Clark Glymour

Department of Philosophy
Carnegie-Mellon University

Costner and The Structure of the Unobserved

Thursday, May 22 at 7:30pm in Physics-Astronomy A110

At the turn of the century Charles Spearman introduced a strategy for finding unobserved causes by examining statistical constraints satisfied by their observed effects. Spearman's ideas were developed over the next thirty years until they were displaced by factor analysis. In the 1960s and after, Herbert Costner and his collaborators generalized and revived latent variable model selection based on constraints satisfied by measured variables. Using directed graphical models, a decade later a group of us at Carnegie Mellon developed general methods for computing a class of constraints implied by latent variable models and used them in a general algorithm for elaborating latent variable models. These Costner-style search procedures considerably outperform searches based on fit statistics in LISREL and similar programs. More recently, Scheines found a constraint based search algorithm that finds univariate linear latent variable models where they exist, and Scheines and Spirtes have developed algorithms that, given purified measurement models for each latent variable, under appropriate conditions describe the causal relations among the latent variables. Still more recent results of Richardson's permit the procedures to be extended to "non-recursive" latent variable models.