Therapist effects in randomised controlled trials: what to do about them

Authors

  • Stephen J Walters

    1. Author:Stephen J Walters, PhD, Professor of Medical Statistics and Clinical Trials, Medical Statistics Group, School of Health and Related Research, University of Sheffield, Sheffield, UK
    Search for more papers by this author

Stephen J Walters, Professor of Medical Statistics and Clinical Trials, Medical Statistics Group, School of Health and Related Research, University of Sheffield, Sheffield S1 4DA, UK. Telephone: +0114 2220730.
E-mail:s.j.walters@sheffield.ac.uk

Abstract

Aims and objectives.  The aim of this study is to describe and compare three statistical methods to allow for therapist effects in individually randomised controlled trials.

Background.  In an individually randomised controlled trial where the intervention is delivered by a health professional it seems likely that the effectiveness of the intervention, independent of any treatment effect, could depend on the skill of the health professional delivering it. This leads to a potential clustering of the outcomes for the patients being treated by the same health professional.

Design.  Retrospective statistical analysis of outcomes from four example randomised controlled trial datasets with potential clustering by health professional.

Methods.  Three methods to allow for clustering are described: cluster level analysis; random effects models and marginal models. These models were fitted to continuous outcome data from four example randomised controlled trial datasets with potential clustering by health professional.

Results.  The cluster level models produced the widest confidence intervals. Little difference was found between the estimates of the regression coefficients for the treatment effect and confidence intervals between the individual patient level models for the datasets. The conclusions reached for each dataset match those published in the original papers. The intracluster correlation coefficient ranged from <0·001–0·04 for the outcomes, which shows only minor levels of clustering within the datasets.

Conclusions.  The models, which use individual level data are to be preferred. Treatment coefficients from these models have different interpretations. The choice of model should depend on the scientific question being asked.

Relevance to clinical practice.  We recommend that researchers should be aware of any potential clustering, by health professional, in their randomised controlled trial and use appropriate methods to account for this clustering in the statistical analysis of the data.

Ancillary