Bayesian Estimation and Model Selection in Group-Based Trajectory Models

• In Event: 4351 - Measurement Issues in Social Statistical Models

Abstract

This paper introduces a Bayesian procedure for estimating the standard group-based single trajectory and dual trajectory models with normally distributed outcomes. We demonstrate three key advantages of the Bayesian approach introduced in this paper over the commonly used maximum likelihood estimation approach. First, our sampling algorithms provide the exact finite sample inference, in contrast to maximum likelihood estimators which rely on asymptotic approximations and are prone to overfitting in small samples. Second, we are able to conduct exact posterior inference on arbitrary functions of model parameters, such as the trajectories and transition probabilities, while maximum likelihood methods must resort to methods such as Taylor series approximations or the Bootstrap. Third, we can utilize Bayesian modeling averaging to determine the functional forms of the trajectories. In addition to providing the researchers with additional flexibility for the shapes of the trajectories, model averaging can save researchers a considerable amount of time as they no longer need to fit each functional form separately and compare model fits.

Authors

• Emma Zang, Duke University

• Justin Max, Capital One Financial Corporation