說明 
99 p 
附註 
Source: Dissertation Abstracts International, Volume: 6003, Section: B, page: 1152 

Adviser: Robert Floden 

Thesis (Ph.D.)Michigan State University, 1998 

The finite sample performance of the independence estimating equations (IEE), which assumes that the responses are independent, and generalized estimating equations (GEE), which take the correlation into account to increase efficiency, methods of estimation proposed by Liang and Zeger (1986), are evaluated when used to fit grouped logistic models. Specifically, the performance of the GEE estimator, under a constant working correlation matrix across clusters, is compared to the IEE estimators in terms, of their bias, their efficiency, and their accuracy in testing a simple hypothesis about the mean parameters 

A Monte Carlo simulation is used to compare the estimators empirically under the various conditions specified in this study. For example, the effects of the population average pairwise correlation of the binary responses and the effect of the distribution of a covariate on the IEE and the GEE estimator associated with that covariate are analyzed. In this regard, a data generating mechanism (DGM) to simulate binary correlated responses is proposed. This DGM has the advantage over the traditional random effects logistic formulation (TRELM) in that the marginal probabilities, conditional only on the covariates, have a simple logistic form and the mean parameters quantify the average effect of the covariates on the population's response 

The finite sample findings show that for high correlated binary responses, when including either a binary or a standard normal within cluster covariate, the GEE outperforms the IEE in estimating the mean parameter associated with that covariate in terms of their efficiency and their accuracy in hypothesis testing, and both the IEE and GEE estimators are equally biased. However, for low correlated binary responses, both methods perform equally well 

In contrast, for highly correlated binary responses, when including a binary cluster specific covariate, the IEE estimator substantially outperforms the GEE in estimating the mean parameter associated with that covariate in terms of their bias, their efficiency, and their accuracy in hypothesis testing. However, for low correlated binary responses the IEE method outperforms the GEE only in terms of their efficiency, although the differences are not as substantial 

School code: 0128 

DDC 
Host Item 
Dissertation Abstracts International 6003B

主題 
Statistics


0463

Alt Author 
Michigan State University

