Data-Dependent Priors for Latent Growth Models with Small Samples

• In Event: 46.079 - Appropriate Modeling of Latent Growth and Relationships

Abstract

In latent growth models (LGMs) with small samples, the selection of a prior for the factor covariance matrix is crucial since the prior has a greater relative impact on the posterior distributions relative to larger sample sizes. This study discusses small samples in LGMs, prior selection for LGMs with Bayesian estimation, and then conducts a simulation to compare the small sample performance of different priors to each other and to traditional maximum likelihood estimates. An Empirical Bayes approach using data-dependent priors for the factor covariance matrix with small samples is outlined and shown in a simulation to provide reasonable credibility intervals, reduced parameter bias, and increased statistical power over both maximum likelihood and commonly used Bayesian prior distributions.

Author

• Daniel McNeish, University of Maryland