지역 독립성 위배 조건에서 잠재계층 수 결정의 정확도 비교: 최대우도 접근과 베이지안 접근의 비교

초록

This study investigates the accuracy of latent class enumeration using maximum likelihood and Bayesian estimation approaches under local dependence in binary latent class models. Simulation data were generated by manipulating sample size, number of latent classes, and item correlation. AIC, BIC, and Sample-size adjusted BIC (SA-BIC) were used for model selection under the maximum likelihood framework, and posterior predictive p-value (PPP) under the Bayesian framework. The accuracy was assessed based on the proportion of correctly identified class numbers across 50 replications. Results indicate that while AIC performed poorly, BIC and SA-BIC showed declining accuracy with increasing item correlation. PPP showed lower overall accuracy but greater robustness under local dependence, with the highest accuracy when the class structure was more complex and correlation was high. These findings suggest that model selection criteria vary in robustness under local dependence and support Bayesian estimation as a viable alternative.

키워드