Investigating methods for inferring causality from observational data: an application to longitudinal cohort data

In health services research, it is vital to know whether an event, such as treatment or modifiable exposure, has a causal effect on the outcome in order to deliver the best treatments and health policy interventions to patients and the public. Due to limited evidence available from randomised controlled trials, policy makers are becoming increasingly reliant on supplementary information from observational data to evaluate the effectiveness of treatments and interventions. Several different methods have been used to derive causal estimates from observational studies including regression models, propensity score adjustment and instrumental variables. Each method requires different assumptions to be satisfied and it is often unclear which methods are the most appropriate for any given situation. Instrumental variable (IV) methods can, under certain assumptions, estimate causal effects even in the presence of unmeasured confounding. IV methods have been well developed for additive no-interaction models for continuous outcomes. The application of IV methods to binary and time-to-event outcomes is more problematic due to non-collapsibility of the relevant effect measures and there has been limited development of IV methods for time-to-event outcomes until recently. Over the last few years, there has been an increase in the availability of electronic health record data such as UK Biobank and Clinical Practice Research Datalink (CPRD). These datasets will provide a huge resource of observational data. Therefore, guidance on the appropriate methods for analysing observational data is required for researchers to obtain reliable results from their analyses. The aim of this thesis is to obtain reliable treatment effect estimates from observational data. The focus is on time-to-event outcomes since causal methods have been less well developed in this area until recently. The relative performance of different causal methods will be assessed and the methods applied to an observational cohort of patients from CPRD.