Discrete Feature Variables with Non-zero Slope Sample Clauses

Discrete Feature Variables with Non-zero Slope. The second set of simulations was performed using data from Section 3.4.2. Recall that, in this case, 10 distinct realizations of longitudinal data were generated for each of 500 subjects. Each subject had six measurement occasions and no missing data. The data was composed of 2 normal, 2 binary, and 2 Poisson feature variables, as well as a binary covariate. In addition, the second simulation incorporated a non-zero slope with respect to time for a subset of the feature variables. Within each latent class, the features were assumed to be independent and data for each feature was generated separately based on an autoregressive (AR1) correlation structure with a correlation coefficient of 0.3. All feature variables were generated with a standard deviation of 5. Table 3.5 provides a summary of the intercept and slope of each feature variable for each of two classes. As was the case in the simulations based on normal feature variables, two scenarios were considered. In the first, the probability of belonging to both classes was equal. In the second, the mixing proportion for belonging to class 1 was 0.80 and the mixing proportion for belonging to class 2 was 0.20. For each scenario, a one-, two-, and three-class finite mixture model was fit to the data. In order to avoid selecting a local solution, 100 random starting values were used for all two- and three-class models. Again, for a fixed number of latent classes, multiple roots were compared and the root that maximized the extended quasi-likelihood under the independence model was selected. Then, of these roots, the solution that minimized the information criteria being considered was selected as the best root and its corresponding number of classes noted. Table 4.5 provides a summary of the number of times that each of the proposed criteria selected a root with the correct number of classes when the mixing proportions were equal. The results suggest that CEQ-BIC performs slightly better than CEQC, EQIC, and BEQC; however, it still only correctly identifies a two-class solution in 60% of the runs. In this simulation, there do not seem to be any improvements in the performance of the information criteria when 5 × 2 cross-validation is used to estimate the extended quasi-likelihood function.