A Critique and Examination of the Polysocial Risk Score Approach: Predicting Cognition in the Health and Retirement Study.

Abstract

Polysocial risk scores were recently proposed as a strategy to improve clinical relevance of knowledge about social determinants of health. The objective of this paper was to assess if the polysocial risk score model improves prediction of cognition and all-cause mortality in middle-aged and older adults beyond simpler models including a smaller set of key social determinants of health. We used a sample of 13,773 individuals aged 50+ at baseline from the 2006 to 2018 waves of the Health and Retirement Study, a US population-based longitudinal cohort. Four linear mixed models were compared: two simple models including a priori selected covariates and two polysocial risk score models which used LASSO regularization to select covariates among 9 or 21 candidate social predictors. All models included age. Predictive accuracy was assessed via R-squared and root mean-squared prediction error (RMSPE) using training/test split and cross-validation. For predicting cognition, the simple model including age, race, gender, and education had an R-squared of 0.31 and an RMSPE of 0.880. Compared with this, the most complex polysocial risk score selected 12 predictors (R-squared=0.35 and RMSPE=0.858; 2.2% improvement). For all-cause mortality, the simple model including age, race, gender, and education had an AUROC of 0.747, while the most complex polysocial risk score did not demonstrate improved performance (AUROC = 0.745). Models built on a smaller set of key social determinants performed comparably to models built on a more complex set of social "risk factors".