Abstract:
|
Latent class analysis (LCA), an application of mixture modeling, is useful when an underlying variable cannot be directly measured. Enumerating clusters of individuals with similar underlying characteristics is difficult; one popular approach is to use the Bayesian Information Criterion (BIC). Missing data further complicates LCA. This simulation study explored whether differential participant dropout affected enumeration. Outcome data for two equal sized groups were sampled from a normal distribution. A period effect was included in one group; in the other group, increasing probabilities of dropout were tested. In a second series of simulations, a period effect was included in both groups to compare nondifferential attrition. Lowest BIC value was used to determine number of latent groups (two - five). PROC TRAJ (SAS) was used for all simulations. The BIC did not perform well in enumerating latent groups, owing perhaps in part to the relatively large sample size. Nine of 72,000 simulations correctly identified two groups when n = 2000; all others suggested five groups. Performance at smaller sample sizes was also investigated.
|