A Comparison of Different Prior Choices for Estimating the Influence of Minor Factors in Bayesian Structural Equation Models

• In Event: Bayesian Estimation of Structural Equation Models

Abstract

Bayesian structural equation models may be estimated under the assumption that minor factors influence the population covariance matrix responsible for available data. This implies that a correct hypothesized model will have less than perfect fit to data. The influence of minor factors may be captured by regularized estimation of a full matrix of residual covariances. Extant literature has explored a normal prior for residual covariances, which returns an index of model fit. In this paper, we compare other shrinkage priors for estimating the influence of minor factors. We find that the generalized double-Pareto, a global-local prior, has optimal performance for estimating the influence of minor factors, but performs poorly for estimating the model fit index.

Author

• James O. Uanhoro, University of North Texas