The social and environmental drivers of corporate credit ratings: international evidence

Abstract

We provide evidence of the exogenous impact of environmental and social performance components on credit ratings in North America, Europe, and Asia. In particular, the product innovation dimension is clearly identified as being the dominating driver of credit ratings within the environmental performance in every subsample region. In the social performance dimension, the extent of diversity is a main driver for firms in North America and Europe, but due to cultural reasons, not in Asia. Our results show that the risk mitigation view holds for all significant corporate social or environmental performance variables, but the magnitude of impact differs regionally.

1 Introduction

We identify the single dimensions of corporate social and environmental performance which have an impact on credit ratings. Our analysis differs from earlier studies through the joint use of more sophisticated and transparent corporate social performance (CSP)¹ measures of Asset4, the identification of the affecting CSP components, the regional differentiation in an international dataset (North America, Europe, and Asia), and the use of an instrumental variables approach in conjunction with commonly employed credit risk models. It is our approach in particular that allows us to provide clearer indications of a causal relationship in terms of how CSP components impact credit ratings, as opposed to the common approaches revealing only correlational relationships.

Dorfleitner et al. (2020) find that out-of-sample prediction quality improves by more than 0.8% in their North America sample if environmental and social performance measures are integrated into an established credit risk model. However, a detailed analysis of the underlying drivers within the social and environmental performance is only available for the USA and suffers from a potential exposure to endogeneity (e.g., Oikonomou et al. 2014) or rather simplistic credit risk modeling (e.g., Attig et al. 2013). Endogeneity, in terms of the reverse causality problem, is crucial to the analysis of the relationship between CSP and credit ratings. On the one hand, CSP is commonly expected to have a positive impact on credit ratings. On the other hand, though, the opposite direction of impact is also conceivable, in the way that firms with better credit ratings save financing costs and are therefore able to increase their spending on CSP. Most studies on this topic use lagged independent variables to deal with the endogeneity problem, which is the first step, but nonetheless appears not to be insufficient. Some (e.g., Bauer et al. 2009; Jiraporn et al. 2014) estimate a two-stage least squares (2SLS) model, which is generally appropriate for reducing endogeneity, but this approach does not meet the standards of current literature on credit risk because credit ratings need to be considered as categorical, and the employed OLS estimation is unable to model this. As a consequence, an international analysis with an adequate credit risk model and a sufficient approach to identify relevant CSP aspects which have a causal impact on credit ratings is still lacking in the literature.

We fill this gap by applying the analysis to both CSP in general and its components in an international dataset including Asset4 CSP measures based on the two-stage predictor substitution (2SPS) with an established credit risk model in the second stage. Asset4 CSP measures are internationally available on a granular level, allowing us to drive our analysis consistently for North America, Europe, and Asia. The environmental performance comprises measures for emission reduction, product innovation, and resource reduction, while the social performance dimension spans the categories product responsibility, community, human rights, diversity, respectively, equal opportunities, employment quality, health, and training. Asset4 scores are, compared to other providers such as MSCI-KLD, methodologically superior and more transparent (Chatterji and Levine 2006). Concerning established credit risk models, endogeneity can be mitigated through the two-stage predictor substitution (2SPS), which is an implementation of the instrumental variable approach for nonlinear models. In the first stage, we regress the CSP scores on instruments such as the average CSP level of firms located in the same area (Jiraporn et al. 2014) and measures for so-called ‘national business systems’ (NBS) (Whitley 1999) in terms of the political, the labor, education, and the cultural system according to Ioannou and Serafeim (2012) as well as on further control variables. All instruments have an impact on CSP as shown in the above studies, but obviously have no direct impact on credit ratings. Hence, they qualify as instruments. Finally, in the second stage, credit ratings are regressed on the CSP estimate of the first stage. We choose the ordered choice model as introduced by Kaplan and Urwitz (1979) and as applied in many studies (e.g., Dimitrov et al. 2015; Baghai et al. 2014; Alp 2013; Jiang et al. 2012; Becker and Milbourn 2011; Blume et al. 1998).

