The effect of non-response data on model selection in latent class analysis: an application of missing and skip response data

• Main Authors: Pei-Ying Lin, 林珮瑩
• Other Authors: TING-HSIANG LIN
• Format: Others
• Language: en_US
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/59095725789280361776

碩士 === 國立臺北大學 === 統計學系 === 93 === Latent class models (LCM) has found important applications in the social and behavioral sciences for modeling responses to sets of categorical variables and non-response is typical when collecting data. In this study, the non-response mainly included “skip pattern” and “missing data”. The primary purpose of this research is to evaluate the effects of several factors (latent class proportions, conditional probabilities, missing data rates, skip pattern rates, sample sizes and number of items) for non-response data across LCM. Some researches have studied latent class with missing items. However, there is no research on the latent class model selection when both skip and missing items occur simultaneously. Therefore, we are interested in investigating in the effect of non-response data on model selection in the latent class analysis. We simulated these non-response data and used the eight information criteria to evaluate the accuracy rates for selecting the best model. The results showed the main factors in model selections of these information criteria are latent class proportions, conditional probabilities, number of sample sizes and items. In addition, when we contrasted with comparison data, the levels of skip and missing rates are also significant. We recommend to use AIC, BIC* and CAIC* when sample size is small and the number of items is four and under more extreme different conditional probabilities. But if item is increased to eight and sample size is still small, only BIC* is not suitable and the other seven indices perform better. AIC and HT-AIC don not improve with large sample sizes and more items situation. With fewer samples and items, BIC and CAIC have very low accuracy rates. The adjusted BIC (BIC*) and adjust CAIC (CAIC*) have noteworthy improvements for selecting correct model.