Online Program

Return to main conference page
Thursday, February 15
PS1 Poster Session 1 and Opening Mixer Thu, Feb 15, 5:30 PM - 7:00 PM
Salons F-I

A Simulation Study of Violations of the Local Independence Assumption in Latent Class Analyses (303592)

View Presentation View Presentation

*Michael P. Chen, U.S. Centers for Disease Control and Prevention 

Keywords: Latent class analysis, modeling, local independence assumption

Latent class analysis (LCA) is used to identify unmeasured class membership among subjects for whom there is a set of observed variables. LCA is often used to evaluate diagnostic tests in the absence of gold standards. However, the degree of bias of LCA is unclear when the fundamental assumption of local independence is violated.

In this study, 5 datasets (size n=200,000) with 3 diagnostic tests (sensitivities/specificities: .95/.75, .92/.80, .90/.85) and known prevalence (10%) were simulated. One dataset was designed to satisfy the local independence assumption whereas the other four were constructed to violate it (correlation coefficients: .1, .2, .3, .6). LCA results for these datasets are compared to gold standards. Results for the locally independent dataset matched the gold standards (prevalence, sensitivities and specificities), but results for the other 4 datasets showed different degrees of bias in prevalence (7%~17%), sensitivity (1%~16%) and specificity (4%~16%). This study suggests that care should be taken when interpreting LCA results when local independence of the observed variables is questionable.