A causal inference framework for comparative effectiveness and safety research using observational data, with application in type 2 diabetes

Abstract

Randomized controlled trials are the gold standard to answer causal questions in health research as the process of randomization ensures balanced treatment groups and therefore makes it possible to compare average group outcomes directly. But they have many limitations with respect to costs, ethical considerations and practicability and therefore may not be suitable to answer all research questions. Evidence on cause and effect relationships from observational studies have the potential to overcome the limitations of trials and close important research gaps as they provide the possibility to study subpopulations of patients which are often excluded due to safety concerns, or can give insights into the risk profile of long-term endpoints. The quality of this real-world evidence depends on the quality of data, their suitability to answer a particular research question and the use of appropriate methods to estimate the treatment effect of interest. Of concern in observational research is bias in the treatment effect estimation due to confounding, as the treatment assignment is not controlled by the researcher and cannot be randomized. It is therefore possible that treatment groups are not balanced and confounding factors exist in the data which influence the treatment choice and the outcome of interest simultaneously. The benefits of observational studies make them attractive for studying the risk and benefit profiles of oral type 2 diabetes treatments, especially of newer agent classes such as Sodium-glucose Cotransporter-2 Inhibitors. Prescribing of this treatment class has increased in recent years and a large proportion of type 2 diabetes patients have become eligible to receive agents from this class after recent treatment guideline changes. More information about treatment effects of Sodium-glucose Cotransporter-2 Inhibitors are needed especially for the large patient population of older adults (e.g. 70 years or older), as possible adverse effects such as osmotic symptoms associated with this class could have severe consequences for these patients. The overall aim of this thesis is to develop a causal inference framework for the exploitation of observational data, needed to derive high quality evidence on the benefit and safety profile of oral type 2 diabetes treatments, with a focus on the widely prescribed treatment class of Sodium-glucose Cotransporter-2 Inhibitors and the patient population of older adults. Chapter 1 and 2 are introductions to causal inference theory including the description of all estimation methods employed in this thesis and an introduction to type 2 diabetes research encompassing important treatment decision considerations, and current research evidence on Sodium-glucose Cotransporter-2 Inhibitors. Chapter 3 presents a triangulation framework of assorted estimation methods to establish the consistency of estimation results from approaches utilizing different parts of the data and relying on different data structure assumptions. Furthermore, an Instrumental Variable approach is introduced which uses data from the period before treatment initiation to mitigate potential bias in case the exchangeability assumption is violated and a history of the outcome of interest previous to treatment initiation has an influence on the treatment decision. Chapter 4 describes a simulation study on the performance of established construction methods for a proxy Instrumental Variable of health care provider prescription preference. The methods are tested under different data conditions such as change in provider preference over time, missing data in baseline covariates and different sample sizes within each health care provider. Additionally, a construction method is introduced that aims to address changes in preference over time and non-ignorabile missingness in baseline characteristics. In Chapter 5 the developed conclusions about a robust Instrumental Variable estimation approach from previous chapters are applied for a causal analysis on the relative benefit and risk profile of Sodium-glucose Cotransporter-2 Inhibitors versus Dipeptidyl peptidase-4 Inhibitors in the patient population of older adults. Chapter 6 provides an overview of the main findings and implications of this thesis and discusses limitations and future research potential of each study.