Handling intra-cluster correlation when analyzing the effects of decision support on health care process measures

The clinical worksite constitutes a naturally clustered environment, posing challenges in the statistical analysis of quality improvement interventions such as computerized decision support. Ignoring clustering in the analysis may lead to biased effect estimates, underestimating the variance and hence type I errors. This paper presents a secondary analysis on data from a previously published, cluster randomized trial in cardiac rehabilitation. We compared six different statistical analysis methods (weighted and unweighted t-test; adjusted χ2 test; normal and multilevel logistic regression analysis; and generalized estimation equations). There were considerable differences in both point estimates and p-values derived by the methods, and differences were larger with increasing intracluster correlation.