Improving causal inference in quasi-experiments : application of common methodologies as instrumental variable analyses and the use of causal diagrams

Abstract

Increasing recognition of the difficulty in eliminating bias from traditional observational studies, and challenges in examining important population-wide exposures, such as dietary exposures and policy interventions, through randomized controlled trials (RCT), has led to growing interest in quasi-experiments in epidemiology.

Quasi-experiments have traditionally been viewed as a separate series of study designs. However, there is recent recognition that major quasi-experimental designs can be viewed as an instrumental variable (IV) designs. Mendelian randomization (MR), a recognized specialized application of IV, has become an important part of the epidemiologist’s toolkit. The use of directed acyclic graphs (DAGs) and associated tools has played a key role in identifying sources of bias in MR studies.

This thesis explored whether conceptualizing a particular type of quasi-experiment, interrupted time series (ITS), as an application of an IV design in conjunction with the use of DAGs, could assist with identification strategies and causal inference. Quasi-experimental designs were used to examine causal questions not easily answered through conventional observational study designs or RCTs, i.e., whether a particular dietary exposure affects cardiometabolic outcomes, and whether a population-wide economic policy has health impacts. Firstly, MR was used to demonstrate a recognized specialized application of IV. Next, an IV approach, with the use of DAGs was applied to ITS studies. Specifically, MR studies were conducted to examine the impact of selenium on ischemic heart disease (IHD), type 2 diabetes (T2D), and their risk factors, and ITS studies were conducted to examine the impact of the introduction of the minimum wage in Hong Kong on suicide rates and birthweight.

The MR studies found weak evidence of an inverse association of genetically predicted selenium with IHD, and strong evidence of a positive association with T2D, although results were not robust to all sensitivity analyses. There was strong evidence of an inverse association with LDL cholesterol, and evidence of a positive association with fasting insulin and indicators of insulin resistance, but little or no evidence of an association with fasting glucose or indicators of beta-cell function. The ITS studies found evidence that the introduction of the minimum wage was associated with a decrease in suicide rates, with the biggest impact in working aged men. There was little evidence of an association with birthweight.

These studies highlight the importance of examining the health impacts of population-wide exposures that are difficult to examine through traditional study designs, in turn underlining the importance of robust quasi-experimental methods. Conceptualizing ITS studies as applications of IV facilitated illustration of the assumptions, strengths and potential threats to validity using DAGs, identification of potential threats to validity as violations of the exclusion restriction assumption, rather than as confounders consistent with the modern definition of confounding and helped to place ITS within a unifying framework. The studies demonstrate how a particular type of quasi-experiment can be conceptualized through an IV analysis and represented in a DAG and provide an example for development of similar conceptualizations for other quasi-experimental designs into the future.