This study investigated how several factors could affect the estimation quality of DINA model. Factors included:
a) structure of the latent classes, 3 kinds --- unstructured, hierarchical, higher-order.
b) type of prior, 2 kinds --- unstructured, hierarchical.
c) level of guessing and slipping parameter, 2 levels --- high (.20--.30), low (.05-- .15).
d) sample size, 3 levels --- 1,000, 2,000, 3,000.
e) estimation method, 2 kinds --- fully Bayes, empirical Bayes.
Results showed that:
1) unstructured prior and empirical Bayes combination provides most robust results.
2) a relatively small sample size (1000) suffices to accurately estimate the DINA model.
3) low level of guessing and slipping resulted in more accurate and precise estimates.
In this study, two important factors for DINA estimation were fixed, one is Q-matrix and the other is test length. Additionally, Q-matrix was well defined then each attribute could be tested by appropriate number of items, and the items measuring an attribute had the same composition. It is well known that, Q-matrix plays an essential part in DINA model estimation, as the Q-matrix has already defined one half of the deterministic part of the DINA model. So the results for this study is reasonable, given the correct specification of Q-matrix, other factors would not obviously affect estimation accuracy. However, some methods were preferred to others under some conditions. It would be more interesting to investigate the interaction between Q-matrix and the other factors.
Second, impacts of these factors were studied when the true model was used. What would happen if the model is mis-specified?
Third, what about using some other CDM models? Like compensatory models, or attribute-level specified models.