Does Digital Commerce Deepening Increase Banking Fragility? Evidence from Household Leverage in a Cross-Country Panel

1. Introduction

Banking fragility remains a central concern for risk management because modern banks transform liquid liabilities into comparatively illiquid assets while operating within interconnected financial systems. This structure makes banking systems socially valuable, but also vulnerable when confidence, liquidity, or asset quality deteriorates. The classic fragility problem is not limited to deposit runs; it also includes the broader possibility that shocks to borrowers, collateral values, or credit quality can move through bank balance sheets and weaken financial intermediation (Diamond & Dybvig, 1983).

A large macro-financial literature shows that credit and leverage cycles are closely linked to financial distress. Credit booms can precede banking crises, and recessions that follow credit-intensive expansions tend to be deeper and more persistent than ordinary downturns. This evidence is especially relevant for household finance because retail credit expansions may initially look like financial deepening while gradually increasing the sensitivity of banks to borrower distress (Schularick & Taylor, 2012; Jordà et al., 2013).

Nonperforming loans are one of the most direct observable indicators of banking fragility because they connect household and firm repayment capacity with the quality of bank assets. Rising nonperforming loans can weaken profitability, absorb capital, restrict new credit, and reinforce macro-financial feedback loops. Empirical evidence on nonperforming loans shows that adverse macroeconomic conditions and bank-level risk exposures are systematically associated with worsening loan quality (Nkusu, 2011; Louzis et al., 2012).

The cross-country relevance of this issue has increased as banking crises have repeatedly imposed large fiscal and output costs. Updated evidence from systemic banking crisis episodes indicates that crisis identification requires attention not only to bank failures but also to severe distress, policy interventions, and the broader costs of financial instability. For this reason, a country-year panel approach is useful for examining whether emerging drivers of retail finance are associated with asset-quality deterioration across different institutional and economic contexts (Laeven & Valencia, 2020).

At the same time, financial intermediation is being reshaped by digital commerce and digital finance. Digital technologies reduce the costs of search, storage, matching, payment, and information processing. These reductions can change how consumers shop, borrow, repay, and interact with financial institutions. In this sense, digital commerce is not only a retail phenomenon; it is part of a wider transformation in the economics of transaction costs, data, and platform-mediated exchange (Goldfarb & Tucker, 2019).

Fintech also changes the structure of financial services by enabling new entrants, new underwriting technologies, faster payments, and alternative distribution channels. These innovations can improve access and efficiency, but they may also complicate supervision if credit expansion occurs outside traditional monitoring frameworks or if digital interfaces accelerate borrower decisions without equivalent improvements in risk assessment. This dual nature makes digital finance directly relevant to banking fragility and financial risk management (Philippon, 2016).

Evidence from technology-enabled lending suggests that digitalization can improve screening and operational efficiency, but the implications for risk are not mechanically one-directional. FinTech mortgage lenders can process applications faster without necessarily increasing default rates, while digital footprints can add predictive information for consumer credit scoring. These findings imply that digital adoption may either support safer credit allocation or expand credit to marginal borrowers, depending on the quality of underwriting and the institutional environment (Fuster et al., 2019; Berg et al., 2020).

A second strand of evidence highlights the regulatory dimension. The rise of technology-based and nonbank lenders can partly reflect technological efficiency, but it can also reflect regulatory arbitrage and shifts in activity toward institutions that are supervised differently from traditional banks. This matters for the present study because app-based commerce and internet banking may proxy not only consumer digital adoption, but also the changing boundary between bank-intermediated and platform-mediated finance (Buchak et al., 2018).

Digital financial inclusion has expanded rapidly, particularly through digital payments, mobile access, and account ownership. This expansion can strengthen resilience by lowering transaction costs and improving access to formal finance, but it may also increase the speed and frequency with which households participate in credit-linked consumption ecosystems. Consequently, the same digital infrastructure that supports inclusion may also increase the importance of monitoring household leverage (Demirgüç-Kunt et al., 2022).

The post-pandemic policy debate reinforces this ambiguity. Fintech can support inclusion, lower service costs, and increase reach, especially in emerging markets, but its benefits depend on adequate consumer protection, supervision, data governance, and macroprudential oversight. Therefore, the relationship between digital finance and banking fragility should be treated as an empirical question rather than assumed to be either stabilizing or destabilizing (Sahay et al., 2020).

