One approach to analysing multi-centre trials of drug treatments works from the premise that it is not plausible for a treatment effect to vary across centres. If it does, it is deemed a fault with the study design. If a significant centre-treatment interaction is not detected, then the design is assumed to be sound and global inference is made from a model not allowing for any variation in treatment effects between centres. However, a drug effect can sometimes vary owing to differences in the centre populations even when they are defined within the constraints of the protocol. For example, one drug may work better on severely ill patients than another but less well on moderately ill patients. Thus, a centre containing more severely ill patients could produce larger treatment effects than centres containing a more even mixture of patients. For this reason it is our belief that centre-treatment effects are always plausible and that if global inference is required from a multi-centre trial, then the random effects model is likely to be the most appropriate.
Interactions are often even more plausible in trials not involving drugs. For example, in a trial of surgical techniques, one centre may have much more expertise with one technique than another. In this type of trial a random effects model should almost always be used in order to provide global inference.
Was this article helpful?