Do ESG strategies enhance bank stability during financial turmoil? Evidence from Europe

Figures & data

Table 1. Variable definitions and expected relationships vs bank stability.

Variable name	Definition	Source	Expected sign
DEPENDENT VARIABLE
DTD	The Merton’s Distance to Default, computed as the inverse standard normal cumulative distribution function of the probability of default (PD).	Bloomberg database (Authors’ calculation)	/
TARGET VARIABLES
ESG	Environmental Social Governance (ESG) score is an overall company score based on the self-reported information in the Environmental (ENV), Social (SOC) and corporate Governance (GOV) pillars. See Table A.1 in Appendix for further details.	Thomson Reuters’ Refinitiv database (Authors’ calculation)	Positive
ENV	Environmental score is an overall company score based on the weighted average of self-reported information in the Resource Use score, Emissions score and Environmental Innovation score. See Table A.1 in Appendix for further details.	Thomson Reuters’ Refinitiv database (Authors’ calculation)	Positive
SOC	Social score is an overall company score based on the weighted average of self-reported information in the Workforce score, Human rights score, Community score and Product Responsibility score. See Table A.1 in Appendix for further details.		Positive
GOV	Governance score is an overall company score based on the weighted average of self-reported information in the Management score, Shareholders score, Corporate Social Responsibility (CSR) Strategy score. See Table A.1 in Appendix for further details.		Positive
D_SHOCK	Dummy for the publication of the Non-financial Reporting Directive 2014/95/EU: equals 1 for years 2015–2017 and 0 otherwise.	Authors’ calculation	Negative
D_TREATED	Dummy equal to 1 for banks above mean ESG values in the year of the shock (2014) and 0 otherwise.	Authors’ calculation	Negative
BANK CONTROLS
SIZE	Natural logarithm of total assets.	Thomson Reuters database (Authors’ calculation)	Positive/Negative
ETA	Equity to total assets.		Positive
LLR_GL	Loan loss reserves to gross loans.		Negative
CIR	Cost-income ratio.		Negative
ROAE	Return on average Equity.		Positive
CASH_TA	Cash to total assets		Positive
DIV	Non-interest income to net operating revenue.		Positive/Negative
D_VLTRO	Dummy equals to 1 for the period in which the European Central Bank (ECB) provided liquidity to a specific bank through the Very Long-Term Refinancing Operations (VLTRO) and 0 otherwise.	Bloomberg database	Positive/Negative
COUNTRY CONTROLS
GDP_GRW	Annual Gross Domestic Product (GDP) growth rate.	World Bank Financial Development database	Positive
HHI	The Herfindahl-Hirschman Index calculated as the sum of the squared market share value (in terms of total asset) of all banks in the country.	Thomson Reuters database (Authors’ calculation)	Positive/Negative
CAP_REG	The strength of capital regulation at the country level.	World Bank (Barth, Caprio, and Levine 2008)	Positive
SUP_PWR	The strength of supervision at the country level.		Positive
D_CRISIS	Dummy equals to 1 for years 2008–2012 and 0 otherwise (2005–2007 and 2013–2017).	Authors’ calculation	Negative

Notes: This table describes the variables used in our model and summarizes their hypothesized relationships with our dependent variables (bank stability).

Table 2. Summary statistics.

	Crisis period (I)		Non-crisis periods (II)
Variables	Mean (Std. Dev.)	Min.–Max.	Mean (Std. Dev.)	Min.–Max.	Difference in means (I)–(II)
Dependent variable
DTD	2.688	0.940–4.689	2.898	0.801–3.897	−0.209***
	(0.535)		(0.488)
Target variables
ESG	0.590	0.166–0.913	0.611	0.159–0.931	−0.021**
	(0.194)		(0.187)
ENV	0.651	0.138–0.966	0.668	0.098–0.973	−0.017**
	(0.255)		(0.241)
SOC	0.594	0.082–0.980	0.614	0.077–0.971	−0.019**
	(0.222)		(0.210)
GOV	0.516	0.084–0.952	0.543	0.044–0.916	−0.026**
	(0.220)		(0.211)
Control variables
SIZE	9.069	3.263–14.763	9.052	2.688–14.733	0.017
	(2.428)		(2.370)
ETA	0.093	0.018–0.449	0.100	0.018–0.449	−0.007**
	(0.065)		(0.059)
LLR_GL	0.038	0–0.261	0.046	0–0.317	−0.008***
	(0.038)		(0.060)
CIR	0.694	0.273–1.619	0.650	0.273–1.619	0.044***
	(0.236)		(0.188)
ROAE	0.045	−0.690–0.520	0.080	−0.689–0.495	−0.035***
	(0.147)		(0.125)
CASH_TA	0.116	0–0.636	0.121	0–0.636	−0.005
	(0.115)		(0.114)
DIV	0.098	−0.058–0.802	0.109	−0.058–0.802	−0.011*
	(0.159)		(0.154)
GDP_GRW	0.002	−0.151–0.083	0.024	−0.098–0.255	−0.022***
	(0.032)		(0.023)
HHI	0.336	0–1.000	0.329	0–1.000	0.007
	(0.232)		(0.207)
CAP_REG	6.479	3–9	6.493	3–9	−0.014
	(1.872)		(1.697)
SUP_PWR	9.911	5–14.5	10.08	5–14.5	−0.169**
	(2.110)		(2.093)

Notes: This table reports the summary statistics (mean, standard deviation, minimum and maximum values) of our variables in (I) crisis years (2008–2012) and (II) non-crisis periods (2005–2007 and 2013–2017). Variable definitions are provided in Table 1. All control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. ***, **, and * denote statistical significance at 1%, 5% and 10% levels for test of differences in means between crisis-period (I) and Non-crisis periods (II).

