 and racial/ethnic differentials in composition and return to risk.