350 – Administrative Data, Record Linkage, and Latent Class Models
Weak Identifiability in Latent Class Analysis
Marcus Berzofsky
RTI International
Paul P. Biemer
RTI International/The University of North Carolina at Chapel Hill
Model identifiability requires a model likelihood with a single global maximum. The ability to find unique parameter estimates for latent class models depends on such identifiability. Weak identifiability occurs when there are regions of the likelihood that are “flat.� In that case, a unique global maximum may exist, but so do many local maxima that provide nearly the same value of the likelihood. In this case, it can be very difficult to find the MLEs, which can lead to erroneous model estimates and conclusions. An important cause of weak identifiability is a violation in one of the model assumptions (e.g., local dependence). This research assesses the likelihood of having a weakly identifiable model based on the type and severity of assumption violation using a simulation approach. It also provides suggestions on how to detect weak identifiability and offers some approaches for avoiding local optima in these situations.