Table 3. Baseline model.

	DTD
	ESG score	ESG score components
Variables	(I)	(II)	(III)	(IV)
DTD (−1)	0.669***	0.669***	0.662***	0.673***
	(0.122)	(0.134)	(0.124)	(0.110)
ESG (−1)	−0.120
	(0.146)
ESG (−1)*D_CRISIS	0.286***
	(0.085)
ENV (−1)		−0.104
		(0.223)
ENV (−1)*D_CRISIS		0.217***
		(0.075)
SOC (−1)			−0.203
			(0.135)
SOC (−1)*D_CRISIS			0.319***
			(0.080)
GOV (−1)				−0.031
				(0.128)
GOV (−1)*D_CRISIS				0.296***
				(0.094)
D_CRISIS	−0.250***	−0.226***	−0.275***	−0.263***
	(0.064)	(0.081)	(0.067)	(0.072)
SIZE (−1)	−0.006	−0.006	−0.004	−0.010
	(0.013)	(0.012)	(0.012)	(0.009)
ETA (−1)	2.989**	3.523**	2.480*	2.112*
	(1.392)	(1.657)	(1.385)	(1.218)
LLR_GL (−1)	−2.344**	−2.454**	−2.323**	−1.926*
	(1.070)	(1.142)	(1.037)	(1.036)
CIR (−1)	0.057	0.079	0.038	0.022
	(0.178)	(0.168)	(0.179)	(0.176)
ROAE (−1)	−0.176	−0.206	−0.188	−0.193
	(0.256)	(0.241)	(0.250)	(0.254)
CASH_TA (−1)	0.461	0.392	0.620*	0.648
	(0.534)	(0.534)	(0.351)	(0.546)
DIV (−1)	−0.216*	−0.220	−0.211*	−0.253**
	(0.117)	(0.139)	(0.113)	(0.104)
D_VLTRO	−0.014	−0.016	−0.013	−0.007
	(0.058)	(0.059)	(0.056)	(0.067)
GDP_GRW (−1)	−0.941	−1.046	−0.743	−0.996
	(0.842)	(1.246)	(0.799)	(1.194)
HHI (−1)	0.047	0.034	0.014	0.054
	(0.101)	(0.105)	(0.095)	(0.089)
CAP_REG (−1)	0.010	0.012	0.013	0.008
	(0.013)	(0.013)	(0.012)	(0.011)
SUP_PWR (−1)	−0.021*	−0.019**	−0.021*	−0.020**
	(0.012)	(0.007)	(0.011)	(0.010)
Bank FE	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes
N. of obs.	439	439	439	439
Hansen p-value	0.894	0.893	0.896	0.896
AR(2)	0.394	0.422	0.367	0.430

Notes: This table reports the estimates of the two-step system GMM model during the period 2005–2017. The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. The variables of interest are: the ESG score – see model (I); and its three components (ENV, SOC and GOV score) – see models (II), (III), and (IV), respectively. Each target variable is interacted with the dummy crisis (D_CRISIS). Variable definitions are provided in Table 1. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 4. Estimation results for disentangled ESG score components.

	DTD
	ENV	SOC	GOV
Variables	(I)	(II)	(III)
DTD (−1)	0.720***	0.680***	0.663***
	(0.140)	(0.126)	(0.107)
Resource use score (−1)	0.283
	(0.337)
Resource use score (−1)*D_CRISIS	−0.537
	(1.431)
Emissions score (−1)	0.088
	(0.187)
Emissions score (−1)*D_CRISIS	−0.563
	(0.583)
Environmental Innovation score (−1)	−0.119
	(0.144)
Environmental Innovation score (−1)*D_CRISIS	0.294***
	(0.089)
Workforce score (−1)		−0.154
		(0.121)
Workforce score (−1)*D_CRISIS		0.242**
		(0.109)
Human Rights score (−1)		0.137
		(0.191)
Human Rights score (−1)*D_CRISIS		−0.481
		(0.551)
Community score (−1)		−0.009
		(0.094)
Community score (−1)*D_CRISIS		−0.053
		(0.232)
Product Responsibility score (−1)		−0.089
		(0.133)
Product Responsibility score (−1)*D_CRISIS		0.372*
		(0.221)
Management score (−1)			0.058
			(0.099)
Management score (−1)*D_CRISIS			−0.075
			(0.201)
Shareholders score (−1)			−0.041
			(0.082)
Shareholders score (−1)*D_CRISIS			0.220**
			(0.098)
CSR Strategy score (−1)			0.191
			(0.181)
CSR Strategy score (−1)*D_CRISIS			−0.345
			(0.327)
D_CRISIS	−0.252***	−0.243***	−0.222***
	(0.074)	(0.084)	(0.068)
Control variables (−1)	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes
N. of obs.	439	439	439
Hansen p-value	0.895	0.899	0.899
AR(2)	0.333	0.469	0.579

Notes: This table reports the estimates of the two-step system GMM model obtained disentangling the ESG score components. Model (I) is referred to the ENV component and includes the following categories: Resource use score, Emissions score, and Environmental Innovation score. Model (II) is referred to the SOC component and includes the following categories: Workforce score, Human Rights score, Community score, and Product Responsibility score. Model (III) is referred to the GOV component and includes the following categories: Management score, Shareholders score, and CSR Strategy score. Each target variable is interacted with the dummy crisis (D_CRISIS). The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. All bank and country controls are included. Variable definitions are provided in Tables 1 and A.1 (in Appendix). The sample period observed is 2005–2017. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 5. Effects of the 2014 EU non-financial reporting directive.