Institutional quality is central to this empirical question. Government effectiveness captures the capacity of the public sector to design and implement credible policies, deliver public services, and maintain institutional execution. In the context of digital finance and household leverage, stronger institutional execution may reduce the probability that credit deepening becomes banking fragility by supporting supervision, consumer protection, enforcement, and macroprudential coordination (Kaufmann et al., 2010).

This study contributes to the literature by estimating a dynamic country-year model that links digital commerce, digital banking, household debt, and nonperforming loans in a single empirical framework. The design also responds to a methodological constraint: although dynamic panel generalized method of moments is attractive for persistent outcomes and potentially endogenous regressors, small cross-sectional samples can make instrument proliferation and overidentification diagnostics unreliable. For that reason, the main specification uses dynamic two-way fixed effects with country and year effects, while generalized method of moments is treated only as an exploratory extension for future larger panels (Roodman, 2009).

The study is organized around four hypotheses. First, higher household debt is expected to increase banking fragility because more leveraged households are more vulnerable to income, interest-rate, and liquidity shocks. Second, app-based retail mobile commerce is expected to influence household leverage by reducing transactional frictions and increasing the convenience of digitally mediated consumption. Third, internet banking may have an ambiguous direct effect on banking fragility: it can improve access, monitoring, and payment discipline, but it may also facilitate broader credit use. Fourth, government effectiveness is expected to attenuate the transmission from household leverage to nonperforming loans by improving institutional execution and financial-sector governance.

The empirical findings refine these expectations. Banking fragility is highly persistent, and household debt shows a positive association with nonperforming loans, although the strength of the result depends on the covariance estimator and robustness specification. App-based mobile commerce does not support a direct fragility-amplification interpretation; instead, its negative coefficient suggests that digital commerce may partly capture formalization, payment efficiency, or digital maturity. Internet banking does not show a robust direct effect, and the institutional moderation term is not statistically stable. These results suggest that digitalization should not be interpreted as a mechanical source of banking fragility; the more consistent risk channel is household leverage.

1.1. Theoretical Framing and Contribution

The theoretical argument developed in this study links three mechanisms. The first is a balance-sheet mechanism: household leverage can increase the vulnerability of banks when borrower repayment capacity weakens. The second is a digital-friction mechanism: app-based commerce, internet banking, and smartphone penetration reduce the costs of consumption, payment, and financial access, thereby potentially affecting the scale and composition of household financial obligations. The third is an institutional-execution mechanism: government effectiveness may condition whether digital financial deepening is absorbed through better formalization and monitoring or translated into weaker loan quality.

The paper advances the literature in three ways. First, it connects digital commerce to banking fragility through the household-debt channel, rather than treating digitalization only as a productivity or inclusion phenomenon. Second, it separates app-based commerce from internet banking, allowing the empirical model to distinguish between digitally mediated retail activity and digitally mediated banking use. Third, it tests institutional moderation explicitly, which is important because the same level of digital adoption may have different consequences in countries with different policy execution and supervisory capacities.

The contribution is also methodological. The analysis begins with a full coverage audit of the Passport country-year series and adapts the model to the data that are empirically supportable. This avoids presenting an overparameterized specification that would be theoretically attractive but statistically fragile. The final design therefore privileges a transparent dynamic two-way fixed effects framework, with robustness checks for cross-sectional dependence, alternative winsorization, housing stress, crisis-year exclusion, mechanism testing, and placebo leads. This structure is suitable for a first cross-country test of the digital commerce-household leverage-banking fragility nexus.

Overall, the study reframes the debate on digital finance and risk. Digitalization may accelerate financial access and consumption, but it can also improve information, formalization, payment discipline, and operational efficiency. The central empirical question is therefore not whether digitalization is inherently risky, but under what conditions it is associated with higher household leverage and whether that leverage translates into nonperforming loans. The results suggest that household leverage remains the more reliable warning channel, while digital commerce and digital banking require a more nuanced interpretation.

2. Materials and Methods

2.1. Study Design

This study uses a cross-country country-year panel design to examine whether the deepening of digital commerce and digital finance is associated with banking fragility through the household leverage channel. The empirical strategy is diagnostic and explanatory rather than predictive. The objective is to estimate whether household debt, app-based retail digitalization, and digital banking use are systematically related to the ratio of bank nonperforming loans to total gross loans, while accounting for persistent banking fragility, country-specific heterogeneity, common year shocks, and institutional conditions.

