QuantDev Brown Bag Seminar Series (In-Person)

Google map

Mathematical Issues in Identifying Latent Variable Models

Abstract: 

Even for simple latent variable models, model identification can be a complex issue. We explore some ideas from algebraic geometry that bear on latent variable model identification to understand challenges in fitting a variety of models. We explore the distinction between generically and strictly identified models. Generically identified models may have regions of the parameter space where the model is difficult to fit, even if technically a unique mode exists. Using examples from factor analysis, latent class analysis, and latent transition analysis, we show that the problematic regions can greatly impact estimation and inference. Researchers can avoid fitting non-identified models, but may not be able to avoid finding them due to empirical under-identification.