Panel A – Logit model: identifying propensity scores
Variables	D_HIGH_ESG
SIZE (−1)	0.0812
	(0.0502)
ETA (−1)	−13.59***
	(2.810)
LLR_GL (−1)	−3.202**
	(1.605)
CIR (−1)	−0.104
	(0.398)
ROAE (−1)	21.24**
	(9.509)
CASH_TA (−1)	−2.029*
	(1.135)
DIV (−1)	−1.237
	(0.866)
Bank FE	Yes
Cluster	Yes
SE Bank	Yes
N. of obs.	326
Panel B – Univariate statistics: effectiveness of matching
			Difference
Variables	Treated (T)	Control (C)	(T-C)	P-value
SIZE	10.496	10.559	−0.063	0.590
ETA	0.098	0.098	0	0.900
LLR_GL	0.056	0.056	0	0.298
CIR	0.652	0.641	0.011	0.380
ROAE	0.001	−0.002	0	0.200
CASH_TA	0.100	0.100	0	0.100
DIV	0.098	0.088	0.010	0.190
Panel C – Difference in difference (DID) estimation
	DTD
Variables	(I)	(II)
D_SHOCK	−0.119	−0.186**
	(0.094)	(0.085)
D_TREATED	−0.121	−2.146
	(0.121)	(1.673)
D_SHOCK*D_TREATED	0.243*	0.203**
	(0.126)	(0.137)
Control variables (−1)	Yes	Yes
Bank FE	No	Yes
Cluster SE Bank	Yes	Yes
N. of obs.	159	159

Notes: This table provides the results of our differences-in-differences strategy based on the Directive 2014/95/EU. Firstly (Panel A) we employ a propensity score matching (PSM) procedure (Rosenbaum and Rubin 1983) to reduce the potential bias arising from heterogeneity of treated (T) and control (C) groups, employing all non-binary bank-level controls (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV). The dependent variable (D_HIGH_ESG) equals 1 for banks above average values of ESG scores (treated) and 0 otherwise. Panel B provides the univariate statistics on the effectiveness of the PSM procedure. Panel C shows the results of the DID estimation for both an OLS (I) and a panel data estimation (II). Finally, Panel D shows the DTD difference between treated and control entities before and after the Directive. The dependent variable is the one-year Merton’s Distance to Default (DTD) as a proxy for stability. Target variables are: D_SHOCK, equal to 1 for years 2015–2017 (after the shock, i.e. the publication of the Directive) and 0 otherwise; D_TREATED, that takes value of 1 for banks above mean values of ESG scores in the year of the shock (2014), and 0 otherwise; the interaction term D_SHOCK*D_TREATED. Variable definitions are provided in Table 1. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% level of each tail. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 6. Accounting for ESG disclosure duration and ESG score level.

	DTD
	Below mean ESG score	Above mean ESG score	N. of years of ESG score disclosure
Variables	(I)		(II)
DTD (−1)	0.691***	0.771***	0.652***
	(0.106)	(0.165)	(0.078)
ESG (−1)	−0.147	−0.067
	(0.341)	(0.179)
ESG (1)*D_CRISIS	0.408**	0.310***
	(0.172)	(0.089)
ESG_Nyears (−1)			−0.025
			(0.020)
ESG_Nyears (−1)*D_CRISIS			0.021***
			(0.005)
D_CRISIS	−0.189	−0.265***	−0.303**
	(0.122)	(0.079)	(0.063)
Control variables (−1)	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes
N. of obs.	156	288	439
Hansen p-value	1.000	0.851	0.921
AR(2)	0.157	0.908	0.495

Notes: This table reports the estimates of the two-step system GMM model obtained using as target variable in model (I) the ESG score level (below vs above the sample mean of ESG score); and in model (II) the number of years of the bank’s ESG score disclosure (ESG_Nyears). Each target variable is interacted with the dummy crisis (D_CRISIS). The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. All bank and country control variables are included in the model. Variable definitions are provided in Table 1. The sample period observed is 2005–2017. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 7. Heckman two-step estimation results.

Panel A. First step of the Heckman two-step model
Variables	D_ESG
D_CRISIS	0.018
	(0.210)
SIZE (−1)	0.387***
	(0.085)
ETA (−1)	−7.339**
	−3.599
LLR_GL (−1)	−2.682
	−2.852
CIR (−1)	0.252
	(0.598)
ROAE (−1)	0.500
	−1.305
CASH_TA (−1)	−6.028***
	−1.777
DIV (−1)	−0.735
	(0.956)
D_VLTRO	0.369
	(0.454)
GDP_GRW (−1)	−9.100***
	−2.721
HHI (−1)	−0.486
	(0.618)
CAP_REG (−1)	0.128**
	(0.057)
SUP_PWR (−1)	−0.113*
	(0.068)
Bank FE	Yes
Cluster SE Bank	Yes
N. of obs.	608
Panel B. Second step of the Heckman two-step model
	DTD
	ESG score	ESG score components
Variables	(I)	(II)	(III)	(IV)
ESG (−1)	−0.434
	(0.284)
ESG (−1)*D_CRISIS	0.193**
	(0.098)
ENV (−1)		0.088
		(0.194)
ENV (−1)*D_CRISIS		0.111*
		(0.065)
SOC (−1)			−0.438
			(0.285)
SOC (−1)*D_CRISIS			0.159*
			(0.092)
GOV (−1)				0.025
				(0.158)
GOV (−1)*D_CRISIS				0.245***
				(0.087)
D_CRISIS	−0.294***	−0.244***	−0.257***	−0.288***
	(0.066)	(0.049)	(0.064)	(0.057)
Control variables (−1)	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes
N. of obs.	608	608	608	608
IMR	0.410	0.129	0.395	0.347
	(0.279)	(0.168)	(0.270)	(0.237)

