Financial Inclusion or Financial Structure? Within- and Between-Country Evidence on Banking Stability in the QUAD Economies

Abstract

Cross-country studies of financial inclusion and banking stability typically report a single coefficient estimated from panels that pool bank-based and market-based financial systems. This study asks what such a coefficient identifies when the inclusion index varies far more across countries than within them. An annual panel of the four Quadrilateral Security Dialogue (QUAD) economies—India, the United States, Japan, and Australia—covering 2004 to 2021 was assembled, and a principal component index of financial inclusion was constructed from International Monetary Fund and World Bank indicators of branch reach and credit and deposit depth and related to the country-level bank Z-score. The between-country variance of the index is approximately 40 times its within-country variance, so any estimator drawing on cross-sectional variation is identified from four country means rather than from 67 country-years. Consistent with this, the between-country association is large and negative (−1.60), whereas the within-country estimate is 0.10 and statistically indistinguishable from zero. A leave-one-country-out analysis traces the negative between-country result entirely to the United States, whose market-based system pairs low bank-based inclusion with the highest measured stability in the sample. A Hansen-type threshold model returns no significant split, and inspection of the regime assignment shows why: The residual-minimizing threshold separates India and the United States from Japan and Australia, with a single observation crossing it, so the model estimates a country grouping rather than a behavioral discontinuity. The one association surviving the within-country specifications is between banking-system capitalization and stability. In panels with few and structurally heterogeneous units, supply-side inclusion indices proxy financial architecture, and the resulting coefficient should not be interpreted as a behavioral parameter.