Estimating causal effects of complex interventions in longitudinal studies with intermediate variables
Lead Research Organisation:
University of Manchester
Department Name: Medical and Human Sciences
Abstract
It is often said we live in an information age, and this includes an increased public awareness of medical issues and knowledge about available treatments. What is often less understood, even by medical professionals, is exactly why some patients respond better to a treatment than others. For example, is it the number of sessions of a particular psychotherapy which is important, or is it the strength of the relationship between the therapist and the clinician? This proposal aims to develop new statistical methods which can be used to answer these types of questions and assess the underlying causal mechanisms of a treatment, using data from recent trials in mental health areas such as schizophrenia and autism as examples.
The proposed research will benefit those suffering from these conditions by providing more clues about how and why their treatments work. Additionally, although the current applications are to mental health areas, the statistical methodology will be applicable to a wide range of medical specialties, enabling studies in all areas to be analysed using these methods, and to better design future studies with these types of causal questions in mind.
The proposed research will benefit those suffering from these conditions by providing more clues about how and why their treatments work. Additionally, although the current applications are to mental health areas, the statistical methodology will be applicable to a wide range of medical specialties, enabling studies in all areas to be analysed using these methods, and to better design future studies with these types of causal questions in mind.
Technical Summary
Valid causal inference remains the key goal motivating the analysis of data from observational studies and randomised trials. Identifying the causal mechanisms is crucial to understand the epidemiology of disease and for developing successful treatments. The motivation underpinning this proposal is the analysis of randomised trials of complex interventions where the estimation of the effects of treatment may be influenced by the presence of indirect effects acting through intermediate variables such as mediators or surrogates; for example, the effect of CBT in a psychotherapy trial may be mediated by the therapeutic alliance between therapist and patient.
Traditional methods of statistical analysis make very strong assumptions requiring the absence (through lack of acknowledgement) of unmeasured confounders between the intermediate variable and outcome. Recent advances recognizing this problem have been proposed that make use of instrumental variables (IVs), where the key instruments are randomisation and its interaction with baseline covariates. An alternative method, principal stratification (PS), stratifies the population into latent classes based on the potential values of the intermediate variables at each level of treatment. Both methods rely on having good predictors of the intermediate variables from baseline covariates. However, there are still further problems; for example, there is likely to be treatment effect heterogeneity, and if a patient‘s idiosyncratic response to treatment is related to their decision to select treatment then the standard IV approach may not estimate a valid casual effect.
This proposal aims to extend the IV and PS methods by addressing this and other complications such as measurement error, and developing the methods to allow for multiple intermediate variables. Additionally, most studies are longitudinal in design and so a particular focus is on extensions to take account of repeated measures and missing data on the outcomes and intermediate variables. The new methods will be applicable to observational studies and randomised trials in all areas, since questions about the causal mechanisms are universal.
Exploratory analysis of randomised trials investigating complex interventions from the field of mental health will be used to illustrate the new techniques, and the empirical results will provide greater understanding of how these particular treatments work. The identification of predictors of intermediate variables such as surrogate outcomes or therapeutic alliance is also of benefit. Wider clinical opportunities will derive from application of the methods to other studies, which will allow a greater insight into the nature of many existing and future treatments.
Traditional methods of statistical analysis make very strong assumptions requiring the absence (through lack of acknowledgement) of unmeasured confounders between the intermediate variable and outcome. Recent advances recognizing this problem have been proposed that make use of instrumental variables (IVs), where the key instruments are randomisation and its interaction with baseline covariates. An alternative method, principal stratification (PS), stratifies the population into latent classes based on the potential values of the intermediate variables at each level of treatment. Both methods rely on having good predictors of the intermediate variables from baseline covariates. However, there are still further problems; for example, there is likely to be treatment effect heterogeneity, and if a patient‘s idiosyncratic response to treatment is related to their decision to select treatment then the standard IV approach may not estimate a valid casual effect.
This proposal aims to extend the IV and PS methods by addressing this and other complications such as measurement error, and developing the methods to allow for multiple intermediate variables. Additionally, most studies are longitudinal in design and so a particular focus is on extensions to take account of repeated measures and missing data on the outcomes and intermediate variables. The new methods will be applicable to observational studies and randomised trials in all areas, since questions about the causal mechanisms are universal.
Exploratory analysis of randomised trials investigating complex interventions from the field of mental health will be used to illustrate the new techniques, and the empirical results will provide greater understanding of how these particular treatments work. The identification of predictors of intermediate variables such as surrogate outcomes or therapeutic alliance is also of benefit. Wider clinical opportunities will derive from application of the methods to other studies, which will allow a greater insight into the nature of many existing and future treatments.