The design follows three sequential stages. First, the raw Passport series were harmonized into a country-year structure and screened for coverage across variables, countries, and years. Second, a main dynamic two-way fixed effects specification was estimated on the largest empirically defensible sample allowed by the data. Third, robustness checks were conducted using alternative covariance estimators, winsorization thresholds, restricted controls, crisis-year exclusions, mechanism regressions, and placebo leads. The final empirical design prioritizes transparency and replicability over overfitting, given the limited cross-sectional size of the available matched sample.

2.2. Data Source, Period, and Sample Construction

The data were obtained from Euromonitor International’s Passport database, using series extracted from the Economies, Channels, and Consumers modules (Euromonitor International, 2026). The initial extraction covered annual country-level observations from 2000 to the latest year available in each series. However, the usable estimation window was restricted by the availability of app-based mobile commerce and institutional indicators. The main model therefore uses the period 2014-2024.

The dependent variable is banking fragility, measured as bank nonperforming loans to total gross loans. The main explanatory variables are household debt, app-based retail mobile commerce, and internet banking use. The control variables are smartphone possession and government effectiveness. The house price-to-income ratio was initially considered as a housing-stress control, but its lower country coverage substantially reduced the estimation sample; therefore, it was retained only for restricted robustness analysis.

The main estimation sample is an unbalanced panel of 171 country-year observations from 16 countries: Australia, Brazil, Canada, France, Germany, Indonesia, Italy, Poland, South Africa, South Korea, Spain, Sweden, Thailand, Turkey, the United Kingdom, and the United States. A more restrictive specification including the house price-to-income ratio reduces the sample to 128 observations from 12 countries. This difference motivated the decision to exclude housing stress from the baseline model and use it only as a robustness check.

Government effectiveness was used as the institutional execution variable because it captures perceptions of public-service quality, policy implementation, bureaucratic capacity, and credibility of government commitments. The variable is conceptually aligned with the role of institutional quality in moderating the transmission of household leverage into banking fragility and is consistent with the Worldwide Governance Indicators methodology (Kaufmann et al., 2010).

2.3. Variable Treatment and Transformations

All variables were transformed before estimation to improve comparability across countries and reduce the influence of scale differences. The dependent variable was transformed as:

$$y_{\left\{it\right\}}=ln\left(1+NP{L}_{\left\{it\right\}}\right)$$

(1)

where NPL_it denotes bank nonperforming loans to total gross loans in country i and year t. The household debt variable was also transformed using ln(1 + x), because it is reported as a monetary level. App-based mobile commerce was not treated as a percentage or share. Since the extracted Passport series is expressed as a monetary value, it was transformed using an inverse hyperbolic sine transformation:

$$AppMCo{m}_{it}^{*}=\mathrm{asinh}\left(AppMCo{m}_{it}\right)$$

(2)

This transformation preserves zero and low values while reducing skewness in monetary variables. Internet banking use and smartphone possession were treated as bounded percentage variables. They were converted to proportions, clipped at the interval [0.0001, 0.9999], and transformed using the logit function:

$$logit\left(x\right)=\ln \left({\displaystyle \frac{x}{1-x}}\right)$$

(3)

All transformed predictors were standardized using z-scores before estimation. This allows the estimated coefficients to be interpreted as the expected change in transformed banking fragility associated with a one-standard-deviation change in each explanatory variable. To reduce the influence of extreme observations, the baseline specification used 1st and 99th percentile winsorization. A 2.5th and 97.5th percentile winsorization rule was used as a robustness check.

All explanatory variables were lagged by one year in the main specification. This lag structure reduces simultaneity concerns and aligns the empirical model with the theoretical mechanism that digitalization and household leverage may affect banking fragility with a delay. The dependent variable was also included with a one-year lag to account for persistence in nonperforming loans.

