Purpose: Latent class analysis (LCA) is commonly used to develop clinical disease typologies designed to assign subjects to subgroups with similar health status profiles. While the validity of such typologies is sometimes examined at a given time point, stability over time is seldom assessed. We examined the consistency of LCA classifications at baseline and after 3 years follow-up for a cohort of chronic alcohol dependent subjects.
Method: We performed a series of clustering analyses using all participants in the National Epidemiologic Survey for Alcohol and Related Conditions who were alcohol dependent both at baseline (Wave 1) and 3 years later (Wave 2) (N=529). We performed two latent class analyses, first with data from Wave 1, second with data from Wave 2, and examined the consistency of individual subject cluster assignments. Next, we used discriminant analysis to assign subjects, using their Wave 2 data, into the clusters defined using Wave 1 data, and again examined the consistency of cluster assignments. To assess the extent to which clusters were distinct and stable over time, we used Euclidean distances to calculate the size of clusters relative to the distances between clusters.
Result: Consistency between cluster assignments based on Wave 1 and Wave 2 data was poor using LCA. Of subjects assigned to the same cluster at Wave 1, fewer than half were assigned to common clusters at Wave 2. When using discriminant analysis, approximately 80% of the subjects were correctly classified. Mean distances for subjects within clusters ranged from 4.18-5.94 for Wave 1 clusters and 4.36-5.74 for Wave 2 clusters, while distances between clusters were relatively small, ranging from 1.69-3.63 for Wave 1 clusters and 1.53-4.33 for Wave 2 clusters.
Conclusion: Typologies for alcohol dependent subjects identified at baseline using LCA were inconsistent with those identified 3 years later for the same subjects. There was very little distance between the clusters at both waves, signifying considerable overlap. Validation of typologies should include not only an examination of stability of identified clusters at a given time point, but also over time.
Candidate for the Lee B. Lusted Student Prize Competition