Seminars

General Departmental Seminar Series

On the Specification of Latent Variable Models

Karen Bandeen-Roche, Ph.D., Department of Biostatistics, Johns Hopkins University

Friday, May 9, 2003, 12-1 pm

6225 Medical Sciences Center, 1300 University Ave.

ABSTRACT

Latent variables have long been utilized by behavioral scientists to summarize constructs that are represented by multiple measured variables or are difficult to measure, such as health practices and psychiatric syndromes. Models incorporating latent variables have been regarded as particularly useful when variables that can be measured are highly imperfect surrogates for the construct of inferential interest. However, many statisticians object to latent variable models as overly abstract, weakly estimable and sensitive to unverifiable modeling assumptions, hence producing findings that inadequately reflect empirical data. My talk describes methods for delineating how a given model's assumptions may be mis-specified, and thereby, what are the targets of parameter estimation in the case of a misspecified model. Theory underlying the methods will be described. Specific approaches will be described both for checking the fit of a model's assumptions and for mapping regions where the target of estimation can be uniquely specified. The methods will be applied to data on post-traumatic stress disorder and on disability in older adults. It is hoped that the proposed methodology will help scientists evaluate the degree to which they have measured health and behavior as intended, better understand the processes that determine health and behavior, and effectively communicate their findings to clinical and policy-making audiences.

Back to General Departmental Seminar Series