We show that within the environmental performance, the innovation dimension has the most significant impact on credit ratings. This is true for North America, Europe, and Asia. However, the magnitude of the effect differs between these regions. The impact of social performance in North America and Europe is mainly driven by diversity, while no social aspects are relevant for Asia. Our findings are important for real-world decision makers, as they enable the identification of those CSP dimensions that have an impact on credit ratings. As the positive link between selected CSP components and credit ratings indicates a lower default risk of firms with high CSP levels, practitioners may profit from this knowledge through a more precise evaluation of credit risk and the resulting incentives to act. Also, as better credit ratings are associated with lower financing costs, our results help to target investments efficiently, leading to cost savings. Particular investments in environmental product innovation are far more impactful than those for emission and resource reduction. Likewise, among the social dimensions, diversity and employment quality are to be prioritized in investment decisions.

The remainder of the paper is organized as follows. We review the related literature and consider theory in Sect. 2. Section 3 describes our international data set and Sect. 4 introduces the employed instumental variable and ordered probit methodology. Sect. 5 presents the empirical results followed by Sect. 6 with robustness tests. Finally, Sect. 7 concludes the paper.

2 Theoretical considerations

A recent stream in literature analyzes the relationship between CSP and credit ratings. Dorfleitner et al. (2020), Stellner et al. (2015), Jiraporn et al. (2014), Oikonomou et al. (2014), Attig et al. (2013), Bauer and Hann (2010), Bauer et al. (2009) and Frooman et al. (2008) all contribute important insights to the prevailing positive link between CSP and credit ratings. However, the combination of a state of the art credit risk model and an econometrical framework to identify causal relationships rather than simple correlations has not yet been pursued.

In theory, there are two possible relationships between CSP and credit ratings. The overinvestment view regards CSP as being a waste of scarce resources, but there is little evidence of this perspective. In contrast, the risk mitigation view is based on the idea that sustainable companies face lower risks.

For US firms, Oikonomou et al. (2014), Attig et al. (2013), Bauer and Hann (2010), and Frooman et al. (2008) find a strong positive link between the KLD environment score and credit ratings. Dorfleitner et al. (2020) report an improved prediction quality in their North America sample if they consider environmental performance in their model. Environmental practices affect the solvency of borrowing firms by determining their exposure to potentially costly legal, reputational, and regulatory risks according to Bauer and Hann (2010). Following the correlation-based evidence of the above-mentioned previous studies, we also conjecture a causal impact of (some of) the components of environmental performance on credit ratings. More concretely, we expect at least one of the environmental performance dimensions of emission reduction, resource reduction, and environmental innovation to have a positive impact on credit ratings.

Bauer et al. (2009) have already evidenced a positive relationship between the social pillar of CSP and credit ratings. Dorfleitner et al. (2020) report an improved prediction quality for North America, regarding a model that considers social performance. Through the breakdown into individual components, Attig et al. (2013) find that KLD social strengths and concerns correlate with credit ratings of US firms and that the individual components of CSP related to primary stakeholder management (i.e., community relations, diversity, and employee relations) matter most in explaining a firm’s creditworthiness. Oikonomou et al. (2014) identify a similar relationship for community, employment, environment, and product safety. The positive link between CSP components and creditworthiness appears plausible especially for employee relations, as these are associated with greater productivity, higher profitability, higher firm value, and superior shareholder returns (e.g., Huselid 1995; Prennushi et al. 1997; Ichniowski and Shaw 1999; Edmans 2011). Bauer et al. (2009) argue that employee relations affect bondholders through their influence on firm risk. Thus, firms with sound and competitive employment practices can enhance their capacity to generate higher and more stable cash flows while simultaneously preempting or mitigating the harmful behavior of dissatisfied employees. In contrast, poor employee relations can limit firms’ access to human capital, lead to the exit of valuable employees, increase both litigation and reputation risks, and raise transaction costs. Hence, we also expect a causal impact of (some of) the components of social performance on credit ratings. More narrowly, at least one of the social performance dimensions of product responsibility, community, human rights, diversity, employment quality, health, or training performance is expected to have a positive impact on credit ratings.