2.4. Baseline Econometric Specification

The baseline empirical model is a dynamic two-way fixed effects panel model with country and year effects:

$$\begin{array}{ll}{y}_{it}=& {\alpha }_i+{\lambda }_t+\rho y_{i,t-1}+{\beta }_{1}Deb{t}_{i,t-1}^{*}+{\beta }_{2}AppMCo{m}_{i,t-1}^{*}\\ & +{\beta }_{3}Bankin{g}_{i,t-1}^{*}+{\beta }_{4}Smartphon{e}_{i,t-1}^{*}+{\beta }_{5}GovEf{f}_{i,t-1}^{*}\\ & +{\beta }_6\left(Deb{t}_{i,t-1}^{*}\times GovEf{f}_{i,t-1}^{*}\right)+{\epsilon }_{it}\end{array}$$

(4)

where y_it is the natural logarithm of one plus the nonperforming-loan ratio, alpha_i denotes country fixed effects, lambda_t denotes year fixed effects, Debt*_i,t-1 is standardized transformed household debt, AppMCom*_i,t-1 is standardized transformed app-based mobile commerce, Banking*_i,t-1 is standardized transformed internet banking use, Smartphone*_i,t-1 is standardized transformed smartphone possession, and GovEff*_i,t-1 is standardized government effectiveness. The interaction term tests whether government effectiveness moderates the household-debt-to-fragility relationship.

Country fixed effects absorb time-invariant structural differences across economies, including long-run institutional, demographic, banking-system, and cultural differences. Year fixed effects absorb common global shocks, such as the COVID-19 period, monetary-policy cycles, global risk sentiment, and international credit conditions. The baseline standard errors are clustered at the country level to account for within-country serial correlation and heteroskedasticity. The use of robust covariance estimation is consistent with the general econometric concern that conventional standard errors may be inconsistent under heteroskedastic disturbances (White, 1980).

2.5. Robustness Strategy and Diagnostic Tests

The robustness strategy was organized around five concerns: dynamic persistence, cross-sectional dependence, sample sensitivity, omitted housing stress, and possible reverse timing. First, the main dynamic two-way fixed effects model was compared with a static two-way fixed effects model that excludes y_i,t-1. This comparison tests whether the digitalization and leverage coefficients are driven by persistence in nonperforming loans.

Second, Driscoll-Kraay standard errors were estimated as a sensitivity check because cross-country panels may exhibit contemporaneous dependence caused by global financial cycles, regional shocks, or common macroeconomic conditions (Driscoll & Kraay, 1998). Third, cross-sectional residual dependence was evaluated using the Pesaran CD test, which is designed to detect cross-sectional dependence in panel-data settings (Pesaran, 2004).

Fourth, the baseline specification was re-estimated after including the house price-to-income ratio as an additional housing-stress control. This restricted model uses a smaller matched sample and is therefore interpreted as a robustness check rather than as the main specification. Fifth, the model was re-estimated after excluding the years 2020 and 2021 to verify whether the estimated relationships were mechanically driven by the COVID-19 shock. Additional robustness checks used alternative winsorization thresholds and placebo leads of digital variables. The placebo models test whether future values of digitalization variables predict current banking fragility, which would weaken the temporal interpretation of the baseline lag structure.

System generalized method of moments was considered as an exploratory robustness option because dynamic panels with lagged dependent variables and potentially endogenous regressors can be estimated using internal instruments derived from lagged levels and differences (Arellano & Bond, 1991; Blundell & Bond, 1998). However, it was not used as the principal estimator because the final matched panel contains only 16 countries. In such a small-N setting, system GMM can become fragile due to instrument proliferation and weak overidentification diagnostics. Therefore, if reported, the GMM estimates should be interpreted as supplementary and should use collapsed instruments, restricted lag depth, and a parsimonious instrument matrix following standard applied guidance (Roodman, 2009).

2.6. Mechanism Model

To evaluate whether digitalization is associated with household leverage before affecting banking fragility, a mechanism regression was estimated with household debt as the dependent variable:

$$Deb{t}_{it}^{*}={\alpha }_i+{\lambda }_t+{\theta }_{1}AppMCo{m}_{i,t-1}^{*}+{\theta }_{2}Bankin{g}_{i,t-1}^{*}+{\theta }_{3}Smartphon{e}_{i,t-1}^{*}+{\theta }_{4}GovEf{f}_{i,t-1}^{*}+u_{it}$$

(5)

This mechanism model tests whether app-based mobile commerce, internet banking, and smartphone penetration are associated with subsequent household leverage. The model is not interpreted as a full mediation design because the available country-year panel does not contain sufficiently rich household-level credit, income, interest-rate, or loan-origination data. Instead, it is used as a supporting test of the theoretical channel linking digital adoption, consumer financial deepening, and household debt.

