Abstract

Abstract: In trials of non-pharmacological treatments clustering of patients may occur due to randomisation or due to the nesting of patients in therapist or therapy groups. This talk will consider the implication of heterogeneity of the clustering effect where the outcome measure is binary.

Heterogeneity of the clustering effect may arise in trials of group administered or therapist treatment where a clustered therapy is compared with a placebo or pharmacological control. This leads to what has been called a partially nested design where there is clustering in one arm but not the comparator.

Heteroscedasticity of the clustering effect can arise in cluster randomised trials. Statistical analysis of cluster randomised trials generally assumes that the intra-cluster correlation coefficient (ICC) is the same in all arms. This assumption can be justified by randomisation provided the clustering effect is due to the baseline characteristics. If, instead, the magnitude of the clustering effect is caused by the intervention, the ICC may differ between trial arms.

After reviewing the situation where the outcome measure is continuous, the robustness of methods of analysis for binary outcomes will be considered.