For the impact of CSP on some types of risk, it was already shown that this relationship varies regionally, e.g., Utz (2018) finds evidence for the risk mitigation view on the impact of CSP on idiosyncratic risk, while the overinvestment view seems to apply in Asia-Pacific. Some previous research on the relationship between CSP and credit ratings is provided for both North America and Europe. Jiraporn et al. (2014) find that the CSP policies of US firms are affected by CSP. Firms with high CSP have better credit ratings, i.e., by 4.5% for a one standard deviation change in the CSP level. In contrast, Stellner et al. (2015) find no relevance of Asset4’s overall CSP rating for credit ratings regarding Europe. Dorfleitner et al. (2020) also confirm regional deviations between North America and Europe in the explanation and prediction quality of credit ratings through CSP. While social performance is a predictor for credit ratings in both North America and Europe, this is only the case for environmental performance in North America in their setting. Given there is an impact, we expect the effect of environmental and social performance categories on credit ratings to differ regionally.

3 Data

Our sample includes S&P credit ratings, Asset4 CSP measures, and some instrumental and control variables. After excluding financial firms based on the Thomson Reuters Business Classification (TRBC), the final data set encompasses 1212 firms with 7032 firm-year observations. Tables 1 and 2 present descriptive statistics of the credit rating variable, respectively, of the Asset4 scores, the instruments, and the control variables. The regional classification into North America, Europe, and Asia is described in Table 3.

Table 1 This table reports on the total number of firms and observations per rating class including the partial quantity of rating upgrades and downgrades compared with the previous period for the entire sample

Table 2 This table reports the descriptive statistics for the asset scores, the instrumental variables, and control variables in our sample covering the period from 2002 until 2018

Table 3 This table reports the breakdown of our data panel on regions and countries which are the base for our panel selection when analyzing regional differences

The dependent variable of the second stage regression is the long-term borrower credit rating of S&P. These credit ratings reflect the creditworthiness of a borrower for a time horizon of at least 1 year. The referring rating grades comprise AAA, AA, A, BBB, BB, B, CCC, CC, and D. The default category D is assigned when obligors are overdue for their required payments. Vazza, and Kraemer (2017) provide further information on the rating methodology.

Table 4 This table presents the description of our selection on Asset4 CSP measures.

Table 5 This table provides an overview of the NBS categories (Whitley 1999) and their variables, which we select as instruments for use during the first stage of our 2SPS regressions based on the work of Ioannou and Serafeim (2012)

Table 6 This table describes used control variables that are firm specific except for GDP growth. All of them are delivered by Worldscope and Thomson Reuters Datastream

Table 7 This table reports on industry classes according to the economic sector level of Thomson Reuters Business Classification (TRBC). Financial firms are excluded

Asset4 publishes annual corporate social and environmental performance scores, which can be interpreted as being external measures for sustainable business models (Ioannou and Serafeim 2012; Chatterji et al. 2016; Humphrey et al. 2012). The scores include information from publicly available sources such as websites, SEC filings such as 10-K, DEF 14A, and 10-Q, sustainability reports, media sources, and NGO reports. The methodology is based on more than 700 questions about the fulfillment of a specific sustainable topic. Each question results in one data point. These pieces of information are aggregated to categories, which again are condensed to pillars. The approach of Asset4 allows us to overcome weaknesses of the KLD, FTSE4Good, and Dow Jones-rating approaches such as lack of transparency (Chatterji and Levine 2006) as far as possible. Following El Ghoul et al. (2017), we also derive the overall CSP performance from aggregating the environmental and social pillars. The final scores range from zero to 100% with high levels reflecting high CSP. The distribution of Asset4 scores may be skewed as the required information to assign a rating is easier to obtain from larger and high-CSP companies as badly performing firms are unlikely to provide the necessary information. As a consequence, we include size and a large set of further control variables in our models. The data is free from survivorship bias as post-bankruptcies, mergers, and other causes of de-listings are accounted for and the corresponding stocks are retained in the sample. A detailed description of the CSP scores is displayed in Table 4.

