Abstract

Objectives: The objective of this study is to compare the performance of generalized estimating equations (GEE) for analysis of clustered binary observations under varying group and observation numbers according to intraclass correlation coefficients (ICC).

Materials and Methods: The comparison of GEE performance was made by using bias of parameter estimations through computer simulations under varying group sizes, ICC, number of clusters, and number of observations per cluster. Simulations were performed in SAS 9.0 by using Monte Carlo simulation method. Analyses were made with SAS GENMOD procedure.

Results: When intraclass correlation coefficient was low (ICC<0.10), there was no significant difference in parameter estimation and their biases and it was observed that GEE gave reliable, consistent and unbiased estimations. However, when ICC increased (ICC>0.10), it was found that the parameter estimations were significantly biased. On the condition that total sample size is fixed; it was observed that, even though the general sample size was constant in all groups while the number of groups was decreasing, when the number of observations per cluster increased, parameter estimations and their biases weren't affected significantly and the effective factor in parameter estimation was ICC.

Conclusion: Because GEE method uses population averaged logistic regression approach, it cannot explain the changes and correlations in clusters completely. The use of GEE method is inconvenient particularly for data sets which have ICC greater than 0.10.