A simulation study is conducted for estimating common odds ratio in multiple 2x2 tables when the data are correlated within clusters. The simulation design imitates scenarios common in the clinical trial setting, with small mean cluster size (m=5), moderate number of cluster (N=10, 20, 30), modest levels of intracluster correlation (ICC=0.0, 0.1, 0.3, 0.5), and two patterns of risk for subjects in the 1st row and kth stratum (P(1k)=narrow, wide band). The cluster size within each table is modeled by the negative binomial distribution truncated below 1 with imbalance parameter, k=0.6, 0.8, and 1.0. The correlated binary response is generated from a beta-binomial distribution. In this paper, we compared unadjusted log Mantel-Haenszel estimator and adjusted estimators based on Rao and Scott, Donald and Donner and Begg as well as common odds ratio based on GEE models, from which Begg’s correction factor is derived. We evaluated these estimators as the number of tables remains fixed but the number of clusters becomes large. All estimators perform well in terms of bias and observed variances; although GEE estimator has slightly larger bias as than other estimators. In terms of coverage proportion and variance estimation, Begg, Rao and Scott, Donald and Donner, and GEE estimator perform much better than Mantel-Haenszel estimator; among them, Begg and GEE estimator perform better than Rao Scott and Donald and Donner estimator. Based on the results of the study, we recommend the use of Begg’s new approach.
Learning Objectives: 1. Evaluate several existing adjusted Mantel-Haenszel estimators in a clinical trial setting. 2. Demonstrate the effects of cluster size, number of cluster and intracluster correlations on the type I error and power
Keywords: Biostatistics, Clinical Trails
Presenting author's disclosure statement:
Organization/institution whose products or services will be discussed: None
I do not have any significant financial interest/arrangement or affiliation with any organization/institution whose products or services are being discussed in this session.