Table 8 This table gives an overview of both stages of our estimated models. The first stage includes instruments and control variables to estimate CSP scores as dependent variables. The second stage includes the estimate of the referring CSP score and the same control variables from the first stage with credit ratings as the dependent variable. The estimation results contain also boundaries needed to assign rating classes based on the linear predictor

Table 9 Pooled estimation

Our first instrument for CSP is selected based on the study by Jiraporn et al. (2014), who ascertain that the CSP policy of surrounding firms to have an impact on firm CSP performance. Thus we apply the average CSP score of all (available) surrounding firms within the same country. Second, a further set on instruments is included, namely the drivers for CSP in terms of ”national business systems” (NBS) according to Whitley (1999), such as the political, labor, education, and the cultural systems. The theoretical NBS category political system is measured with the aid of a regulations index, an anti-self-dealing index, an absence-of-corruption index, and an index for left/center political orientation. The education and labor system is modeled by union density and a skilled labor index while the cultural system involves indices for power distance and individualism. A detailed description of the variables of each NBS category is presented in Table 5.

We add further control variables based on previous research. Following Standard&Poor’s (2013) and Merton (1974), we include the three-year averages of the operating margin, the total debt, and the interest coverage ratios. The interest coverage ratio is transformed as suggested by Blume et al. (1998). We set negative values to zero because these could be due to low interest payments or high negative earnings, while both explanations have a contradictory impact on credit ratings. By assuming decreasing marginal effects for high levels of interest coverage, we cap the three-year average at 100. To model a non-linear shape, we transform the interest coverage $C_{it}$ of a company i in year t into four subvariables $c^A_{it}$, $c^B_{it}$, $c^C_{it}$, $c^D_{it}$ acording to:

	$c^A_{it}$	$c^B_{it}$	$c^C_{it}$	$c^D_{it}$
if $C_{it} \in [0, 5)$	$C_{it}$	0	0	0
if $C_{it} \in [5, 10)$	5	$C_{it}-5$	0	0
if $C_{it} \in [10, 20)$	5	5	$C_{it}-10$	0
if $C_{it} \in [20, 100)$	5	5	10	$C_{it}-20$ .

We control for firm size for two reasons. On the one hand, larger companies are less likely to default (Altman et al. 1977). On the other hand, the CSP scores are likely to be skewed with respect to firm size. Referring to Blume et al. (1998), we also control for systematic risk (market model beta) as well as idiosyncratic risk. The firms’ willingness to pay dividends can also be an indicator of credit risk (Hoberg and Prabhala 2009). Furthermore, firms with a high market-to-book ratio may be less likely to default (Pástor and Pietro 2003). Retained earnings are used to proxy a company’s life cycle phase (DeAngelo et al. 2006), whereas established companies tend to have better ratings (Fons 1994). Additionally, capital expenditure has been evidenced to influence credit risk (Tang 2009). We include cash among the controls because firms in distress tend to hold precautionary savings (Acharya et al. 2012). Furthermore, tangibility may have an impact on credit risk (Rampini and Viswanathan 2013). As S&P credit ratings appear to change at least to some extent pro-cyclically, the gross domestic product (GDP) growth rate is employed to model the business cycle. A detailed description of the above control variables is presented in Table 6. Time fixed effects are intended to catch all remaining systematic effects (Elton et al. 2001). Finally, we also control for industry-fixed effects. An overview of industries is delineated in Table 7.

Table 10 Panel North America

Table 11 Panel Europe

Table 12 Panel Asia

Table 13 First stage panel North America

Table 14 First stage panel Europe

In order to control for multicollinearity, we calculate variance inflation factors (VIF) for overall CSP scores, instruments, and control variables. If necessary, input variables are discarded in a selection process in order to maintain only VIFs below 10 indicating sufficient low exposure to multicollinearity. The variable ’individualism’ is discarded in that process for the combined dataset of all three regions. An estimation based on the full set of instruments is presented in the robustness checks.

Table 15 First stage panel Asia

Table 16 Robustness checks

Table 17 Robustness checks

Table 18 Marginal effects panel North America

Table 19 Marginal effects panel Europe

4 Methodology

As CSP and credit ratings are likely to be highly endogenous, our analysis is based on the instrumental variable approach to mitigate the bias due to the endogeneity of the input variables. Thus in the first stage, we regress the respective CSP factor on selected instruments and controls. All factors that can explain variation in CSP but do not affect credit ratings qualify as instruments.