2.7. Software, Reproducibility, and Data Availability

The data were processed and analyzed using Python 3.13.5. The main packages used were pandas 2.2.3, NumPy 2.3.5, statsmodels 0.14.6, SciPy 1.17.0, scikit-learn 1.8.0, matplotlib 3.10.8, and openpyxl 3.1.5. The workflow included data import, country-year harmonization, missingness and coverage diagnostics, variable transformation, lag construction, winsorization, model estimation, diagnostic testing, robustness checks, and figure/table generation.

The replication materials include the processed model-ready panel, the long-format extracted data, coverage tables, coefficient tables, diagnostic outputs, robustness tables, and figures. Because the raw data were obtained from a third-party commercial database, restrictions may apply to redistributing the original Passport extraction files. The processed analytical files and code can be made available as supplementary materials subject to the licensing conditions of the data provider.

2.8. Use of Generative Artificial Intelligence

During manuscript preparation, generative artificial intelligence was used to assist with language drafting, editorial structuring, and academic style refinement. It was not used to generate the raw data, fabricate results, or independently conduct the statistical analysis. All empirical results were produced from the extracted dataset and verified by the author. The author reviewed, edited, and takes full responsibility for the final manuscript content.

3. Results

3.1. Coverage, Estimation Sample, and Model Scope

The merged Passport panel confirms that the original broad conceptual model required adaptation to the actual data coverage. Although the nonperforming-loan indicator has broad coverage across countries and years, the matched sample is constrained by the availability of household debt and app-based mobile commerce. The final baseline sample contains 171 country-year observations from 16 countries over the effective estimation period 2015-2025, after transformation, lag construction, and listwise deletion for the variables required by the main dynamic specification. The restricted model that additionally includes the house price-to-income ratio contains 128 observations from 12 countries, which confirms that this housing-stress variable is better treated as a robustness control than as a baseline regressor.

The empirical design therefore uses a dynamic two-way fixed effects model as the main specification. Country fixed effects control for time-invariant cross-country heterogeneity, year fixed effects control for common global shocks, and the lagged dependent variable captures persistence in banking fragility. All reported coefficients are based on transformed and standardized predictors, so the main coefficients are comparable in standard-deviation units.

3.2. Main Dynamic Two-Way Fixed Effects Estimates

Table 1 reports the main dynamic estimates and key robustness variants. The lagged nonperforming-loan variable is positive and highly significant across all dynamic specifications. In the baseline dynamic model with country-clustered standard errors, the coefficient of lagged NPL is 0.461 and is significant at the 1% level. The result remains virtually identical under Driscoll-Kraay standard errors and remains stable in the restricted housing-stress model, the alternative winsorization model, and the model excluding 2020-2021. This finding indicates strong persistence in banking fragility.

Household debt has a positive coefficient in the main dynamic model. The coefficient is 0.192 and is statistically significant at the 10% level with country-clustered standard errors. Under Driscoll-Kraay standard errors, the same coefficient becomes significant at the 5% level. The coefficient remains positive under the 2.5/97.5 winsorization robustness check, where it is also significant at the 5% level. However, it loses statistical significance in the restricted house-price/income model and in the model excluding 2020-2021. The evidence therefore supports a moderate, but not uniformly robust, positive association between lagged household leverage and banking fragility.

App-based mobile commerce has a negative and statistically significant coefficient in the main dynamic model. The baseline coefficient is -0.155 and is significant at the 5% level. The result strengthens under Driscoll-Kraay standard errors and remains negative and significant in the alternative winsorization and no-2020-2021 specifications. This result does not support the initial hypothesis that app-based retail digitalization directly increases banking fragility. Instead, it suggests that app-based commerce may capture formalization, transaction efficiency, or broader digital-market maturity rather than a direct risk-amplifying channel.

Internet banking is not statistically significant in the dynamic models. While the static model shows a weakly positive coefficient, this effect disappears once lagged banking fragility is included. Smartphone possession is also generally not robustly significant, although the Driscoll-Kraay specification reports a negative coefficient significant at the 5% level. Government effectiveness is not significant in the dynamic models, and the interaction between household debt and government effectiveness is not statistically significant. Therefore, the proposed institutional moderation hypothesis is not supported by the available sample.

