| Summary: | 碩士 === 國立臺灣大學 === 流行病學與預防醫學研究所 === 100 === From the aspects of ethic and cost-effective principles, many clinical trials and longitudinal social science studies usually involve in relatively small number of study subjects with a moderate to large number of observations per subject during follow-up. When the number of subjects or the cluster size is finite, the robust variance estimator for generalized estimating equations parameter estimates of regression models for marginal means proposed by Liang and Zeger (1986) exhibits considerable bias and may result in inflated type 1 error. Various modifications of the robust variance estimator for analysis of clustered data have been proposed in literature. However, little work has been done for longitudinal data.
In this paper, we adopt the existing robust variance estimators, proposed for analyzing the clustered data, in the analysis of longitudinal data. In our simulation study, not only exchangeable correlation structure but also a time-related correlation matrix such as first-order autoregressive structure are considered as the working correlation structures for the longitudinal data. Our numerical results suggest that when the number of subjects is larger than 10, the robust variance estimator proposed by Wang & Long (2011) for continuous responses, and estimators proposed by Kauermann & Carroll (2001) or Fay & Graubard (2001) for binary responses perform relatively well in terms of mean squared error and coverage rate of resulting t confidence interval. When the number of subjects is smaller than 10, we need to use jackknife or bootstrap estimators instead of robust variance estimators in order to infer more information about the population characteristics by taking resamples.
|
|---|
