Designs and methods of explanatory (causal) analysis for randomised trials of complex interventions in mental health

The general public is exposed through the media to countless claims concerning the efficacy of counselling and various forms of psychotherapy for the treatment of depression and other mental health and personal problems. Health service providers are under increasing pressure to increase the availability of counselling and psychotherapy. There is clearly a need to design and implement controlled clinical trials to test whether these therapies work. It is equally clear that we need to develop and implement research projects that can tell us how these therapies work, what are the sources in the variability in responses to therapy and how these might be manipulated so that the therapies can be refined and improved. Equally, if a particular form of therapy does not appear to be very effective, we can use the same sources of information to develop an improved version that might be.

This is a proposal for the development, evaluation and dissemination of statistical and econometric methods for the explanatory analysis (causal inferences) of the joint effects of the following intermediate variables (mediators) in psychological treatment trials: programme participation (sessions of therapy attended); the quality of the therapy ? treatment fidelity and the strength of the therapeutic alliance ? including possible therapist effects; mediator variables arising from the treatment model (in the context of cognitive behaviour therapy for psychotic delusions, for example, these might include strength of the delusion, beliefs about the nature of the delusion and strategies for coping with the problem). A key feature of the project will be the design of trials for the evaluation of these explanatory pathways.


Technical challenges specifically associated with the analysis of causal pathways arise from the fact that virtually all of the intermediate variables listed above will be subject to: (a) selection effects ? patients with a good treatment-free prognosis are likely to be those who can develop a good relationship with a therapist and attend all of the planned sessions; (b) measurement error; and (c) missing data. The methods to cope with these challenges will be based on the use of instrumental variables (the key instrument being randomization and its interactions with baseline covariates), using maximum likelihood, two-stage least squares or structural mean models (G-estimation).