Table 20 Marginal effects panel Asia

The first stage regression includes the CSP measure $x_{i,t-1}$ as a dependent variable, and instrument variables $\varvec{z}_{i,t-1}$ and controls $\varvec{c}_{i,t-1}$ as explanatory (vectorial) variables with referring coefficients vectors $\varvec{\beta }_z$ and $\varvec{ \beta }_c$ as described by:

$$\begin{aligned} x_{i,t-1} = \varvec{z}_{i,t-1} ' \varvec{\beta }_z + \varvec{c}_{i,t-1} ' \varvec{\beta }_c + \epsilon _{1,i,t}. \end{aligned}$$

(1)

This estimation is based on OLS. To account for the panel structure of our data, we include time-fixed effects among the controls and clustering of standard errors at the firm level.

The second stage regression is based on a model that was initially introduced by Kaplan and Urwitz (1979) and further developed by (e.g. Blume et al. (1998)). This model is applied in many studies (e.g., Dimitrov et al. 2015; Baghai et al. 2014; Alp 2013; Jiang et al. 2012; Becker and Milbourn 2011). Our threshold model is based on an unobserved linking variable $y_{it}^*$, which represents the creditworthiness of a firm i and year t and calculates

$$\begin{aligned} y_{it}^* = {\hat{x}}_{i,t-1} \beta _{{\hat{x}}} + \varvec{c}_{i,t-1} ' \varvec{\beta }_c + \epsilon _{2,i,t}, \end{aligned}$$

(2)

where ${\hat{x}}_{i,t-1}$ is the CSP estimate of the first stage and $\varvec{c}_{i,t-1}$ represents the vector of observed explanatory variables of firm i in the year $t-1$. Accordingly, $\beta _{{\hat{x}}}$ is the CSP coefficient while $\varvec{\beta }_c$ is a vector of coefficients for control variables. The linking variable $y_{it}^*$ is continuous and its range comprises the set of real numbers. In our study, we consider nine different levels of credit ratings (i.e., AAA, AA, A, BBB, BB, B, CCC, C, and D). The variable $R_{it}$ defines the rating category of firm i and year t. It takes the value 9 if firm i has a rating of AAA, 8 if AA, 7 if A, 6 if BBB, 5 if BB, 4 if B, 3 if CCC, 2 if CC and 1 if D in year t. Thus the first stage of our estimation maps the credit ratings into a partition of the unobserved linking variable $y_{it}^*$ as follows:

$$\begin{aligned} R_{it} = {\left\{ \begin{array}{ll} 9 & \text {if }\quad y_{it}^* \in [\mu _8,\mu _9) \quad \quad (AAA)\\ 8 & \text {if }\quad y_{it}^* \in [\mu _7,\mu _8) \quad \quad (AA)\\ 7 & \text {if }\quad y_{it}^* \in [\mu _6,\mu _7) \quad \quad (A)\\ 6 & \text {if }\quad y_{it}^* \in [\mu _5,\mu _6) \quad \quad (BBB)\\ 5 & \text {if }\quad y_{it}^* \in [\mu _4,\mu _5) \quad \quad (BB)\\ 4 & \text {if }\quad y_{it}^* \in [\mu _3,\mu _4) \quad \quad (B)\\ 3 & \text {if }\quad y_{it}^* \in [\mu _2,\mu _3) \quad \quad (CCC)\\ 2 & \text {if }\quad y_{it}^* \in [\mu _1,\mu _2) \quad \quad (CC)\\ 1 & \text {if }\quad y_{it}^* \in (\mu _0,\mu _1) \quad \quad (D) , \end{array}\right. } \end{aligned}$$

(3)

where $\mu _1,\dots ,\mu _8$ are partition points independent of time t while $\mu _0 =-\infty$ and $\mu _9=\infty$. Thresholds are not given ex-ante but instead determined in the statistical estimation procedure. The assumption that $\epsilon _{it}$ is normally and independently distributed with a mean of 0 and a variance of 1 is ensured in the estimation. We obtain a certain rating (i.e., a realization of $R_{it}$) and a realization of the input variables for each company and each year during the observation period.² The explanatory variables are lagged by one period to model the status of information at the time of prediction. Table 8 provides an overview of the input factors, boundaries, and outputs of the estimated models.

