Tailoring health policies to improve outcomes using machine learning, causal inference and operations research methods

To maximise the impact of health policies on population health and improve the equitable distribution of health, policymakers require answers to questions such as: does the policy work for the intended recipients? Who benefits most? Does the policy reduce health inequalities? Who should be eligible for a programme? To generate evidence to answer these questions, policy evaluations need to go beyond the average population impact and consider how impacts differ across different types of individuals (treatment effect heterogeneity across subgroups). While such subgroup analysis has been done before, the previously used approaches are limited in that they open the door to the researcher cherry-picking the subgroups on the basis of what turns out as statistically significant in the estimates.
By contrast, machine learning techniques - automated algorithms that learn from the data - can reveal patterns in the policy impact that may not be expected beforehand. This is important for policymakers who need to understand who benefits most (and who does less, or not at all) from the implemented policy in question. Once we are able to assess how a given policy affects different sub-groups of the population and by how much, we can take these insights and design the eligibility criteria of a health policies, so that they maximise a decision maker's objective, for example by generating the most total health benefit given a fixed health care budget.

The combination of machine learning with methods that can estimate causal impacts of policy is relatively recent area of research, in particular their application to learn about "treatment effect heterogeneity" and the targeting of policies. Hence, there is no methodological guidance available on how to apply these recent tools in health policy evaluations. The proposed research aims to make a contribution by assessing and extending the available approaches to address the specific challenges that typically arise when evaluating health policies.
These include statistical challenges, such as the need to account for potential biases due to observed and unobserved differences between the treated and control groups; but also challenges to make evaluations relevant to decision making, by considering not just the benefits but also the costs of an intervention (cost effectiveness), and also considering budget constraints or considerations of equity when designing which population subgroups should be targeted with a policy.

This project proposes to address these challenges, by assessing and extending recently proposed machine learning and causal inference methods in the context of health policy evaluations and also by combining tools from different disciplines: causal inference, machine learning and cost-effectiveness modelling, for the first time. By successfully addressing these challenges, this project will deliver methods that will help researchers and policymakers carry-out more comprehensive evaluations of country-wide health policies. This could help support significant improvements to population health and reduce the health gap between the rich and poor within countries. The methodological developments are motivated by two case studies from a low- and middle-income country context, where the gains in terms of improving health and reducing health inequalities are particularly large. The case studies focus on two large scale health policies with ongoing relevance: major public health insurance reform in Indonesia and the country-wide Family Health Programme in Brazil.

To maximise impact on current health policy making, design of the specific research questions in the case studies will benefit from on-going input from Indonesian and Brazilian collaborators as well as policymakers. With extensive communication and impact activities, this project will make its methodological insights available for researchers working on health policy evaluations, in academia and beyond.

To maximise the impact of health policies, policy evaluations need to go beyond the average effects and consider treatment effect heterogeneity. Recently, machine learning approaches have been proposed to estimate treatment effect heterogeneity in a flexible yet principled way, combined with rigorous causal inference approaches. These methods have been further extended to examine which individuals should be targeted by a policy, based on their observed characteristics (optimal treatment regimes). However, these methods have not yet been translated to the challenging context of the evaluation of large-scale health policies. Here, due to non-experimental evaluation designs, researchers need to adjust for observed and unobserved confounding, and to inform real-world policy making, there is a need to estimate interpretable health policy assignment rules, to account for budgetary and equity constraints, and to consider the long term costs and benefits of a health policy.

To address these challenges, we plan to first compare combined causal inference/machine learning approaches to estimate heterogeneous treatment effects for subgroups and (conditional) individual level treatment effects. The methods will be extended to adjust for confounding in quasi-experimental studies, and we will test their performance by applying them in a case study and in a simulation study. Building on this work, we develop a framework that can answer the overarching question: which subgroups of individuals or communities should be 'treated' to maximise the health policy maker's objective function, subject to constraints? Third, to consider the full health benefits and costs of a policy, we combine with the above framework the tools of decision modelling.

Outputs from this project will equip researchers with advanced tools to generate impact evaluation evidence that can inform the design of efficient health policies to promote population health and health equity objectives.

The project is motivated by the need to better use existing data to better tailor large scale health policies, with a view to maximise population health given scarce resources. This research is anticipated to benefit a larger audience beyond the academic community.

The potential non-academic beneficiaries include:

1. Direct beneficiaries:

1.a. Researchers and technical analysts involved in impact evaluation, working for policy organisations in Brazil and Indonesia, and in the UK.

The project will benefit this group by resulting in methodological guidance to assess heterogeneity in impact evaluations and economic evaluations of health policies. The methodological guidance will be accompanied by training events.

1.b. Policy makers in Indonesia and Brazil who will be engaged throughout the research, in the following stages

- Refining the policy evaluation questions (e.g. providing inputs on relevant policy levers, equity considerations, and constraints in each country

- Ensuring there is a deep understanding of the contexts in which the case study policies were designed and implemented

- Being informed about emerging results, interpreting final results

In Indonesia, the impact evaluation of the Universal Health Insurance scheme is an ongoing activity by Indonesian and international researchers. With a large portion of the population still not covered by health insurance, learning about heterogeneous policy impacts, and the relative benefit of policy targeting rules can have a real impact on the ongoing re-design of the health insurance programmes. We aim to achieve this impact by engaging Indonesian Policy makers (Ministry of Health [see letter of support], the Health Technology Assessment Committee, and Social Security Administrator for Health) throughout.

In Brazil, potential plans of disinvestment in the Family Health Programme highlights the policy relevance on generating evidence on "for whom" the programme worked best, and can feed into future developments. This will be ensured by the engagement of policy makers such as the Paolo State Secretary of Health, the Ministry of Health, and the Brazilian Primary Care Secretariat, tasked with PSF policy formulation at the federal level. We will be supported in policy maker engagement by the World Bank (Brazil) [see letter of support].

By broadening the scope of evidence generated from policy evaluations and economic evaluations, on the long term, this research will help ensure that scarce societal resources are allocated in the best ways to improve population health.

2. Potential users of the outputs of the project include:

2.a. Researchers and technical analysts involved in impact evaluation in LMICs outside of Brazil and Indonesia (e.g. analysts working for non-governmental organisations (e.g. 3ie, Innovations for Poverty Action), intergovernmental organisations (World Bank [see letter of support], Global Fund, WHO).

2.b Researchers and technical analysts involved in health policy evaluation and health technology assessment in the UK (e.g. analysts working for Public Health England, NICE, Department of Health, Health Foundation [see letter of support]), by providing methodological guidance on the methods developed in the project, which are relevant outside of the LMIC context