Notes: This table reports the results obtained from the Heckman two-step model over the period 2005–2017, with Panel A showing the first step and Panel B the second one. The first step estimates the decision equation using a multinomial probit model, whose parameters are used to calculate the Inverse Mills Ratio (IMR). In this setting the dependent variables are dummies (D_ESG) equal to 1 from the year in which a bank of our sample started to be involved in ESG practices; and 0 in the previous years. The second step estimates the stability regression with the Inverse Mills Ratio (IMR) generated by the first step. The dependent variable used in the second step is the one-year Merton’s Distance to Default (DTD) which measures bank stability. The variables of interest in Panel B are: the ESG score – see model (I); and its three components (ENV, SOC and GOV score) – see models (II), (III), and (IV), respectively. Each target variable is interacted with the dummy crisis (D_CRISIS). Variable definitions are provided in Table 1. All non-binary independent variables are lagged by one year with respect to the dependent variable. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 8. IV 2SLS estimation results.

Panel A – First stage of the IV 2SLS model
	ESG	ENV	SOC	GOV
Variables	(I)	(II)	(III)	(IV)
Political_orient (−1)	0.063**	0.059*	0.055*	0.076***
	(0.025)	(0.033)	(0.031)	(0.027)
D_CRISIS	(−0.021)	(−0.019)	(−0.025)	(−0.021)
	(0.017)	(0.026)	(0.022)	(0.020)
Control variables (−1)	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes
N. of obs.	547	547	547	547
F-Cragg Donald test	474.0***	2851.32***	400.36***	468.27***
Panel B – Second stage of the IV 2SLS model
	DTD
	ESG score		ESG score components
Variables	(I)	(II)	(III)	(IV)
ESG (−1)	−0.077
	(0.195)
ESG (−1)*D_CRISIS	0.183**
	(0.091)
ENV (−1)		1.538
		(1.264)
ENV (−1)*D_CRISIS		0.132*
		(0.078)
SOC (−1)				−0.094
				(0.179)
SOC (−1)*D_CRISIS				0.219**
				(0.093)
GOV (−1)				0.166
				(0.173)
GOV (−1)*D_CRISIS				0.183*
				(0.094)
D_CRISIS	−0.313***	−0.340***	−0.328***	−0.292***
	(0.052)	(0.062)	(0.053)	(0.050)
Control variables (−1)	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes
N. of obs.	445	445	445	445
Sargan p-value	0.100	0.101	0.931	0.414

Notes: This table reports the results of the instrumental variables (IV) two-stage least squares (2SLS) estimator model over the period 2005–2017. In Panel A we report the results of the first stage of the IV 2SLS estimator, where we instrument each of our target variables (the ESG score and its three components) with country political orientation (Political_orient) (Cheung, 2016; Hasan et al. 2018), showing their correlations and goodness (the Cragg-Donald weak instruments F-test statistics) with reference to the original target variable being instrumented (ESG, ENV, SOC and GOV scores). The dependent variables for the first stage are: ESG, ENV, SOC, and GOV scores in columns (I), (II), (III), (IV), respectively. Panel B provides the results of the second stage of the IV 2SLS where we replace the potentially endogenous interest variables with the results of the first stage, including the Sargan test (Sargan p-value) of over-identifying restrictions. The dependent variable for the second stage is the one-year Merton’s Distance to Default (DTD) which measures bank stability. Variable definitions are provided in Table 1. All bank and country control variables are included in the model. All non-binary independent variables are lagged by one year with respect to the dependent variable. Accounting-based control variables (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 9. Baseline model with an alternative bank stability measure.

Panel A: The Z-score
	Z-score
	ESG score	ESG score components
Variables	(I)	(II)	(III)	(IV)
Z-score (−1)	0.849***	0.825***	0.827***	0.800***
	(0.078)	(0.088)	(0.091)	(0.088)
ESG (−1)	0.046
	(0.392)
ESG (−1)*D_CRISIS	0.791***
	(0.284)
ENV (−1)		0.090
		(0.465)
ENV (−1)*D_CRISIS		0.721***
		(0.235)
SOC (−1)			0.065
			(0.577)
SOC (−1)*D_CRISIS			0.782***
			(0.258)
GOV (−1)				0.081
				(0.365)
GOV (−1)*D_CRISIS				0.824***
				(0.262)
D_CRISIS	−0.680***	−0.668***	−0.658***	−0.649***
	(0.202)	(0.199)	(0.202)	(0.187)
Control variables (−1)	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes
N. of obs.	469	469	469	469
Hansen p-value	0.867	0.866	0.881	0.878
AR(2)	0.543	0.537	0.595	0.433
Panel B: Z-score components
	STD_ROAA	ROAA	ETA
Variables	(I)	(II)	(III)
Dependent Variable (−1)	0.721***	0.347***	3.089
	(0.037)	(0.073)	(4.904)
ESG (−1)	0.001	−0.001	−0.007
	(0.001)	(0.001)	(0.049)
ESG (−1) * D_CRISIS	−0.001**	0.002***	0.012
	(0.0003)	(0.001)	(0.037)
D_CRISIS	−0.312***	−0.288***	−0.292***
	(0.093)	(0.093)	(0.081)
Control variables (−1)	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes
N. of obs.	481	504	504
Hansen p-value	0.959	0.999	0.880
AR(2)	0.290	0.338	0.949