Following the assumption that $\epsilon _{2,i,t,}$ is normally and independently distributed with a mean of 0 and a variance of 1 and given ${\hat{x}}_{i,t-1}$ and $\varvec{c_{i,t-1}}$, the probability of assignment to a specific rating class can be calculated according to:

$$\begin{aligned} P(R_{it}=j| {\hat{x}}_{i,t-1}, \varvec{c}_{i,t-1}) ={\varPhi } (\mu _j - {\hat{x}}_{i,t-1} \beta _{{\hat{x}}} + \varvec{ c_{i,t-1} \beta _c } ) - {\varPhi } (\mu _{j-1} - {\hat{x}}_{i,t-1} \beta _{{\hat{x}}} + \varvec{ c_{i,t-1} \beta _c } ) \end{aligned}$$

(4)

with $j = 1,...,9$, $\mu _0 =-\infty$ and $\mu _9=\infty$.

5 Empirical tests

To test our hypotheses, we estimate a total of 13 different model specifications. Starting with a model of overall CSP, two further models include the environmental or the social pillar respectively. Further models focus on each of the components contained in the pillars, respectively. Concerning environmental performance, we estimate models for emission, environmental innovation, and resources. Referring to social performance, additional models include product responsibility, community, human rights, diversity, employment quality, health, and training. All of these models are estimated on the pooled dataset of North America, Europe, and Asia in two stages based on the 2SPS approach. Each model considers one CSP score as a dependent variable in the first stage regressed on instrumental and control variables. The corresponding second step includes the credit rating as the dependent variable with both the CSP estimate and the same controls from the first stage as independent variables. The regression results for both stages of all models are presented in Table 9. Moreover, we test for weak instruments in the first stage and report adjusted $R^2$ values as goodness-of-fit measures for both stages of every model.

5.1 The impact of CSP and its components

The first stage regression results for the overall CSP, the environment, and the social model in our pooled sample of North America, Europe, and Asia show that some of our instruments are significant and hence add an important explanation to the CSP scores. The test on weak instruments delivers p values close to zero, implying that the null hypothesis of weak instruments can be rejected. In the second stage, we find coefficients of overall CSP in all three regions to be positive and significant on a 1% level. The sign indicates that strong CSP performance tends to be linked to better credit ratings. Thus increases of firm CSP also tend to go along with credit rating improvements. Hence, these results confirm the risk mitigation view. By implementing the argument of Galema et al. (2008) that aggregating multiple categories of CSP may hide confounding effects among the single components of corporate social and environmental performance, we focus on CSP components in the following.

When targeting the environmental category level of Asset4 CSP scores, we find that all environmental categories (emission, environmental innovation, and resources) are relevant. A consideration of the most distinct result regarding environmental innovation raises the question of why conventional control variables such as R&D expenses cannot catch the effect. First, we argue that CSP aims to measure future long-term development while the accounting ratios included in controls represent solely the status quo. Second, CSP also catches intangible assets which are likely not (fully) reflected in accounting ratios. Previous research reveals some reasons for the potential relationship between environmental innovation and firm performance. Environmental innovation may increase efficiency and hence decrease total material cost (Porter and Van der Linde 1995). Additionally, businesses can gain competitive advantages through green product and green process innovation (Chen et al. 2006). Moreover, Kammerer (2009) argues that product innovation also increases the customer benefits and thereby also the demand. Furthermore, a positive impact on the market performance is confirmed by Pujari (2006), including reputation among the potential drivers of this (Eiadat et al. 2008).

