The non-linear relationship between ESG performance and bank stability in the digital era: new evidence from a regime-switching approach

The paper examines how changes in ESG scores affect bank stability in the context of increasing digitization of European banking. Motivated by post-crisis attention to sustainability and the rapid diffusion of ICT, the study addresses whether the impact of digitization on bank stability is linked to ESG activities and whether bank stability depends on the level of ESG scores. The authors posit that the ESG–stability relationship may be non-linear and contingent on ESG levels, and that digitization (ICT diffusion and technological endowment) interacts with ESG to influence stability. Focusing on European banks (15 countries, 2005–2022), the study aims to clarify mixed theoretical predictions (shareholder vs stakeholder views) and incorporate the digital environment ignored by most prior work.

Theoretical debates on digitization and bank stability contrast the innovation–stability hypothesis (ICT enhances performance, information dissemination, and stability) with the innovation–fragility hypothesis (financial innovation increases risk-taking and systemic fragility). Empirical evidence is mixed and often assumes linearity; few studies measure digitization via ICT diffusion/endowment. The ESG–stability literature contrasts the classical (overinvestment, agency costs, reduced value and stability) and stakeholder (moral and reputational capital, transparency, lower risk) perspectives. Many studies find ESG reduces bank risk, with governance and social pillars often salient, especially during crises. Recent work suggests possible non-linear relationships (e.g., U-shaped effects between social responsibility and performance), but few empirically test non-linearity for banks; Lupu et al. (2022) is a notable exception. The authors formulate hypotheses: H1-H2 (ICT diffusion/endowment positively affect stability), H3-H4 (ESG negatively vs positively affects stability), H5 (ESG–stability relationship is non-linear), and H6 (the effect of digitization on stability depends on ESG levels).

Design: Panel econometric study of European banks operating in 15 countries (France, Germany, Belgium, Denmark, Czechia, Netherlands, Finland, Spain, Italy, Poland, Portugal, Norway, United Kingdom, Switzerland, Sweden) over 2005–2022. The sample selection relies on data availability and countries ranked within the top 30 of the 2022 Global Innovation Index. Variables: Dependent variable is banking Z-Score (BZ_Score) as a stability proxy. Key explanatory variables for digitization: IUI (internet users, % of population), MCS (mobile cellular subscriptions per 100 people), ATM (ATMs per 100,000 adults). ESG is the Refinitiv ESG score for each bank. Controls: NIM (net interest margin), ROA (return on assets), CIR (cost-to-income ratio), GDPG (GDP growth). Baseline model (linear): BZ_Score_it = β0 + β1 IUI_it + β2 MCS_it + β3 ATM_it + β4 ESG_it + β5 NIM_it + β6 ROA_it + β7 CIR_it + β8 GDPG_it + bank and time effects + ε_it. Estimation approach: Panel Smooth Transition AutoRegressive (PSTAR) model to allow regime switching and non-linear effects driven by ESG. The PSTAR specification includes lagged BZ_Score as an autoregressive term and a logistic transition function G(q_it; γ; c) with ESG as the transition variable. The model permits gradual transitions between regimes defined by ESG thresholds. Diagnostics and preprocessing: Descriptive statistics indicate non-normality and serial correlation. Multicollinearity assessed via correlations and VIF (≈1.24) shows no serious issues. First- and second-generation unit root tests (LLC, IPS, Hadri; Pesaran 2003, 2007) indicate variables are I(1); break unit root tests (Karavias–Tzavalis) identify structural breaks around major crises (2008–2012; COVID-19). Cross-sectional dependence and slope heterogeneity are confirmed by multiple tests (Friedman, Breusch–Pagan, Frees, Pesaran). Panel cointegration is supported by Kao, Pedroni, Westerlund, and Persyn–Westerlund tests. Heteroscedasticity and serial autocorrelation are present; the PSTAR order p=1 is selected (Chi-square 1751.124, p=0.000). Non-linearity tests: Wald, Fisher, and LR tests reject linearity, supporting a PSTAR with order m=1 and r=2 thresholds. Optimal ESG thresholds: c1 = 33.314 and c2 = 63.348; transition parameters γ1 = 18.557 and γ2 = 975.09; RSS ≈ 7.6e+7; AIC = 8.108; BIC = 8.183. PSTAR estimation: Three ESG regimes are defined by ESG_{t−1} ≤ 33.314 (Regime 1), 33.314 < ESG_{t−1} < 63.348 (Regime 2), and ESG_{t−1} ≥ 63.348 (Regime 3). Coefficients for digitization, ESG, and controls are estimated per regime (see Key Findings). Robustness: Panel Granger causality tests (Dumitrescu–Hurlin, Juodis et al. 2021) assess directions of causality among variables, with bidirectional causality documented between ESG and BZ_Score and between several digitization proxies and BZ_Score.

