This study investigate effect of utilizing hierarchical linear growth model for missing at random data (MAR) and not missing at random data (NMAR) with common missing data approach. This paper introduced the traditional missing data approach and their advantage and disadvantage. The type of missing data with their special assumption was also identified and interpreted. The major approach this study used is modeling the NMAR as a discrete time survival process and single indicator model. The simulation study investigates the bias and parameter coverage, and an empirical data was analyzed to explain the influence of risk factor and disadvantage minority group. The result showed that the difference between the MAR and NMAR was small. The incorrect assumption for NMAR resulted in the great bias for variance and covariance. The more proportion of missing data and stronger dependence leaded the coverage less and serious bias.
comments:
1) We can observe that the statistic fit indices didn’t be involved the bias and coverage when dependence and small missing rate. When we analyze the empirical data, the correct model which is unknown can’t be found out by fit indices. This paper investigate the influence but the solution still not be found. In simulation, it is still expected that correct model corresponding with generated model will perform well.
2) In empirical study, MAR and NMAR have similar parameter estimate. The author recommended NMAR model rather than exploratory fashion. I agree this view because we always need to do theoretical assumption to model our data.