Table 1. Main dynamic estimates and robustness checks.

Variable	M2 Dynamic TWFE cluster	M3 Dynamic TWFE DK	M4 Dynamic + HPI	M6 Winsor 2.5/97.5	M7 Excl. 2020-2021
NPL t-1 (z, log)	0.461*** (0.029)	0.461*** (0.031)	0.492*** (0.038)	0.446*** (0.028)	0.480*** (0.020)
Household debt t-1 (z, asinh)	0.192* (0.097)	0.192** (0.079)	0.229 (0.302)	0.191** (0.084)	0.157 (0.090)
App M-commerce t-1 (z, asinh)	-0.155** (0.061)	-0.155*** (0.031)	-0.091 (0.125)	-0.168*** (0.054)	-0.155** (0.058)
Internet banking t-1 (z, logit)	0.022 (0.021)	0.022 (0.023)	0.067 (0.055)	0.025 (0.021)	0.005 (0.016)
Smartphone possession t-1 (z, logit)	-0.051 (0.032)	-0.051** (0.023)	-0.031 (0.037)	-0.055 (0.036)	-0.039 (0.028)
Government effectiveness t-1 (z)	0.016 (0.035)	0.016 (0.030)	-0.017 (0.034)	0.027 (0.033)	0.017 (0.035)
Household debt × Gov. effectiveness	-0.013 (0.026)	-0.013 (0.044)	0.025 (0.083)	-0.030 (0.028)	-0.020 (0.027)
House price/income t-1 (z)			0.046 (0.029)
Observations	171	171	128	171	139
Countries	16	16	12	16	16
R-squared	0.974	0.974	0.983	0.973	0.977

¹ Coefficients are reported with standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. All main predictors are transformed, lagged one year, and standardized. Country and year fixed effects are included in all models shown.

Table 1. Main dynamic estimates and robustness checks.

Variable	M2 Dynamic TWFE cluster	M3 Dynamic TWFE DK	M4 Dynamic + HPI	M6 Winsor 2.5/97.5	M7 Excl. 2020-2021
NPL t-1 (z, log)	0.461*** (0.029)	0.461*** (0.031)	0.492*** (0.038)	0.446*** (0.028)	0.480*** (0.020)
Household debt t-1 (z, asinh)	0.192* (0.097)	0.192** (0.079)	0.229 (0.302)	0.191** (0.084)	0.157 (0.090)
App M-commerce t-1 (z, asinh)	-0.155** (0.061)	-0.155*** (0.031)	-0.091 (0.125)	-0.168*** (0.054)	-0.155** (0.058)
Internet banking t-1 (z, logit)	0.022 (0.021)	0.022 (0.023)	0.067 (0.055)	0.025 (0.021)	0.005 (0.016)
Smartphone possession t-1 (z, logit)	-0.051 (0.032)	-0.051** (0.023)	-0.031 (0.037)	-0.055 (0.036)	-0.039 (0.028)
Government effectiveness t-1 (z)	0.016 (0.035)	0.016 (0.030)	-0.017 (0.034)	0.027 (0.033)	0.017 (0.035)
Household debt × Gov. effectiveness	-0.013 (0.026)	-0.013 (0.044)	0.025 (0.083)	-0.030 (0.028)	-0.020 (0.027)
House price/income t-1 (z)			0.046 (0.029)
Observations	171	171	128	171	139
Countries	16	16	12	16	16
R-squared	0.974	0.974	0.983	0.973	0.977

¹ Coefficients are reported with standard errors in parentheses. * p < 0.10, ** p < 0.05, *** p < 0.01. All main predictors are transformed, lagged one year, and standardized. Country and year fixed effects are included in all models shown.

Figure 1. Coefficient plot for the main dynamic two-way fixed effects model.

Figure 1. Coefficient plot for the main dynamic two-way fixed effects model.