Notes: This table reports the estimates of the two-step system GMM model during the period 2005–2017 substituting the market-based bank stability measure (DTD) with an accounting-based widely used alternative (Z-Score) (Panel A) and its components (Panel B). It is calculated as the sum of equity to total assets (ETA) and return on average assets (ROAA), scaled by the three–year standard deviation of ROAA (STD_ROAA). Since the Z-score is highly skewed, we use its natural logarithm (Laeven and Levine 2009; Liu et al., 2013). Data for its computation is collected from Thomson Reuters. The variables of interest are the ESG score (see model (I)), and its three components (ENV, SOC and GOV score) (see models (II), (III), and (IV)), respectively. Each target variable is interacted with the dummy crisis (D_CRISIS). All bank and country control variables are included in the model. Variable definitions are provided in Table 1 and A1 (in Appendix). The sample period observed is 2005–2017. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 10. Baseline model with an alternative ESG score definition.

		DTD
	BESG score	BESG score components
Variables	(I)	(II)	(III)	(IV)
DTD (−1)	0.593***	0.624***	0.633***	0.614***
	(0.102)	(0.115)	(0.112)	(0.116)
BESG (−1)	−0.026
	(0.155)
BESG (−1)*D_CRISIS	0.397***
	(0.154)
BENV (−1)		−0.198
		(0.164)
BENV (−1)*D_CRISIS		0.481***
		(0.170)
BSOC (−1)			−0.195
			(0.179)
BSOC (−1)*D_CRISIS			0.368***
			(0.141)
BGOV (−1)				−0.031
				(0.180)
BGOV (−1)*D_CRISIS				0.354***
				(0.115)
D_CRISIS	−0.268***	−0.276***	−0.275***	−0.280***
	(0.053)	(0.056)	(0.058)	(0.056)
Control variables (−1)	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes
N. of obs.	316	316	316	316
Hansen p-value	0.983	0.991	0.987	0.980
AR(2)	0.509	0.704	0.747	0.704

Notes: This table reports the estimates of the two-step system GMM model obtained using an alternative definition of ESG score provided by Bloomberg database. In model (I) BESG is the ESG score that captures the transparency level of information disclosed about sustainable practices on Environmental, Social and Governance aspects. In model (II) BENV consists of corporate rating about the level of disclosed information on the following issues: carbon emission, climate change effect, pollution waste disposal, renewable energy, and resource depletion. In model (III) BSOC consists of corporate rating about the level of disclosed information on the following issues: supply chain, discrimination, political contributions, diversity, human rights, and community relations. In model (IV) BGOV consists of corporate rating about the level of disclosed information on the following issues: cumulative voting, executive compensation, shareholder’s right, takeover defence, staggered boards, and independent directors. Each target variable is interacted with the dummy crisis (D_CRISIS). The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. All bank and country control variables are included in the model. Variable definitions are provided in Table 1. The sample period observed is 2005–2017. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 11. Baseline model for different sub-samples.

	DTD
Variables	EBA Banks (I)	No-EBA Banks (II)	Market-oriented financial systems (III)	Bank-oriented financial systems (IV)	Below mean GDP pre-capita (V)	Above mean GDP per-capita (VI)
DTD (−1)	0.892***	0.064	1.533	0.702***	0.625***	0.601***
	(0.165)	(0.449)	(1.246)	(0.125)	(0.178)	(0.177)
ESG (−1)	−0.086	−0.643	−2.943	−0.031	−0.127	−0.341
	(0.197)	(0.909)	(8.437)	(0.244)	(0.309)	(0.626)
ESG (−1)*D_CRISIS	0.415***	0.066	−5.203	0.277***	0.222*	0.375***
	(0.106)	(0.339)	(6.520)	(0.101)	(0.131)	(0.138)
D_CRISIS	−0.284***	−0.220*	−0.418	−0.328***	−0.207	−0.328***
	(0.089)	(0.122)	(0.381)	(0.059)	(0.139)	(0.071)
Control variables (−1)	Yes	Yes	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes	Yes	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes	Yes	Yes
N. of obs.	299	140	113	326	253	186
Hansen p-value	0.590	0.997	0.874	0.880	0.972	0.543
AR(2)	0.666	0.523	0.885	0.960	0.122	0.132

Notes: This table reports the estimates of two-step system GMM model tested employing different sub-samples that control for: the bank size (banks subjected to the EBA stress testing 2014 exercise vs those that are not subjected); the orientation of the financial system (below vs above the sample mean of the ratio of domestic credit provided by banking sector to GDP); and the income level of the country (below vs above the sample mean of GDP per capita). The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. The target variable is the ESG score. It is interacted with the dummy crisis (D_CRISIS). All bank and country control variables are included in the model. Variable definitions are provided in Table 1. The sample period is 2005–2017. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorised at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are included in all specifications. Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 12. Alternative measure of the crisis: the country-level index of financial stress (CLIFS).

