Design and analysis of clinical trials with clustering effects due to treatment

Abstract

Where patients receive therapy as a group, there are good theoretical reasons to believe that variation in the outcome will be smaller for patients treated in the same group than for patients treated in different groups. Similarly, where different therapists treat different groups of patients, outcome for patients treated by the same therapist may differ less than outcome for patients treated by different therapists. Clinical trials evaluating such therapies need to consider this potential lack of independence. As with cluster-randomized trials, this has implications for the precision of treatment effects estimates and statistical power. There are nevertheless differences between clustering due to the organization of treatment and that due to randomization. In cluster-randomized trials the distribution of cluster sizes in each treatment arm should be similar as a consequence of randomization unless there is differential loss to follow-up. With clustering due to therapy group or therapist, cluster size may differ systematically between treatment arms, due to size of therapy groups or differing health professional caseload. Intra-cluster correlation may also differ between treatment arms. The implications of differential cluster size and intracluster correlation for design and analysis will be illustrated by data from two trials, the first comparing nurse practitioner care with general practitioner care, and the second comparing a group therapy with individual treatment as usual. The special case where a group therapy or therapist is compared with an unclustered treatment is examined in detail using a simulation study. The implications of differential clustering effects for sample size and power are addressed. It is argued that the design and analysis of this type of trial should take account of possible heterogeneity in cluster size and intracluster correlation.