3.3. Mechanism Model: Digitalization and Household Debt

The mechanism model uses household debt as the dependent variable to evaluate whether digitalization is associated with subsequent leverage. App-based mobile commerce has a positive coefficient of 0.499, but it is not statistically significant at conventional levels. Smartphone possession has a positive coefficient of 0.194 and is significant at the 10% level. Internet banking is close to zero and not significant, while government effectiveness is negative and not significant. These results suggest that the proposed digitalization-to-debt channel is plausible but empirically weak in the current panel. The mechanism requires richer credit-market variables, such as household debt relative to disposable income, consumer-loan growth, credit-card penetration, or household debt-service burden, to be tested more directly.

Table 2. Mechanism regression with household debt as the dependent variable.

Variable	M5 Mechanism: household debt
App M-commerce t-1 (z, asinh)	0.499 (0.286)
Internet banking t-1 (z, logit)	-0.011 (0.088)
Smartphone possession t-1 (z, logit)	0.194* (0.107)
Government effectiveness t-1 (z)	-0.317 (0.212)
Observations	171
Countries	16
R-squared	0.991

Table 2. Mechanism regression with household debt as the dependent variable.

Variable	M5 Mechanism: household debt
App M-commerce t-1 (z, asinh)	0.499 (0.286)
Internet banking t-1 (z, logit)	-0.011 (0.088)
Smartphone possession t-1 (z, logit)	0.194* (0.107)
Government effectiveness t-1 (z)	-0.317 (0.212)
Observations	171
Countries	16
R-squared	0.991

3.4. Diagnostic Evidence

The diagnostic tests are consistent with the use of robust covariance estimation. The Pesaran CD test does not detect statistically significant residual cross-sectional dependence in the main dynamic model. The residual AR(1) pooled correlation is also not statistically significant. However, the Breusch-Pagan test indicates heteroskedasticity. This confirms the need to report robust standard errors and supports the inclusion of Driscoll-Kraay standard errors as a sensitivity check.

Table 3. Diagnostic tests for the main dynamic model.

Diagnostic	Statistic	p-value	Detail
Pesaran CD residual cross-sectional dependence	-1.493	0.135	avg pairwise corr=-0.042; pairs=120
Residual AR(1) pooled correlation	0.069	0.392	n pairs=155
Breusch-Pagan heteroskedasticity	49.137	0.027	F=1.739; F p=0.016

Table 3. Diagnostic tests for the main dynamic model.

Diagnostic	Statistic	p-value	Detail
Pesaran CD residual cross-sectional dependence	-1.493	0.135	avg pairwise corr=-0.042; pairs=120
Residual AR(1) pooled correlation	0.069	0.392	n pairs=155
Breusch-Pagan heteroskedasticity	49.137	0.027	F=1.739; F p=0.016

4. Discussion

The first and most robust empirical result is the strong persistence of banking fragility. The positive and highly significant coefficient on lagged NPLs suggests that problem loans do not adjust instantaneously; rather, they display inertia across time. This is consistent with the view that loan-quality deterioration can persist through bank balance sheets, risk-management practices, and cost-efficiency channels. The result also validates the decision to use a dynamic model rather than relying only on a static fixed effects specification (Berger & DeYoung, 1997).

The second important result is the positive association between household debt and future banking fragility. Although this effect is moderate and not uniformly significant across all restricted specifications, it is positive in the dynamic baseline and becomes statistically stronger under Driscoll-Kraay standard errors and alternative winsorization. Substantively, this supports the broader macro-financial argument that private credit expansions can leave banking systems more exposed when household balance sheets become strained. The result is aligned with historical evidence showing that credit booms and leverage cycles are central to financial-crisis risk (Schularick & Taylor, 2012).

The evidence on household leverage is also consistent with the broader literature on the macroeconomic costs of credit-intensive expansions. The fact that household debt remains positive after controlling for country and year fixed effects suggests that the leverage channel is not merely capturing stable cross-country differences or common global shocks. However, the loss of significance in the restricted housing-stress model implies that the estimated leverage effect should be interpreted cautiously. The model supports household debt as a relevant fragility channel, but not as a standalone sufficient explanation of banking risk (Jordà et al., 2013).

The most unexpected result is the negative association between app-based mobile commerce and banking fragility. This finding does not support the original hypothesis that app-based retail digitalization directly increases nonperforming loans through easier consumption and credit access. A more plausible interpretation is that app-based commerce captures digital-market maturity, payment formalization, transaction traceability, or more efficient consumer ecosystems. This interpretation is consistent with the idea that financial and digital innovation may have both stabilizing and destabilizing effects depending on the institutional and banking context in which it develops (Beck et al., 2016).

