Bayesian Statistics Seminar
North Carolina State University
presents
Liansheng Zhu
North Carolina State University
"Analyzing Longitudinal Discrete Data with Nonignorable Missing"
ABSTRACT
It is not unusual to have missing responses in longitudinal studies. When the missingness is not random, analysis of the data can be very challenging and satisfactory estimates and valid inference may not be available. Pattern-mixture models are often used to model the joint distribution of observed and missing responses to reduce the bias in estimates. However, marginalization over the missing patterns makes the inference difficult to interpret the marginal parameters. Methods of marginalization over dropout patterns are not available for discrete data, especially when there are several important covariates and interpretation of their estimates is desired. In avoiding marginalization, a sensitivity analysis with the multiple imputation regarding various missing mechanisms is introduced for a special study with binary responses. It may provide a nice tool to explore robustness of results. Further, in the light of random pattern-mixture models proposed by Wang et al. (2004), we present a pseudo-imputation method to obtain estimates of parameters at the marginal level that can be easily implemented by most statistical software. The proposed method seems to provide acceptable parameter estimates.
Tuesday, September, 27, 2005
4:00 - 5:00 pm
208 Patterson Hall