The dissertation investigated how classification accuracy and parameter estimation of a nonparametric multilevel latent class analysis with covariates (hereafter referred as to conditional NP-MLCA) are affected by six study factors: 1) the quality of latent class indicators (i.e., conditional response probabilities: CRPs), 2) the number of latent class indicators, 3) level-1 covariate effects, 4) level-2 covariate effects, 5) the number of level-2 units, and 6) the group size of level-2 units. A total of 384 conditions were examined. Among the conditions and study factors explored in this dissertation, the results suggested four important implications. First, level-1 classification accuracy was acceptable when the sample size is 9,000, quality of indicators is 0.8, and the number of latent class indicators was 12. Second, the covariate estimates can be extremely biased when the quality of indicators was 0.6, especially when only six indicators were used. Third, for the level-2 covariate effect, larger samples size can compensate for the condition when the CRPs of indicators were 0.7, given the conditions explored in this dissertation. In addition, among the conditions where CRPs of indicators was 0.8, the biases of level-2 covariate effects were smaller when the number of groups was 150 than those when the number of groups was 50. Fourth, when the latent classes had indicators that showed consistent patterns, CRPs (0.6, 0.7, and 0.8) had the strong effect on CRP estimates in terms of biases and 95% CI coverage rate. When the latent classes had indicators that showed mixed patterns, the number of indicators (6 vs. 12) had the strongest effect in terms of biases and 95% CI coverage rates of CPR estimates of latent classes. An empirical study was included to illustrate the model.
Read