Internet banking does not show a robust direct relationship with NPLs once persistence in banking fragility is included. This suggests that digital financial usage, at least as captured by aggregate internet banking penetration, is not equivalent to risky credit expansion. Digital finance may expand access, lower transaction costs, and improve financial inclusion, but its effect on financial stability depends on credit underwriting, supervision, consumer protection, and the type of financial product being adopted. For this reason, the findings should be read as evidence against a simple digitalization-equals-fragility mechanism, rather than as evidence that digital finance is risk-free (Sahay et al., 2020).

The institutional moderation hypothesis is not supported. Government effectiveness is not statistically significant in the main dynamic specification, and its interaction with household debt is also insignificant. This does not imply that institutions are irrelevant to banking stability. Instead, it suggests that the specific annual country-level measure used here may be too broad to capture the supervisory, regulatory, and consumer-credit channels through which institutions shape the debt-to-NPL transmission. Future versions of the model should consider more targeted indicators of financial regulation, bank supervision, consumer-credit protection, macroprudential policy, and insolvency enforcement.

Overall, the findings require a refinement of the theoretical framing. The original model assumed that digital commerce deepening could increase banking fragility by stimulating household leverage. The empirical evidence instead suggests a more nuanced argument: household leverage is the clearest fragility channel, whereas digitalization has ambiguous effects. App-based mobile commerce appears to be associated with lower, rather than higher, subsequent NPLs in the main dynamic model, while internet banking does not exhibit a robust direct effect. The paper should therefore be framed around the conditional and ambiguous relationship between digital deepening and banking stability, rather than around a linear risk-amplification narrative.

The main limitation is data coverage. The country-year sample is defensible for two-way fixed effects estimation, but it is not large enough to support system GMM as a primary estimator. The household-debt and app-commerce series are the binding constraints, and the housing-stress control further reduces the sample. In addition, the available variables are aggregate and do not distinguish between unsecured consumer credit, mortgage credit, credit-card debt, buy-now-pay-later products, or digital credit originated through fintech platforms. These limitations restrict the ability to identify a precise micro-level credit mechanism.

Despite these limitations, the results are useful for financial risk management because they caution against treating digital commerce as mechanically destabilizing. Digitalization can coexist with lower observed banking fragility when it proxies for formalization, better payment infrastructure, or broader economic maturity. At the same time, the positive household-debt coefficient indicates that household balance-sheet expansion remains a relevant warning signal for banking systems. Regulators and banks should therefore monitor digital adoption together with credit quality, household repayment capacity, and underwriting standards rather than relying on digital usage indicators alone.

5. Conclusions

This study examined whether digital commerce deepening, digital banking use, and household leverage are associated with banking fragility in a cross-country country-year panel. The model was adapted to the actual availability of Passport data and estimated using a dynamic two-way fixed effects specification with country and year effects. The final baseline sample included 171 country-year observations from 16 countries.

The strongest conclusion is that banking fragility is persistent. Lagged nonperforming loans are positive and highly significant across all dynamic specifications. The second conclusion is that household debt has a positive, moderate, and partially robust association with subsequent banking fragility. This supports the relevance of household leverage as a macro-financial risk channel.

The third conclusion is that digitalization does not have a simple destabilizing effect. App-based mobile commerce is negatively associated with banking fragility in the main dynamic model, while internet banking does not show a robust direct effect. These findings require a more cautious theoretical interpretation: digital commerce may reflect efficiency, formalization, and digital-market maturity, not only easier consumption and potential overborrowing.

The fourth conclusion is that government effectiveness does not significantly moderate the household debt-NPL relationship in the current sample. This does not rule out institutional effects, but it suggests that broader governance indicators may not be sufficiently precise to capture financial-supervisory mechanisms.

For publication, the paper should present the baseline model as dynamic TWFE rather than system GMM. System GMM can be mentioned only as an exploratory robustness option because the cross-sectional dimension of the matched sample is small. Future research should expand the model with credit-to-GDP ratios, disposable income, interest rates, household debt-service ratios, digital credit indicators, bank capital, and macroprudential-policy measures. A stronger test of the mechanism would also require separating consumer credit, mortgage debt, credit-card debt, and fintech-originated credit.