Bayesian Comparison of Growth Mixture Models: To Better Pursue Truth

• In Event: Marriott Roundtable Session Seven
  In Roundtable Session: Data Analysis and Data-Driven Decision Making (Table 1)

Abstract

In Growth Mixture Models selection, it's important to find out how many latent classes are there that represent different trajectory patterns. Researchers often use ad-hoc approaches for the sake of convenience, but these default approaches often do a poor job on class enumeration. First, the frequentist approach hardly rules out the local maxima. Second, the Bayesian model selection results are highly sensitive to the choice of what part of the model is labeled as "the likelihood", and performing model selection via conditional likelihood is hard to justify. Thus, we recommend and justify in this paper that performing Bayesian model selection via marginal likelihood is the most rigorous way to recover the true parameter values and the number of latent classes.

Authors

• Xingyao Xiao, University of California - Berkeley
  Presenting Author

• Sophia Rabe-Hesketh, University of California - Berkeley
  Non-Presenting Author

• Yihong Cheng, Boston College
  Non-Presenting Author

• Li Zhu, University of California - Berkeley
  Non-Presenting Author