Alternative regression approaches to the analysis of medical care survey data

Abstract

In a multivariate analysis of ambulatory care utilization of a subsample of the 1970 National Health Interview Survey (NHIS) data the dependent variables representing utilization, acute conditions and chronic conditions were found to have discrete variable properties violating normality assumptions of standard regression analysis. Focusing on the utilization variable, alternative multivariate approaches were compared with results obtained from standard least squares analysis. These were Poisson-based multivariate regression, logit analysis, and discriminant analysis. While the fixed interval measure of utlization had an L-shaped frequency distribution with considerable departure from normality, it was found that more theoretically appropriate alternatives provided only marginal gains over the standard least squares techniques.