Exchangeable cluster binary data correlation coefficient estimation with generalized estimating equations

This study used simulated data to evaluate the performance of distinct conditional generalized estimating equations (CGEE) for the analysis of exchangeable correlation for binary data. The CGEE differs from the usual generalized estimating equations (GEE) in that, instead of marginal expectations, the conditional expectations of the responses were used in the estimating equations. The major distinction among the CGEEs compared was the sizes of the conditioning events used in the conditional expectations. The results show that, for the estimation of correlation coefficient, the bias decreases, and the variance increases when more members in a cluster are included in the conditioning event. The increase of variance is, however, only moderate for small intracluster correlation coefficient. On the other hand, for the estimation of regression parameters, the bias and variance of the estimates both increase when the size of the conditioning event increases. The increase, however, is also insignificant when the correlation coefficient is small.