Predicting changes in U.S. county health: What does machine learning tell us about multiple determinants of health.

Abstract

Objective: Understanding multiple determinants of health and their contributions to premature mortality is essential for policymaking in population health. As an exploratory forecasting exercise, we aimed to analyze the associations between multiple determinants of health and improvement in premature mortality in U.S. counties.

Methods: Using County Health Rankings data on health determinants from 2010-2014, we predicted changes in premature mortality from 2015 to 2019. We used Extreme Gradient Boosting, which allows for flexible functional estimates of a multi-dimensional health production function, to assess the association between improvement in premature mortality and pre-period health determinants. We made Accumulated Local Effects plots to illustrate these associations and conducted a conventional spatial regression as a robustness check.

Results: We find a negative association between the baseline some-college rate (percentage of adults aged 25-44 with some post-secondary education) and change in premature mortality. A one-percentage-point increase in the baseline some-college rate in 2010-2014 is associated with a reduction of premature mortality by 9.57 years of potential life lost per 100,000 population from 2015 to 2019. The magnitude of the association for each county varied based on each county's baseline some-college rate.

Conclusions: Our findings reveal that improvements in post-secondary education are associated with relatively large, yet variable, gains in population health across U.S. counties. Applying machine learning methods to estimate a county-level health production function can inform tailored policies to improve local population health.