Statistical and methodological considerations when using cluster analysis in neuropsychological research

Abstract

Multivariate classification of variables into mathematically definable and homogenous subsets is often a useful first step in pattern recognition prior to formal statistical analyses of data sets. One such methodology, cluster analysis, has the main goal of clustering entities that share common characteristics and data structure. For example, one goal of such an analysis is to gain insight into the variables that are important in determining group membership so that new data can be easily classified; additionally, one may wish to develop subsets of data that share certain characteristics to facilitate statistical analysis of variables that are hypothesized to be related to clustered entities. As such, cluster analysis can be useful when applied to neuropsychological variables, particularly when an empirical statistical approach to classification is desirable or when significant interindividual differences in neuropsychological function exist within clinical populations. The goal of this chapter is to provide a review of clustering methods, including hierarchical agglomerative methods and iterative partitioning methods. Recommendations for determining the appropriate number of clusters and for comparing clustering methods also will be discussed. Further, validation techniques will be addressed. The chapter will conclude with a discussion of data issues commonly encountered in neuropsychological research, such as non-normality of data and incomplete data records from patients, and techniques for handling these situations. Finally, a table outlining various statistical packages is provided for the reader interested in the availability of software for computation.