- Non-linearity and regimes: Linearity is rejected; the PSTAR identifies three ESG regimes with thresholds at ESG 33.314 and 63.348, confirming H5. - ESG effects by regime (Table 11): - Regime 1 (ESG ≤ 33.314; 385 obs): A 1% increase in ESG decreases BZ_Score by 0.245% (t = -2.123), indicating reduced stability. - Regime 2 (33.314 < ESG < 63.348; 970 obs): A 1% increase in ESG decreases BZ_Score by 0.171% (t = -2.217), still destabilizing but less so. - Regime 3 (ESG ≥ 63.348; 499 obs): A 1% increase in ESG increases BZ_Score by 0.173% (t = 2.236), enhancing stability. Interpretation: Low-to-moderate ESG aligns with the classical overinvestment view (H3 supported in Regimes 1–2), while high ESG aligns with stakeholder theory (H4 supported in Regime 3). - Digitization effects: - ICT diffusion (IUI): Positive effect on stability in Regimes 1 and 3 (e.g., Regime 1: 0.333, t=3.459; Regime 3: 0.102, t=2.783). - Mobile subscriptions (MCS): Positive in Regimes 2 and 3 (Regime 2: 0.157, t=2.372; Regime 3: 0.153, t=2.363). - Technological endowment (ATM): Negative in low ESG (Regime 1: -0.294, t=-2.813), positive in higher ESG regimes (Regime 2: 0.422, t=2.146; Regime 3: 0.427, t=2.178). Thus H1 is supported; H2 is rejected in Regime 1 but supported in Regimes 2–3, evidencing H6 (digitization’s impact depends on ESG level). - Controls: ROA is negatively associated with stability in Regime 1 and positively in Regimes 2–3; CIR negative in Regime 2 and positive in Regime 3; GDPG positive in Regimes 2–3. - Causality: Dumitrescu–Hurlin tests show bidirectional causality between ESG and BZ_Score; between IUI/MCS/ATM and BZ_Score; unidirectional causality from GDPG and CIR to BZ_Score. Juodis et al. (2021) tests corroborate causality from IUI, MCS, ESG, NIM, ROA, and CIR to BZ_Score, with ATM and GDPG not causal in that specification. - Policy-relevant thresholds: ESG ≥ 63.348 is the zone where ESG strengthens stability; below ≈33.314, tech endowment (ATM) increases fragility. - Sample/context: 15 European countries, 2005–2022, encompassing the GFC, European sovereign debt crisis, and COVID-19 period; structural breaks are documented.

The findings answer the core question by demonstrating that the ESG–stability nexus is non-linear and regime-dependent in a digital banking environment. At low-to-moderate ESG levels, ESG expenditures behave like overinvestment or agency costs that depress stability; beyond a high threshold (≈63.35), reputational and stakeholder capital dominate, translating ESG investments into improved stability. This reconciles divergent theories: both classical and stakeholder perspectives hold under different ESG regimes. Digitization generally supports stability via ICT diffusion (IUI, MCS), but technological endowment (ATM) can heighten fragility when ESG is weak, suggesting that ESG commitment conditions the benefits of digitization. Thus, ESG strengthens the stabilizing effect of digital infrastructure and mitigates its potential fragility, aligning with the hypothesis that digitization’s impact depends on ESG engagement. These results underscore the importance of integrating sustainability into digital strategies and adapting regulatory frameworks to both digitization and sustainability objectives.

The study contributes by (i) documenting a non-linear, regime-dependent relationship between ESG performance and bank stability; (ii) integrating the digital context via ICT diffusion and technological endowment; and (iii) applying a PSTAR regime-switching approach to estimate ESG thresholds relevant for stability. Practically, banks should pursue digitization while elevating ESG performance to at least the identified stability-enhancing threshold, reconfiguring business models and organizational sustainability. Policymakers should ensure supervisory and regulatory frameworks keep pace with digitization and sustainability, strengthening legal infrastructure to support both. Future research should decompose ESG into its pillars to gauge their differential roles, expand measures of digital transformation (beyond ATMs, internet, and mobile proxies), and test generalizability across broader bank and country samples.

- ESG not decomposed into environmental, social, and governance pillars, limiting insights on pillar-specific effects. - Digitization captured by proxy variables (IUI, MCS, ATM), not direct measures of banks’ internal digital investments and capabilities. - Sample coverage and bank selection depend on data availability across 15 European countries, limiting direct comparability with other regions and potentially affecting external validity. - Presence of structural breaks and cross-sectional dependence managed econometrically, but residual heterogeneity may remain. - ATM as a technological endowment proxy may conflate access with legacy infrastructure, especially amid digital channel migration.