Methods to address confounding and heterogeneity in cost-effectiveness analyses using real-world data

This thesis is concerned with improving methods for cost-effectiveness analyses (CEA). Real-World Data (RWD), for example, from routine data sources such as electronic health records, is used to generate comparative effectiveness and cost-effectiveness evidence in settings where appropriate evidence from Randomised Controlled Trials (RCTs) is not available. However, studies using RWD face fundamental issues pertaining to the study design, in particular around the risk of bias due to confounding and treatment effect heterogeneity. The aim of this thesis is to contribute to the literature on CEA methods for those settings. The thesis considers recent advancements in the causal inference and econometrics literature to examine the following objectives: (i) to identify challenges for comparative- and cost-effectiveness studies in applying the ‘target trial’ framework, (ii) to evaluate a novel local instrumental variable (LIV) approach in a CEA, (iii) to evaluate the performance of the LIV approach according to varying levels of instrument strength in a simulation study. The first paper in the thesis considers the main challenges in applying the target trial framework in comparative effectiveness and cost-effectiveness studies that use RWD, and offers recommendations, in particular around the interrelated issues of defining the study population and the comparator groups. The second paper is concerned with methods to address unmeasured confounding and heterogeneity, which are major challenges in CEA that use RWD. In this paper, I evaluate LIV methods in the context of a CEA that uses routine data from the ‘Emergency Surgery OR noT’ (ESORT) study. In the third paper, I extend this assessment of LIV methods with a simulation study that assesses the performance of LIV in realistic scenarios, defined by varying levels of instrument strength, and different forms of heterogeneity and sample sizes. The findings from these papers suggest that, in addressing both confounding and heterogeneity, LIV methods can provide accurate estimates of treatment effects of direct decision-making relevance. I find that, provided the instrument is strong, or the sample size is at least moderate, the LIV approach reports estimates with low bias and that are statistically efficient, regardless of the form of treatment effect heterogeneity that is present. The thesis concludes that by directly addressing confounding and heterogeneity the proposed methods can mitigate concerns about studies using RWD. Findings from this thesis can help future CEA that use RWD, to provide more useful evidence for decision-making.