Predicting women's height from their socioeconomic status: A machine learning approach

摘要

The social determinants of health literature routinely deploy socio-economic status (SES) as a key factor in accounting for women's height-an established indicator of human welfare at the population level-using traditional regression. However, this literature lacks a systematic identification of the predictive power of SES as well as the possible non-linear relationships between the measures of SES (education, occupation, and material wealth) in predicting variation in women's height. This study aims to evaluate this predictive power. We used the Demographic and Health Surveys (DHS) from 66 low- and middle-income countries (women = 1,273,644), sampled between 1994 and 2016. The analysis consisted of training seven machine-learning algorithms of different function classes and assessing their predictive power out-of-sample, vis-a-vis OLS regression. In an OLS framework, SES accounts for 0.7%, R-2, of the total variance in women's height (from sigma(2)(OLSFix) = 31.82 to sigma(2)(OLSSES) = 31.57), adjusting for country, community, and sampling year fixed effects. The country-specific variances range from as low as 25.10 units in Egypt to as high as 74.46 units in Sao Tome and Principe. With the same set of SES measures, the best performing learner, a Bayesian neural net, produces a predictive variance of sigma(2)(BnnSES) = 31.52. This is a negligible improvement in variance explained by 0.3% (sigma(2)(BnnSES) - sigma(2)(OLSSES)). Given our selection of algorithms, our findings indicate no relevant non-linear relationships between SES and women's height, and also the predictive limits of SES. We recommend that scholars report both the average effect of SES on health outcomes as well as its contribution to the variance explained. This will improve our understanding of how key social and economic factors affect health, deepening our understanding of the social determinants of health.