	DTD
Variables	(I)	(II)	(III)	(IV)	(V)	(VI)
DTD (−1)	0.722***	0.691***	0.682***	0.840***	0.800***	0.800***
	(0.077)	(0.051)	(0.069)	(0.072)	(0.067)	(0.067)
ESG (−1)	−0.098	−0.123	−0.0705	0.00890	0.0211	0.0211
	(0.126)	(0.077)	(0.123)	(0.112)	(0.111)	(0.111)
ESG * CLIFS (−1)	0.869**	0.674**	0.730*
	(0.417)	(0.322)	(0.407)
ESG * ΔCLIFS (−1)				0.182***	0.172***	0.172***
				(0.053)	(0.052)	(0.052)
CLIFS (−1)	−0.959***	−0.914***	−1.027***
	(0.256)	(0.160)	(0.247)
ΔCLIFS (−1)				−0.144***	−0.153***	−0.153***
				(0.033)	(0.033)	(0.033)
Control variables (−1)	Yes	Yes	Yes	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	No	Yes	Yes	No	Yes	Yes
Country FE	No	No	Yes	No	No	Yes
Cluster SE Bank	Yes	Yes	Yes	Yes	Yes	Yes
N. of obs.	368	368	368	368	368	368
Hansen p-value	0.989	0.891	0.880	0.150	0.765	0.795
AR(2)	0.401	0.827	0.817	0.467	0.525	0.535

Notes: This table reports the estimates of the two-step system GMM model during the period 2005–2017 by using the Country-Level Index of Financial Stress (CLIFS) as alternative measure of financial market turmoil. The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. The variables of interest are the ESG score and its interaction with either the level (CLIFS) or the annual variation of the index (ΔCLIFS). Variable definitions are provided in Table 1. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorised at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are always included, whereas year and country FE are either included or excluded in columns from (I) to (VI). Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table 13. Subprime vs Sovereign crisis.

	DTD
Variables	(I)	(II)	(III)
DTD (−1)	0.714***	0.726***	0.717***
	(0.140)	(0.127)	(0.084)
ESG (−1)	−0.041	−0.017	−0.016
	(0.246)	(0.178)	(0.167)
ESG (−1) * SUB_CRISIS	0.591***	0.506**	0.504**
	(0.213)	(0.225)	(0.207)
ESG (−1) * SOV_CRISIS	0.228**	0.211**	0.229***
	(0.099)	(0.094)	(0.080)
SUB_CRISIS	−0.271***	−0.269***	−0.269***
	(0.083)	(0.097)	(0.069)
SOV_CRISIS	−0.312***	−0.288***	−0.292***
	(0.093)	(0.093)	(0.081)
Control variables (−1)	Yes	Yes	Yes
Bank FE	Yes	Yes	Yes
Year FE	No	Yes	Yes
Country FE	No	No	Yes
Cluster SE Bank	Yes	Yes	Yes
N. of obs.	439	439	439
Hansen p-value	0.959	0.999	0.880
AR(2)	0.290	0.338	0.949

Notes: This table reports the estimates of the two-step system GMM model during the period 2005–2017. The dependent variable is the one-year Merton’s Distance to Default (DTD) which measures bank stability. The variables of interest are: the ESG score (see model (I), (II), and (III)), respectively, and its interaction with the following dummy variables: (1) the SUB_CRISIS which takes the value of 1 for the year 2008–2009 and 0 otherwise; (2) the SOV_CRISIS which takes the value of 1 for the years 2010–2012 and 0 otherwise. Variable definitions are provided in Table 1. All non-binary independent variables are lagged by one year with respect to the dependent variable. The control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorised at the 1% of each tail. The Hansen p-value is that of the Hansen test statistic of over-identifying restrictions, while AR(2) are the p-value of the first and second-order autocorrelation test statistic. Bank fixed-effects (FE) are always included, whereas year and country FE are either included or excluded in columns from (I) to (III). Bank clustered standard errors (SE) are reported in parentheses. The superscripts ***, **, and * denote coefficients statistically different from zero at the 1%, 5%, and 10% levels, respectively, in two-tailed tests.

Table A1. Calculation of ESG score and its components.

Pillar	Category	Category definition	Indicators in rating	Category weights	Sum of category weights	New category weights	ESG score components
Environmental	Resource Use score	Resource use category score reflects a company’s performance and capacity to reduce the use of materials, energy or water, and to find more eco-efficient solutions by improving supply chain management.	19	0.110 (=19/178)	0.3400 (=0.110+ 0.120+ 0.110)	0.3235 (=0.110/ 0.340)	ENV = Resource Use score*0.3235 + Emission score*0.3529 + Environmental Innovation score*0.3235
	Emissions score	Emission category score measures a company’s commitment and effectiveness towards reducing environmental emission in the production and operational processes.	22	0.120 (=22/178)		0.3529 (=0.120/ 0.340)
	Innovation score	Environmental innovation category score reflects a company’s capacity to reduce the environmental costs and burdens for its customers, and thereby creating new market opportunities through new environmental technologies and processes or eco-designed products.	20	0.110 (=20/178)		0.3235 (=0.110/ 0.340)
Social	Workforce score	Workforce category score measures a company’s effectiveness towards job satisfaction, healthy and safe workplace, maintaining diversity and equal opportunities, and development opportunities for its workforce.	29	0.160 (=29/178)	0.3550 (=0.160+0.045 +0.080+0.070)	0.4507 (=0.160/ 0.3550)	SOC = Workforce score*0.4507 + Human Rights score*0.1268 + Community score*0.2254 + Product Responsibility score*0.1972
	Human Rights score	Human rights category score measures a company’s effectiveness towards respecting the fundamental human rights conventions.	8	0.045 (=8/178)		0.1268 (=0.045/ 0.3550)
	Community score	Community category score measures the company’s commitment towards being a good citizen, protecting public health and respecting business ethics.	14	0.080 (=14/178)		0.2254 (=0.080/ 0.3550)
	Product Responsibility score	Product responsibility category score reflects a company’s capacity to produce quality goods and services integrating the customer’s health and safety, integrity and data privacy.	12	0.070 (=12/178)		0.1972 (=0.070/ 0.3550)	GOV = Management score*0.6230 + Shareholders score*0.2295 + CSR Strategy score*0.1475
Governance	Management score	Management category score measures a company’s commitment and effectiveness towards following best practice corporate governance principles.	34	0.190 (=34/178)	0.3050 (=0.190+0.070 +0.045)	0.6230 (=0.190/ 0.3050)
	Shareholders score	Shareholders category score measures a company’s effectiveness towards equal treatment of shareholders and the use of anti-takeover devices.	12	0.070 (=12/178)		0.2295 (=.070/ 0.3050)
	CSR Strategy score	CSR strategy category score reflects a company’s practices to communicate that it integrates the economic (financial), social and environmental dimensions into its day-to-day decision-making processes.	8	0.045 (=8/178)		0.1475 (=0.045/ 0.3050)
TOTAL			178	1.00	1.00		ESG = (ENV*0.3400)+ (SOC*0.3550)+(GOV*0.3050)

