Analyzing the Effect of Treatment Actually Received in Randomized Clinical Trials

In randomized clinical trials, we are often faced with variation in the level (or 'dose') of treatment actually received among patients assigned to experimental treatment. While the intent-to-treat comparison provides an appropriate primary analysis of treatment effect, there is often interest in using information about treatment received to study aspects of the dose-response relationship. Due to the non-random nature of treatment actually received, na�ïve approaches such as an 'as-treated' analysis will generally produce biased results. In this talk, we discuss 'structural model' approaches to this problem. Structural models, which describe treatment effects using potential (latent) outcomes, provide a clear and intuitive approach to causal inference. The assumptions, strengths and weaknesses of these methods and possible extensions will be discussed. Practical issues will be illustrated with data from a clinical trial examining the effect of skin-to-skin holding on premature infants.