Learning more from complex psychological and social interventions in mental health

Abstract

University of ManchesterAbstract of thesis submitted by Clare FlachFor the degree of Doctor of Philosophy (PhD) Learning more from complex psychological and social interventions in mental health 2014Complex interventions by definition consist of many parts. Once it has been established that a complex intervention is effective the next step is to determine how it is effective and what the active ingredients are. In a randomised controlled trial the causal effect of the intervention can be determined simply but when the mechanism of the intervention is investigated the exposure is no longer randomised. Instrumental variable methods have been developed to overcome problems of unmeasured confounding which result from the loss of randomisation but they require additional assumptions. When applying instrumental variable techniques in real data sets with finite samples the identification of effective instruments and the use of weak instruments can cause bias in estimation. This thesis compares methods for the identification of instruments and estimation in the presence of many weak instruments. Shrinkage techniques, the LASSO (Least Absolute Shrinkage and Selection Operator) and Elastic Net, utilised in data mining are applied to the context of instrumental variable selection and compared to a single instrument, all instruments and backward stepwise selection. The commonly used two stage least squares estimator is compared to the limited information maximum likelihood estimator with and without Fuller's adjustment in the presence of many weak instruments. The selection and estimation methods are compared in simulated data replicating the design of a clinical trial of a complex intervention with varying levels of instrument strength and number. The simulation results indicate that when there are multiple true instruments using multiple instruments is preferred to a single instrument. The benefit of multiple instruments increases as the individual instruments become weaker. Selection by the LASSO increases the first stage F-statistic and reduces bias but precision can suffer in the more parsimonious models. Estimation by two-stage least squares is preferred over limited information maximum likelihood in terms of the mean-squared error in the presence of many weak instruments but the maximum likelihood estimators perform better in terms of median bias. When the process variable is categorical the two-stage least squares is preferred in terms of bias and precision.The statistical methods identified to be effective in the simulated data are applied to clinical trial datasets to answer substantive questions regarding the important components of cognitive behavioural therapy and to determine if the therapy works through the expected processes. The results indicate that formulation is a key component of CBT therapy for the prevention of psychosis and suggests that homework and active change strategies are also important. However due to the high correlation between these factors it is not possible to distinguish the importance of one aspect over another.  

Date of Award	1 Aug 2014
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Graham Dunn (Supervisor) & Anthony Morrison (Supervisor)