Notes: This table summarize the category scores and their weights used by Thomson Reuters’ Refinitiv database to compute the ESG score and its three components (ENV, SOC, and GOV). The ‘Category Weights’ are determined by the number of indicators (so-called ‘Indicators in Rating’) that make up each category in comparison to all indicators used in the Thomson Reuters’ Refinitiv ESG Score framework. This means that higher weight is assigned to themes that are more mature in terms of disclosure and the relative performance scores of companies is calculated with a higher degree of confidence. As a result, categories that contain multiple issues with relatively higher transparency like Management (composition, diversity, independence, committees, compensation, etc.) and companies reporting more information across these topics will have higher weight than lighter and less reported categories, such as Human Rights or CSR Strategy. The ‘Sum of Category Weights’ is obtained by summing each category weight of respective pillars. The ‘New Category Weights’ are computed by dividing category weights to the sum of category weights of respective pillar.

For further information, see the detailed description on ESG scores available at the following website: https://www.refinitiv.com/content/dam/marketing/en_us/documents/methodology/esg-scores-methodology.pdf.

Table A2. Correlation Matrix.

	Variables	1	2	3	4	5	6	7	8	9	10	11	12	13	14
1	DTD	1
2	Z-score	0.2408*	1
3	ESG	−0.1738*	0.0396	1
4	ENV	−0.2728*	−0.0124	0.8935*	1
5	Resource use	−0.0051	−0.0209	0.5447*	0.5722*	1
6	Emission	0.0264	0.0144	0.4946*	0.5812*	0.6668*	1
7	Innovation	−0.0783*	−0.0652	0.3452*	0.3701*	0.4149*	0.3716*	1
8	SOC	−0.1716*	−0.0215	0.8901*	0.7857*	0.4960*	0.4092*	0.2914*	1
9	Workforce	−0.1368*	0.0118	0.7507*	0.6657*	0.4141*	0.3624*	0.2141*	0.8709*	1
10	Human Rights	0.0052	−0.0751*	0.4465*	0.3709*	0.5011*	0.4059*	0.4064*	0.4203*	0.2718*	1
11	Community	0.0462	−0.0087	0.4635*	0.3617*	0.5095*	0.4367*	0.2642*	0.5462*	0.2997*	0.3545*	1
12	Product Resp.	−0.0288	−0.0968*	0.4350*	0.3270*	0.4780*	0.3321*	0.3446*	0.5398*	0.2996*	0.3983*	0.5117*	1
13	GOV	0.0586	0.1504*	0.6584*	0.3437*	0.2373*	0.1844*	0.1703*	0.3614*	0.2697*	0.3035*	0.2195*	0.1906*	1
14	Management	0.0221	−0.0215	0.4386*	0.1847*	0.2757*	0.1923*	0.1868*	0.2423*	0.1498*	0.3563*	0.2900*	0.2651*	0.7203*	1
15	Shareholders	0.0168	0.0277	0.0241	−0.0631	−0.1204*	−0.0769*	−0.0546	−0.0945*	−0.0412	0.0082	−0.1494*	−0.1843*	0.2553*	0.0529
16	CSR strategy	−0.0649	0.0482	0.7007*	0.6762*	0.4505*	0.4012*	0.3168*	0.6251*	0.5221*	0.3534*	0.3424*	0.3220*	0.3961*	0.2347*
17	SIZE	0.0307	−0.0397*	0.3629*	0.3432*	0.2716*	0.2377*	0.2348*	0.3689*	0.2808*	0.3228*	0.2512*	0.2570*	0.1618*	0.1882*
18	ETA	0.2156*	0.0227	−0.2635*	−0.3222*	−0.0749*	−0.0638	−0.0741*	−0.2798*	−0.2451*	−0.1083*	−0.0153	−0.1235*	−0.0112	−0.0763*
19	LLR_GL	−0.1180*	−0.3935*	−0.0459	0.0077	−0.0163	−0.0601	−0.0289	−0.0397	−0.0734	−0.0457	−0.0031	−0.0371	−0.0929*	−0.0591
20	CIR	−0.2895*	−0.2729*	0.0601	0.0965*	0.0328	−0.0087	0.0715*	0.0641	0.0136	0.0614	−0.0037	0.0621	−0.0272	0.0340
21	ROAE	0.2878*	0.1850*	0.0443	−0.0215	0.0306	0.0429	−0.0438	0.0636	0.0701*	0.0037	0.0972*	0.0225	0.0788*	0.0339
22	CASH_TA	0.1236*	−0.0808*	−0.0098	0.0062	−0.0110	0.0616	−0.0311	−0.0154	0.0258	−0.0417	−0.0390	−0.0123	−0.0181	−0.0614
23	DIV	−0.0132	−0.1160*	−0.0243	0.0124	−0.0595	−0.0407	−0.0656	0.0164	0.0160	−0.0490	−0.0458	−0.0507	−0.1045*	−0.1170*
24	GDP_GRW	0.2581*	0.1411*	−0.0737*	−0.1301*	0.0015	0.0020	0.0008	−0.0644	−0.0720*	0.0020	0.0029	0.0021	0.0316	0.0013
25	HHI	−0.0070	0.1106*	0.1616*	0.1157*	−0.0025	−0.0030	−0.0004	0.1514*	0.1330*	−0.0002	−0.0050	−0.0055	0.1349*	−0.0028
26	CAP_REG	−0.1285*	0.0461*	0.0576	0.0118	0.0054	0.0060	−0.0019	0.0326	0.0050	0.0055	0.0017	−0.0007	0.1096*	−0.0015
27	SUP_PWR	0.0949*	0.0384*	−0.1209*	−0.0744*	−0.0003	0.0007	−0.0014	−0.1412*	−0.1627*	0.0004	−0.0025	−0.0029	−0.0838*	−0.0018

