Abstract
BackgroundProcess evaluation (PE) focuses on assessing the causal relationships between interventions, outcomes of interests and the processes that bring about changes to the outcomes of interest. Since causal relationships are of interest in such a study, methods of causal inference should generally be applied to PE. This thesis focuses on the development of methods of causal inference to overcome existing barriers to the application of these methods in a PE with a focus on PE in mental health related trials. The focus on mental health related trials is because many of the interventions used in mental health related services affect the outcome via a series of processes and so require the use of PE to investigate the causal relationships between the intervention, processes and outcome of interest. The methodological gaps addressed in this thesis were chosen following a systematic review on methods of causal inference used within PE in mental health related trials. The gaps in methods of causal inference centre around the estimation of sequentially mediated causal effects and the estimation of the causal odds ratio (OR) for binary outcomes.
Overview and main findings
This thesis is broadly divided into four main parts. The first part describes a systematic review of existing methods of causal inference in use for PE and identified existing methodology gaps in the use of causal inference within PE. The review confirmed the existence of two methodological gaps: estimation of sequentially mediated causal effects with two or more mediators, and estimation of causal OR under scenarios where a continuous mediator is nested within a binary outcome.
The second part of this thesis reviewed the existing methods used to address these problems. From this review the potential outcome (PO) framework was adopted to define causal effects of interest within this thesis. This definition would subsequently guide the direction for the development of the methods. Additionally from this review, the estimation method developed by Imai et al. (2010) was chosen to propose novel methodological solutions in this thesis.
The third part of the thesis focused on the development of the methods, focusing on how the foundation of the methods were formed from prior work in causal inference and adaptations to existing methods. A set of procedures was simultaneously developed to validate the new methods. This was done to provide confirmation that the novel methods, as implemented in this thesis, provided results in line with what was expected. The procedures largely verified that these were within expectations. Additionally, sensitivity analyses approaches were proposed for the newly developed methods. To estimate the causal effects using the novel methods, certain assumptions were made and the sensitivity analyses served to assess the impact on the causal effect estimates should the assumptions be violated. A new R programme was developed as an implementation of the novel estimation methods as well as the sensitivity analyses and allows these methods to be easily accessible.
The fourth and last part of the thesis consisted of an application of the methods using data from a real randomised controlled trial (RCT), the Carers’ Assessment, Skills and Information Sharing (CASIS) trial. The application demonstrated that the novel methods had real world utility and enabled the testing of hypotheses which previously could not be tested.
Conclusions
This thesis uniquely enables the estimation of sequentially mediated causal effects and the estimation of causal OR under scenarios that previously had no existing solutions. Additionally, sensitivity analyses were developed for the newly developed estimators to enable an assessment of assumptions used in the estimation of the causal effects. An important limitation in the current implementation of the estimators is the inability to include any interaction terms. This can be rectified in a future modification of the developed estimators. An important strength of this thesis lies in the rigorous procedures used to demonstrate the correctness of the implementation of the estimators. Also, a demonstration of the newly developed estimators on a real trial demonstrated their capabilities in testing hypotheses relating to sequentially mediated causal effects and causal OR within the conduct of PE in an RCT.
| Date of Award | 1 Dec 2022 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Sabine Landau (Supervisor), Kimberley Goldsmith (Supervisor) & Richard Emsley (Supervisor) |