Estimating the potential impact of behavioral public health interventions nationally while maintaining agreement with global patterns on relative risks

Abstract

This paper introduces a novel method to evaluate the local impact of behavioral scenarios on disease prevalence and burden with representative individual level data while ensuring that the model is in agreement with the qualitative patterns of global relative risk (RR) estimates. The method is used to estimate the impact of behavioral scenarios on the burden of disease due to ischemic heart disease (IHD) and diabetes in the Turkish adult population. Methods Disease specific Hierarchical Bayes (HB) models estimate the individual disease probability as a function of behaviors, demographics, socio-economics and other controls, where constraints are specified based on the global RR estimates. The simulator combines the counterfactual disease probability estimates with disability adjusted life year (DALY)-perprevalent-case estimates and rolls up to the targeted population level, thus reflecting the local joint distribution of exposures. The Global Burden of Disease (GBD) 2016 study metaanalysis results guide the analysis of the Turkish National Health Surveys (2008 to 2016) that contain more than 90 thousand observations.

Targeting resources efficiently and justifiably by combining causal machine learning and theory

“…While monotonicity and sign constraints can be imposed on linear models using prior distributions, the estimation of these Bayesian models, especially with large datasets, is computationally intensive and may not converge (Ali et al, 2020 ). GLM_QC is fit with the efficient glmnet algorithm with well-established convergence properties (Friedman et al, 2010 ).…”

“…One of the objectives in the 2019-2023 Strategic Plan of the Turkish Ministry of Health is to increase the physical activity of the citizens, including individual incentives for achieving certain physical activity levels (strategy 1.2.2 ). Governments around the world are implementing measures to increase physical activity (World Health Organization, 2019) that increasing physical activity has a substantial impact on decreasing diabetes and IHD prevalence (Ali et al, 2020). These two diseases constitute the fastest growing and the largest source of the burden of disease in the country, respectively.…”

“…Both models contain the following independent https://vizhub.healthdata.org/gbd-compare/ https://data.tuik.gov.tr/Bulten/Index?p=Turkey-Health-Survey-- Table 2 summarizes the qualitative constraints on the model coefficients: for example, the first line under the "Sign of the Effect" block indicates that all levels of physical activity (Insufficiently Active, Active, and Very Active) reduce disease risk relative to the Inactive, whereas the first line under the "Shape of the Effect" block indicates that the disease risk increases with age (not necessarily linearly)-all else being the same. These constraints are based on the 2016 GBD (Global Burden of Disease) meta-analysis results of hundreds of global studies on the relative risk (RR) of risk factors (BMI, Age, Gender), behaviors (Smoking, Secondhand Smoke, Physical Activity) for diabetes and IHD (Gakidou et al, 2017), as in Ali et al (2020). Note that when the reference risk is low, RR is approximately equal to the odds ratio (Zhang and Kai, 1998), and thus, the appropriate ln(RR) sign and monotonicity can be used for the GLM_QC model coefficients.…”

“…Incorporating domain knowledge with monotonicity constraints has been shown to improve the accuracy of machine learning models in Bayesian networks (Altendorf et al, 2012 ) and with shape constraints in generalized additive models (Pya and Wood, 2015 ). Ali et al ( 2020 ) empirically states that incorporating domain knowledge in Bayesian models with the sign and monotonicity constraints on the prior distribution (qualitative informative) rather than the traditional informative prior approach of setting the prior mean and standard deviation of the relevant model parameters leads to better models, i.e., with expected sign and monotonicity and holdout prediction performance.…”

“…The observational data consist of five waves of the biannual individual level nationally representative health surveys conducted by the Turkish Statistical Institute (TUIK) with more than ninety thousand observations, providing self-reported socioeconomic, demographic, behavior, and disease information. The theory-driven initial model terms and the qualitative constraints on sign and monotonicity are taken from the literature (Ali et al, 2020 ).…”