Analysis of Clustered Binary Responses Using GEE Estimation with Bias Corrected Covariance Estimator

• Activity Number: 226
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Sadia Mahmud*+ and Naseem K. Afridi and Juanita Hatcher and Debra Nanan
• Companies: Aga Khan University and Aga Kahn University and Aga Khan University and Pacific Health & Development Sciences
• Keywords: binary ; correlated ; bias corrected

Abstract

The method of Generalized Estimating Equations (GEE) is a regression approach for correlated responses. GEE uses a robust estimator to estimate the covariance matrix of the regression coefficients. However, the robust estimator may be biased when the number of clusters is small. Mancl and DeRouen (2001) proposed a bias-corrected robust covariance estimator for regression coefficients. A cross-sectional study was conducted in 2002 in Peshawar, Pakistan, to assess the tetanus toxoid vaccination status among women of reproductive age. The sampling technique employed was stratified two-stage cluster sampling. A total of 40 clusters were selected. We analyzed the clustered binary responses using GEE estimation with bias-corrected covariance estimator, implemented by a GEE SAS macro written by Mancl. The working correlation structure specified was "exchangeable." The multiple logistic regression model contained seven covariates. For all the regression coefficients the bias corrected standard errors were larger than the robust standard errors, though the former and the corresponding p values were not markedly different. For only two covariates the bias corrected estimator (in conjunction with using t-distribution to compute confidence intervals for odds ratio) led to a different inference as compared to that from the robust estimator.