A comparative study of cluster analytic methods for development and validation of typologies for somatoform disorders in primary care
Description
The problem of somatization in primary care medicine has received considerable attention from researchers in the past few years. Somatizing patients visit the physician frequently with a variety of physical symptoms for which no organic cause can be found. Many researchers consider the current classifications of somatization disorder and undifferentiated somatoform disorder to be arbitrary and inadequate for describing the range of somatizing patients seen in medical settings. An important clinical research goal would be to identify and describe relatively distinct homogeneous subgroups if they exist Cluster analytic techniques, which have been used for classification problems in medicine and psychiatry, have been controversial due to inadequate theoretical basis. Methods have been developed for a statistical test for the presence of cluster structure, employing an internal criterion measure and Monte Carlo simulation of a null distribution. Different clustering outcomes can be compared using bootstrapping. Although these approaches offer solid statistical credibility they have not been widely used in clinical research The purpose of this study was to investigate the efficacy of four hierarchical agglomerative clustering techniques for a clinical research problem. The use of a statistical test for cluster structure is presented. All four methods provided solutions with evidence to support the presence of distinct clusters. Five stable clusters were identified and concordance among the 'best' solutions from each method was high. Bootstrapping techniques were used for parameter estimation and comparison of the clustering methods. The group average method gave the best overall performance when all patients were to be classified. The centroid method performed well when a few outliers could remain unclassified. Ward's minimum variance and complete link methods appeared to be more seriously affected by the presence of outliers and unequal cluster size. Nevertheless, cluster analytic methods provided a reasonable means of identifying empirical subtypes of somatizers in this study