Using the Control Rate to Explore Heterogeneity of Treatment Effects in Clinical Trials

• Activity Number: 141
• Type: Topic Contributed
• Date/Time: Monday, August 12, 2002 : 2:00 PM to 3:50 PM
• Sponsor: Section on Health Policy Statistics*
• Author(s): Christopher Schmid*+ and Paul Stark
• Affiliation(s): New England Medical Center and Tufts New England Medical Center
• Keywords: Meta-analysis ; Multilevel models ; Bayesian inference ; MCMC ; Meta-regression

Abstract

Previously, we showed that weighted regression of treatment effects on event rates in the control group overestimated significance of the slope, compared with a more appropriate hierarchical model fit with an EM algorithm. Here, we compare the hierarchical model to two Bayesian versions, using normal and binomial error distributions fit by MCMC. We assembled 232 meta-analyses of binary outcomes published between 1990 and 1998 in twenty-one medical journals and in the Cochrane database. The control rate was significantly correlated with the log odds ratio in 13% of meta-analyses using the non-Bayesian model, but in only 6% using either Bayesian model. Treatment efficacy generally increased with higher control rates. Control rate effects were much more likely when treatment effects were heterogeneous and when control rates exhibited sufficient variation. The control rate effect did not vary with medical specialty, publication source, year of publication, significance of treatment effect, and number of studies included. The method seems generalizable, and for certain types of data, a useful tool for relating heterogeneity among clinical trials to underlying risk.