Impact of Economic Factors on Life Expectancy: A Machine Learning Approach

Abstract

Accurate estimation of population longevity is a critical input for macroeconomic planning, health-sector budgeting and international development monitoring. Leveraging a harmonised cross-sectional data set for 156 sovereign states, this study undertakes a rigorous comparative evaluation of three predictive frameworks: multiple linear regression, compact multilayer-perceptron neural networks and radial-basis support-vector regression applied to a common panel of economic, demographic and child-mortality indicators. Two parsimonious perceptron configurations (5–3 and 1 × 7 hidden-unit topologies) are trained with resilient back-propagation and subjected to hold-out testing. Forecast accuracy is scrutinised through mean error, mean absolute error (MAE), root-mean-squared error (RMSE), normalised RMSE and per cent bias. Both neural architectures decisively outperform the linear and kernel baselines, yielding out-of-sample MAE values of 0.17 year and 0.20 year, respectively, compared with 0.26 year for ordinary least squares and 0.32 year for the support-vector estimator; RMSE shows a commensurate hierarchy. Given the 16-year range of life expectancy in the sample, these sub-quarter-year deviations attest to the ability of even modest neural frameworks to capture non-linear interactions, most notably between external debt, crude birth rate, population scale and infant mortality proxies, that elude conventional models. Residual diagnostics confirm homoscedastic, unbiased errors for the multilayer perceptrons, whereas the support-vector regressor exhibits systematic under-prediction at the upper tail. The evidence underscores the methodological and practical utility of lightweight artificial neural networks for national longevity forecasting, furnishing policymakers with more precise baselines for targeted economic and public health interventions