MODEL DEVELOPMENT OF A PATIENT CLASSIFICATION SYSTEM USING GRAPHIC RESIDUAL ANALYSIS (ARIZONA).

Abstract

The purpose of the research was to explicate the use of graphic residual analysis as one strategy for model development in nursing. The research question concerned the use, decisions made, criteria for those decisions and issues resulting from the use of graphic residual analysis in model building. Graphic residual analysis was used as an exploratory methodology to test and respecify the Arizona Health Sciences Center Patient Classification model. A sample of 852 patient classifications, covering all inpatient units at Arizona Health Sciences Center, with the exception of psychiatry, was used for model building. The model developed was a causal model using regression analysis as the statistical technique. Two major sets of assumptions concerning this approach were tested. The mathematic assumptions of the regression analysis consisted of a zero mean, equal variance, independence and normal distribution of the residuals. The causal model assumptions considered were that residuals from each equation met the mathematic assumption, residuals from one equation were uncorrelated with residuals of any other equation, all relevant variables were in the model, there was no measurement error and the functional relationship was correct. Both sets of assumptions were tested by using graphic residual analysis. The primary contribution of this study to nursing research was to begin to evolve criteria to determine when threats to assumptions were sufficient to cause difficulty with the modeling process and criteria for actions to be taken to correct those threats. Findings included that graphic residual analysis was; effective with models of several levels of explained variance, of assistance in determining the parameters of design matrices in exploratory research and valuable in determining the creation of categorical variables as indicators of populations for which the model did not perform. Criteria for dealing with problems such as measurement error, model redundancy, model closure, and parsimony were evolved.