Exploiting instrumental variables to estimate the effects of time-varying treatments using routine data

Evidence from clinical trials plays a central role in the evaluation of the benefits and harms of treatments, but for many of them, trial-based evidence is unavailable or insufficient. Government agencies are increasingly using patient data collected routinely in hospitals, general practices, and disease registers, for complementing evidence from trials. However, studies that use routine data are prone to biases due to both observed and unobserved differences between patient groups receiving different treatments. For example, patients in different treatment groups often differ according to important prognostic variables, which may be measured or unmeasured, that affect both the treatment patients receive and how well they respond to treatment. Policy makers are therefore worried that these potential biases in the use of routine data may lead to wrong treatment decisions and poor allocation of healthcare resources.

This research addresses these concerns by extending an approach, known as instrumental variables (IV) analysis, widely used in social sciences for addressing both observed and hidden biases. IV analysis essentially involves using variables (known as 'instruments') that affect the treatment patient receives but have no effect on health outcomes, except through the treatment itself. For example, IV analysis often uses genetic factors for estimating the effect of risk factors on disease; genes are associated with risk factors but only affect health outcomes through the risk factor because they are assigned at random at birth. IV methods can therefore play an important role in obtaining robust evidence from routine data, but the application and usefulness of IV methods in these studies remains poorly understood. In particular, policy makers are often interested in evaluating the effects of treatments sustained over time, which requires controlling for both observed and unobserved differences between treatment groups, at different points in time. Existing IV methods are not appropriate for addressing this type of biases that vary over time.

In addressing these challenges, the objectives of this research are:
i) to assess the validity and usefulness of IV analysis in studies that use routinely collected data. In particular, the research will demonstrate how to assess the plausibility of potential instruments to control for the biases over time.
ii) to address analytical issues in the implementation of IV methods in this context. This will include addressing issues related to the quality of the instruments, for example, how well these variables predict the treatment received.
iii) to illustrate the flexibility and usefulness of the proposed IV methods across different clinical settings; these will include studies evaluating biological treatments for rheumatoid arthritis, intensive glycaemic control for type 2 diabetes, and second-line treatments for chronic heart failure.

The findings of this research will be disseminated beyond the academic community, for example, by delivering seminars and practical workshops to help applied researchers, clinical experts and policy analysts understand how derive robust estimates of the effects of treatment strategies sustained over time using routinely collected data. This research will also prioritise dissemination activities to those directly involved in the analysis and interpretation of evidence from routine data sources to inform resource allocation decisions, for example NICE Science, Policy and Research advisers. By helping address major concerns with both observed and hidden biases in the use of routine data, this research will help future studies provide more robust evidence to inform treatment decisions in the best ways for improving population's health.

Decision makers are increasingly using routine data to evaluate health interventions when evidence from clinical trials is limited or unavailable. This has been particularly relevant to the evaluation of treatment strategies sustained over time due to the limitations of trials regarding the duration of follow-up and treatment adherence. The ability to derive valid estimates of sustained treatment effects from routine data critically depends on adequate control of both measured and unmeasured confounding, at each point in time. Instrumental variable (IV) methods have been widely used to address the confounding by exploiting sources of variation, known as 'instruments' (e.g. genetic factors), that explain treatment assignment but have no direct effect on outcomes. However, existing IV methods are not appropriate for addressing the time-varying nature of the confounding in the estimation of sustained treatment effects.
This research considers a novel approach that incorporates IV analysis into the marginal structural model (MSM) framework, which can adjust for the time-varying confounding. The project tackles outstanding methodological and practical issues that must be addressed before this combined IV-MSM approach can be adopted to derive valid inferences from routine data. We will achieve this in four related work packages:

1) To assess the validity of time-varying instruments through the use causal diagrams, graphical tools and falsification tests.
2) To address the recurrent issue of 'weak instruments', by exploring matching approaches and multiple robust estimators to improve the strength of time-varying instruments.
3) To extend the standard IV-MSM approach to deal with continuous instruments, as these are often considered in studies using routine data.
4) To translate the proposed methods to a broad audience by illustrating substantive applications of these across different clinical settings, and by providing software tools and practical guidance.