Worker and workplace predictors of self-reported health status: An application of econometrics and machine learning

This paper applies an array of econometric and machine learning techniques to individual-level data from the 2008 National Study of the Changing Workforce to identify worker and workplace features that appear to be strong predictors of self-reported health status. Marginal effects and log-odds ratios are discussed, as is prediction performance. Of all the methods used, the (gradient) boosted-tree and support-vector-machine models delivered the best prediction accuracy. In assessing variable importance, several features consistently stood out cross the array of analyses: depression, sleep difficulties, home stress, race, Hispanic ethnicity, life satisfaction, work-family conflict, education, and access to specific types of flexible working arrangements. Earnings, age, and female variables, while not significant in the baseline analyses, emerged as relevant in several sensitivity analyses. Some of these findings are policy items (e.g., flexible working arrangements, work-family conflict issues as well as race, gender, and ethnicity issues), which might be considered by firms, unions, and/or policy makers to improve self-reported health results moving forward; other significant features such a depression and sleep difficulties are more within the individual's purview to remediate.