Wang Yan, Kim Eunsook, Hsu Hsien-Yuan
Department of Psychology, University of Massachusetts Lowell, 850 Broadway St, Lowell, MA, 01854, USA.
Department of Educational and Psychological Studies, University of South Florida, Tampa, FL, USA.
Behav Res Methods. 2025 Feb 26;57(4):103. doi: 10.3758/s13428-025-02619-0.
Factor mixture modeling (FMM) has been increasingly adopted in social, behavioral, and health sciences to identify population heterogeneity by incorporating both continuous latent variables (i.e., latent factors) and categorical latent variables (i.e., latent classes). FMM is known to face a variety of methodological challenges given its model complexity, and this study evaluates the potential of Bayesian estimation, particularly prior specifications, in addressing two challenges of FMM: classification accuracy and parameter recovery. We considered possible scenarios in applied research where subjective beliefs regarding class separation were incorporated into prior specifications such that subjective class separation might be greater or smaller than the true class separation in the population. Results of comprehensive Monte Carlo simulations showed adequate model performance using a moderately informative prior with subjective class separation greater than the true class separation. Practical implications for researchers are provided.
因子混合模型（FMM）在社会、行为和健康科学中越来越多地被采用，通过纳入连续潜变量（即潜因子）和分类潜变量（即潜类别）来识别群体异质性。鉴于其模型复杂性，FMM面临各种方法学挑战，本研究评估贝叶斯估计的潜力，特别是先验规范，以应对FMM的两个挑战：分类准确性和参数恢复。我们考虑了应用研究中的可能情况，即将关于类别分离的主观信念纳入先验规范，使得主观类别分离可能大于或小于总体中的真实类别分离。全面的蒙特卡罗模拟结果表明，使用主观类别分离大于真实类别分离的适度信息先验时，模型性能良好。为研究人员提供了实际应用建议。