Next, we analyze which single categories of the social performance dimension drive the impact on credit ratings. Our findings show a significant positive impact of health and diversity, while the latter is more important in terms of significance. A considerable number of empirical studies identifies a positive relationship between gender diversity in the boardroom and firm performance for North America (Carter et al. 2003; Erhardt et al. 2003; Miller and del Carmen Triana 2009) and European countries (Campbell and Mínguez-Vera 2008; Reguera-Alvarado et al. 2017; Lückerath-Rovers 2013). A similarly positive relationship can be formulated between gender diversity in management and firm performance if moderated by a firm’s strategic orientation and the organizational culture (Dwyer et al. 2003). Contrasting views (e.g., Adams and Ferreira 2009; Marinova et al. 2016) exist, but are less widespread. Possible explanations include the conjecture that diversity may help in decision processes by introducing other perspectives and information and additionally a different assessment of risk (Gul et al. 2011). Moreover, a diverse mindset within firms helps to catch up with business and society trends of the customer base and attract talented personnel (Li and Chen 2018).

5.2 The region matters

Tables 10, 11 and 12 show the second stage results for separate estimations on the panels of North America, Europe, and Asia.³ When focusing on North America, we find all dimensions (emission, environmental innovation, resources, product responsibility, community, human rights, diversity, employment quality, health, and training) to be positively significant. Concerning Europe, we find the dimensions environmental innovation and diversity to be significantly positively related to credit ratings. The measures community and training are weakly significant on a 10% level and the first reveals a negative sign. Coefficients in the Asia subsample are significant in the dimensions of emission, environmental innovation, and resources. Among the social categories, no dimension is significant. Except for the community category in Europe, all significant CSP coefficients show positive signs indicating the positive link between the referring CSP scores and credit ratings. Our results once more generally support the risk mitigation view.

As the link function in our model limits the interpretability of the CSP impact, marginal effects (at means of the controls) according to Greene (2011) are calculated. In the Tables 18, 1920, one can observe the practical implications of our results. Predominantly, we see increases of the probability to obtain a better rating class if the CSP score is significant and is increased by 1% point (ceteris paribus). At the same time, the probability to obtain a worse rating class decreases. For example, the probability of an actual BBB rated North American counterparty to upgrade to an A rating increases by 0.43% points if the overall CSP score increases by 1% point under otherwise identical circumstances, while the probability of a downgrade to BBB decreases by 0.48%.

In general, we support the argumentation of Attig et al. (2013) that CSP helps to generate intangible assets such as reputation and relationships with stakeholders, which again improve a firm’s competitiveness (Orlitzky et al. 2003). This may explain the relevance of all CSP scores in North America. The same argumentation may apply also for Europe. However, besides environmental innovation and diversity no further CSP component is significant on the 1% level, which is likely caused by the comparably high mean levels and low variation of CSP of European firms. For example, emission exhibits a higher mean of 81.3% and a lower standard deviation of 19.5% when compared to the mean (52.3%) and standard deviation (31.8%) of North American companies. As a result, European firms can differentiate less from each other through emission reduction. In contrast, environmental innovation shows a higher standard deviation (29.3%) and turns out to be significant. In the case of diversity, the pressure of the market seems so strong that smaller variation suffices for a significant impact. In the Asia panel, not even diversity is significant, although obtaining the most intensive effects among social components in Europe and North America—likely due to cultural reasons. Previous literature has provided similar implications for Asia by finding limiting or reducing the effects of diversity aspects on firm performance. Based on a sample of Asian countries (Hong Kong, South Korea, Malaysia, and Singapore), Low et al. (2015) primarily find a positive effect of the numbers of female board directors on firm performance, although it is substantially reduced in countries with higher female economic participation and empowerment likely due to tokenism. Li and Chen (2018) only find a positive relationship between board gender diversity and firm performance for Chinese firms if they do not exceed a critical size. Darmadi (2011) even found a negative relationship between the diversity of board members and financial performance for Indonesia.

When comparing our results with those of earlier studies on credit risk, we find accordance with Jiraporn et al. (2014) in the sense that overall CSP has a positive impact on credit ratings in North America. Stellner et al. (2015) do not find such a relationship for their Europe sample. In contrast, we find a significant positive impact of both overall CSP and some of its components. In agreement with Oikonomou et al. (2014), product characteristics are relevant in this context. Further, we identify the workforce categories of employment quality and diversity as being drivers inside the workforce pillar. We can confirm the first empirical evidence of Attig et al. (2013), with CSP strengths and concerns related to primary stakeholder management (i.e. community relations, diversity, employee relations, environmental performance, and product characteristics) being linked to credit ratings and extending their work in terms of causality and a more sophisticated CSP measurement approach, respectively.

