Improving Small-Sample Inference in Group Randomized Trials and Other Sources of Correlated Binary Outcomes.

Estimating Equations;Estimation Performance;Generalized Estimating Equations;Test Size;Science;Biostatistics

Westgate, Philip MichaelSong, Peter Xuekun ;

University of Michigan

关键词: Estimating Equations;    Estimation Performance;    Generalized Estimating Equations;    Test Size;    Science;    Biostatistics;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/89710/pwestgat_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Group Randomized Trials (GRTs), along with many other types of studies, commonly can becomposed of a small to moderate number of independent clusters of correlated data. In thisdissertation, we focus on statistical inference in these settings. Particularly, we concentrateon test size and estimation variability when a marginal model is employed.Our first focus is in a general GRT setting in which a logistic regression only implementsan indicator of treatment assignment. A Wald test, using a model-based standard error, fora marginal treatment effect can tend to have a realized test size smaller than its nominalvalue. We therefore propose a pseudo-Wald statistic that consistently produces test sizes attheir nominal value, therefore increasing or maintaining power.Our second focus is on the estimation performance of QIF as compared to GEE when thenumber of clusters is not large, with a focus on GRT settings. GEE is commonly used for theanalysis of correlated data, while QIF is a newer method with the theoretical advantage ofbeing equally or more efficient. Therefore, it would be reasonable to believe that QIF shouldmaintain or increase power in GRTs, which typically have low power. We show, however,that QIF may not have this advantage in GRT settings, and estimates from QIF can have greater variability than estimates from GEE due to the empirical impact of imbalance incluster sizes and covariates, therefore concluding GEE is a more appropriate method inthese settings.We finally focus on improving the small-sample estimation performance of QIF. Specifically,we propose multiple alternative weighting matrices to use in QIF that combat itssmall-sample deficiencies. These weighting matrices are expected to perform better in small-sample settings, such as for GRTs, but maintain QIF;;s large-sample advantages. We comparethe performances of the proposed QIF modifications via simulations, which show they canimprove small-sample estimation. We also demonstrate that two of the proposed QIF versionswork best.