Statistical approaches to modeling symptom clusters in cancer patients

Abstract

This study examined statistical methods to identify and quantify symptom clusters in diverse disciplines, discussed methodological issues in symptom cluster research in oncology, and provided guidance to researchers and clinicians as to the choice and conceptual implications of particular methods. Correlation and related measures of association show the mathematical evidence of a concurrent tendency for 2 or more symptoms. Graphical modeling reveals a more concrete image of possible symptom clusters and provides an idea as to how and why they are correlated. Structural equation modeling can be used to identify symptom clusters with a large number of symptoms, complex relationships, and/or directional relationships. Factor analysis can identify groups of symptoms which are interrelated due to a common underlying cause. Cluster analysis can group symptoms which have similar patterns across patients and find clinical subgroups based on symptom experience. The best strategy to study symptom clusters is to combine various methods while recognizing the strengths and limitations inherent in each method. A tight partnership of clinicians, clinical oncology researchers, and statisticians is essential. Designing a research to identify symptom clusters involves practical issues related to levels of measurement, dimensionality, confounding variables, symptom selection, and heuristic versus deterministic search.