台大心理系

In latent class analysis simulations assessing performance of parameter estimation, the estimated classes in every replication need to be mapped to the true classes set in the process of data generation. Three procedures for class mapping were compared in this study, including the starting value method (Collins & Lanza, 2010), the membership assignment method (Tueller et al., 2011), and the class similarity method we proposed. A simulation that manipulated class proportions, conditional probabilities, number of variables, and sample size was conducted to evaluate the likelihood that the classes were successfully mapped and the mean squared error of the estimated conditional probabilities for each method. The preliminary results indicate that the membership assignment and the class similarity methods yielded similar performance and outperformed the starting value procedure. Future studies may extend the simulation design to cover a wider range of conditions so as to enhance the generalizability of the findings.