6 Robustness checks

We prove the robustness of our results concerning the specification of instruments in the first stage, to the period selection, missing data, and the relevance of environmental sensitive industries. Regression coefficients of CSP variables are presented in Tables 16 and 17. At first, we address the average CSP performance of surrounding firms used as an instrument based on the research of Jiraporn et al. (2014). While in the standard analysis, the average CSP for the USA is calculated based on the country level, we demonstrate the robustness of our results when surrounding firms are defined as located in the same state. All results remain almost unchanged. Also, in the main analysis, instruments and controls are subject to a selection process based on VIFs. To prove the robustness of our results for the entire sample, we include the individualism variable, which was discarded in the selection process. Referring to the three regional panels, we replace one instrument in each. Again, we derive similar results to our main findings.

During recent years, there have been several changes in the political alignment of some countries, e.g., since 2017, US climate politics have shifted from renewable energies back to a stronger focus on fossil fuels. Hence, we analyze whether our findings are subject to any development in recent years. We run estimations with a sample reduced by observations of the most recent year in the sample, and also the same for the second and the third recent year. As a result, we see no substantial deviations in the CSP effects for any of those time variations in the sample. In this context, we also address the case of missing data. After matching the final dataset, each combination of credit rating, CSP, and control variables per time and company is dismissed if any data value relating to these variables is missing. To measure the impact of the missing control variables’ data, we implement a mean imputation according to Schafer (1997). Instead of discarding missing observations, we substitute them by the mean. Again, the corresponding estimations support our main result.

As the industry appears to be significant in terms of the impact of environmental CSP dimensions (Khan et al. 2016), we additionally analyze the impact of the industry through an interaction of CSP with a dummy variable expressing whether a firm belongs to the “NAICs Codes of Environmental Sensitive Industries” published by the US Small Business Administration. In our sample, we find no evidence that the impact is stronger there.

7 Conclusion

While the corresponding literature has researched the general impact of overall CSP on credit risk, the identification of the actual drivers on a lower aggregation level of CSP has so far not been addressed adequately. We supplement earlier studies by using CSP measures based on the more sophisticated and more transparent methodology of Asset4. Moreover, international data coverage allows us to analyze (and compare) the three regions of North America, Europe, and Asia with a consistent methodology and data set. Compared with the majority of previous studies, our analysis focuses on single components of CSP. We account for the requirements of both the consideration of endogeneity regarding the impact of CSP on credit ratings and recent credit risk modeling by applying the instrumental variable approach in terms of the two-stage predictor substitution with an established credit risk model in the second stage. This approach allows us in particular to provide clearer indications of a causal relationship in terms of how CSP components impact credit ratings in contrast to the common approaches, which only reveal correlational relationships.

We initially estimate the impact of overall CSP on credit ratings to confirm the findings of the previous literature. Then we investigate which of the CSP dimensions can improve the quality of credit rating predictions. Each of the three environmental categories has a significant positive impact while environmental innovation is most distinct. As part of social performance, measures for community and diversity (involving equal opportunities) are significant. With respect to differences between North America, Europe, and Asia, the impact of social performance is driven by the extent of diversity only in North America and Europe, which has no impact in Asia and is likely due to cultural reasons. Product innovation is still the determining driver within the environmental performance of all regions.

The identification of the drivers of impact for CSP on credit ratings has important implications for practice. Some of the CSP dimensions generally act in a risk-mitigating manner in terms of default risk, for which credit ratings are a proxy. From this point of view, investments in CSP are not a waste of resources. Moreover, because better credit ratings are associated with lower financing costs, our results help to target investments in CSP for the purpose of referring cost reductions efficiently. In particular investments in product innovation and diversity appear to have the highest impact.

With the identification of these CSP components that lead to lower credit risk, our analysis shows that some, but not all aspects of CSP produce favorable effects beyond a philanthropic rationale. However, as a limitation, it has to be noted that real-world causality in the context of this relationship can only be proven by means of natural or quasi-experiments, therefore confirming the necessity for continued research in the future.