Explaining Digital Payments Adoption with Econometrics and Explainable Machine Learning: Cross-Country Evidence from a Global Household Survey
By Saida Hajjaji
This paper studies the determinants of individual adoption of digital payments using the Global Findex Database 2025, released by the World Bank and based on nationally representative household surveys conducted in 2024 across 141 economies. We combine a standard econometric approach with explainable machine learning (ML) methods in order to provide both transparent global benchmarks and granular, policy-relevant insights into digital payment behaviours.
We first estimate a parsimonious logistic regression model on a broad multi-country sample of 5,189 adults from 97 economies, focusing on core socio-demographic and structural characteristics commonly used in the financial inclusion literature. While this baseline specification confirms well-established patterns related to education, income, and access to digital tools, its predictive performance remains limited, as expected given the restricted set of covariates and the objective of cross-country comparability.
We then exploit a richer subsample of 5,183 employed individuals from 74 economies for whom detailed information on digital income reception, payment use cases, and digital connectivity is available. On this enriched sample, we estimate regularized logistic regression, random forest, and gradient boosting models. Predictive performance improves substantially, with the area under the ROC curve increasing from approximately 0.61 in the baseline logit to up to 0.94 in the best-performing ML model. Using SHapley Additive exPlanations (SHAP), we show that adoption of digital payments is primarily driven by participation in digital income and payment ecosystems—such as receiving wages, transfers, pensions, or agricultural income digitally, making domestic remittances, and paying merchants and utilities—together with a synthetic index capturing digital connectivity.
Overall, our results illustrate a strong complementarity between traditional econometrics and explainable machine learning. While the econometric model provides an interpretable global benchmark, the ML–SHAP framework uncovers heterogeneous mechanisms that are not captured by standard specifications. These findings have direct implications for the sequencing of digital finance reforms, the design of digital payment strategies, and the transition toward retail central bank digital currencies, particularly in emerging and developing economies.
Source SSRN