Continued.

	Variables	15	16	17	18	19	20	21	22	23	24	25	26	27
15	Shareholders	1
16	CSR strategy	−0.1217*	1
17	SIZE	−0.0589	0.3738*	1
18	ETA	0.0974*	−0.2962*	−0.4311*	1
19	LLR_GL	−0.0222	−0.0961*	−0.0814*	0.2583*	1
20	CIR	−0.0470	0.0299	−0.0727*	−0.0282	0.1376*	1
21	ROAE	0.0223	0.0512	0.0520*	−0.0028	−0.1697*	−0.5334*	1
22	CASH_TA	0.0307	0.0106	−0.1691*	0.2048*	0.1338*	0.1475*	0.0129	1
23	DIV	−0.0235	0.0005	−0.0058	0.1165*	0.0195	0.1734*	−0.0646*	0.1499*	1
24	GDP_GRW	−0.0016	0.0036	0.0026	0.0594*	−0.1051*	−0.0769*	0.0005	0.0496*	0.0140	1
25	HHI	0.0056	0.0216	0.0020	0.0466*	−0.1152*	−0.1217*	0.0017	0.0233	−0.0090	0.0281	1
26	CAP_REG	−0.0021	−0.0183	−0.0001	0.0036	0.0933*	−0.0187	−0.0019	−0.0265	−0.0101	−0.0332*	0.1607*	1
27	SUP_PWR	0.0000	−0.1570*	−0.0018	0.0501*	0.0779*	0.0870*	−0.0015	0.1312*	−0.0398*	0.0800*	−0.1315*	0.1521*	1

Notes: This table shows the correlation matrix of the variables used in the empirical analysis over the period 2005–2017. Variable definitions are provided in Tables 1 and A1 (in Appendix). All control variables based on accounting data (SIZE, ETA, LLR_GL, CIR, ROAE, CASH_TA, DIV) are winsorized at the 1% of each tail. The symbol * indicates statistically significance at the 5% level.

Table A3. Sample description by country.

Country of headquarter	N. of Obs.	Percent
Austria	15	3%
Belgium	17	4%
Czech Republic	10	2%
Denmark	13	3%
Finland	11	3%
France	26	6%
Germany	8	2%
Greece	36	8%
Ireland	27	6%
Italy	46	10%
Netherlands	4	1%
Norway	11	3%
Poland	61	14%
Portugal	11	3%
Russia	23	5%
Spain	50	11%
Sweden	32	7%
Switzerland	6	1%
United Kingdom	32	7%
Total	439	100%

Notes: This table describes our sample by providing the breakdown of observations at the country level.

Figure A1. Parallel Trends. Notes: This figure illustrates the behaviour of the average one-year Merton’s Distance to Default (DTD) before and after the shock or treatment (i.e. the publication of the Non-financial Reporting Directive 2014/95/EU in October 2014) for both the treated and the control group. The treated (control) group is represented by banks above (below) the average values of ESG scores in the year of the shock (2014).

Barth, J. R., G. J. Caprio, and R. Levine. 2008. “Bank Regulation are Changing: But for Better or Worse?” World Bank Policy Research Working Paper, No. 4646.

 Google Scholar

Rosenbaum, P., and D. Rubin. 1983. “The Central Role of the Propensity Score in Observational Studies for Causal Effects.” Biometrika 70: 41–55.

 Web of Science ®Google Scholar

Hasan, I., N. Kobeissi, L. Liu, and H. Wang. 2018. “Corporate Social Responsibility and Firm Financial Performance: The Mediating Role of Productivity.” Journal of Business Ethics 149: 671–688.

 Web of Science ®Google Scholar

Laeven, L., and R. Levine. 2009. “Bank Governance, Regulation and Risk Taking.” Journal of Financial Economics 93: 259–275.

 Web of Science ®Google Scholar