The analysis of continuous outcomes in multi-centre trials with small centre sizes

The standard analysis of clinical trials stratified by centre is to include centres as fixed effects, but if many centres contribute small numbers of patients, this approach results in a loss of power. Assuming no treatment by centre interaction, we used simulation to examine power and coverage of confidence intervals from three approaches to the analysis of continuous outcome in multi-centre trials: ignoring centres, including centres as fixed effects, and including them as random effects. The simulation incorporated eight sizes of centre effects; randomization in blocks of size 2 or 4; and two sample sizes, namely 100 and 200 patients per treatment arm in a parallel groups design. All simulated data sets included many centres with few subjects. The three different approaches were unbiased and had similar coverage. Fixed effects analysis was less powerful, particularly when centre effects were small. Incorporating block randomization with larger block size increased non-orthogonality in the design, contributing to loss of power. Where centre effects are small and recruitment in many centres is low, the approaches of ignoring centres or incorporating them as random effects have better performance than the traditional fixed effects analysis.