An empirical method to refine personality disorder classification using stepwise logistic regression modeling to develop diagnostic criteria and thresholds.

This study of DSM-III-R personality disorder (PD) classification provides an empirical approach to determine (1) the discriminative power of each criterion and (2) the optimal number of criteria needed to diagnose the presence of each PD. A semistructured assessment of 110 outpatients was performed for the 11 PDs and their 104 diagnostic criteria. Sensitivity, specificity, and predictive powers were calculated for each criterion item. Logistic regression was performed to determine (1) univariate weightings of the individual criteria as applied to a given diagnosis, and (2) multivariate measures of the criteria that significantly improved the chi-square value in a stepwise fashion. The significant items were then equally weighted to calculate the optimal number needed to diagnose category membership. Of 104 PD criteria, 41 discriminated at a significance level of .05 or less, and each PD could be optimally diagnosed with fewer criteria than currently required. We can empirically reduce the number of criteria combinations comprising individual categories, decrease heterogeneity, and narrow diagnostic boundaries. This increases the likelihood of identifying etiological factors, predictors of clinical course, specific treatments, familial aggregation, and neurobiological correlates for the PD taxa.