How Well Do Objective Regional Conditions Predict Life Satisfaction? Evidence from 41 OECD Countries

• In Event: Performance and its Discontents: Contemporary Challenges in Public Management

Abstract

Place-based well-being policy rests on two rarely tested assumptions: that the objective conditions captured in regional dashboards track how residents evaluate their lives, and that regions can be meaningfully benchmarked on common metrics across national borders. We examine both using the October 2025 release of the OECD Regional Well-Being database, covering 468 TL2 regions in 41 OECD member and accession countries, with life satisfaction observed for 446 of them. Nine learners, from linear regression and the elastic net to random forests, gradient boosting, and stacked ensembles, are compared under two contrasted validation designs: random five-fold cross-validation, in which regions of the same country inform one another's predictions, and country-blocked cross-validation, in which entire countries are held out of training. Gradient boosting predicts roughly three-quarters of the cross-regional variance under random validation (R2 = 0.76), nearly double the linear benchmark (0.48), but only one-fifth when countries are held out (0.20). The gap between the two figures indexes how much of the apparent predictability of regional well-being is carried by country membership rather than by a portable region-level mapping from conditions to satisfaction. Leave-one-country-out analysis sharpens the point: the model transfers a country-level shift in mean satisfaction while failing to recover the internal ordering of a country's regions, and national biases are large, with Korea, Japan, and Turkiye reporting less satisfaction than their measured conditions predict and New Zealand and Israel reporting more. SHAP decompositions identify household disposable income as the dominant contributor, entering negatively below approximately USD 25,000 (PPP) and flattening beyond USD 30,000, a regional counterpart to the individual-level satiation debate. Conformal prediction intervals at the 90 percent level flag 45 regions as significantly mismatched, in capital and peripheral regions alike. The analysis is predictive and descriptive rather than causal, but it suggests that cross-national well-being comparisons rest largely on country-level structure, and that the choice of validation design is itself a substantive analytic decision.

Author

• JungHwan Ryu, Research Center for Public Human Resource Development / SungKyunKwan University
  Presenting Author