Abstract
Cluster analysis may have theoretical and practical value in illness perception research. It is not clear, however, which of the many methods available is the most appropriate for use with illness perception data. A Monte Carlo study was conducted, whereby 420 artificial datasets with a predetermined cluster structure were generated to resemble Revised Illness Perception Questionnaire (IPQ-R) data. Sample size and equality in cluster size were manipulated. Average Linkage, Complete Linkage, Ward's method and K-means (using the number of clusters and cluster centroids derived from Ward's method) were applied to the artificial datasets and the percentage of cases correctly classified in each dataset by each method was recorded. A 4×3×2 ANOVA revealed that K-means cluster analysis was the most appropriate method for use in illness perception research. It is plausible that these results are generalisable to cluster analysis in other similar types of health psychology data.
Acknowledgements
The author would like to thank Dr Rona Moss-Morris for kindly providing the IPQ-R validation dataset and the reviewers for their constructive advice.
Notes
Notes
[1] K-means analysis is not included here as the number of clusters are specified a priori, rather than being deciphered from the cluster output. The number of clusters identified from Ward's cluster analysis in each dataset were subsequently specified for the K-means analysis.
[2] The Greenhouse-Geisser correction works by adjusting the degrees of freedom, not the F-value. For this reason, the degrees of freedom reported are non-integer values.