The causal relation between corporate social performance and corporate financial performance
A study on the role of different models, methods and measures
Abstract
This study examines the causal relationship between corporate social performance (CSP) and corporate financial performance (CFP). Our aim is to analyze how different methodologies and different measures of CSP and CFP affect results. This study analyzes the CSP-CFP relationship by applying three methodologies and three measures of CFP. In addition, the analysis is conducted for two measures of CSP: the generic measures (aggregate measures of CSP provided by ASSET4) and the material measure (a self-constructed outcome-based measure of corporate environmental
performance). The study finds that there is no strong evidence for a causal relationship between CSP and CFP. Secondly, we find that results are inconsistent across methodologies and measures of CFP when the analysis is conducted using the generic measures of CSP. This inconsistency however disappears, when the material measure is used. We conclude that the way in which CSP is measured strongly affects the results on the CSP-CFP relationship.
Key words: Corporate social performance, corporate financial performance, CSP/CFP link, Granger causality
Word count: 16923
Student: Elizabeth van der Wagen
Student number: s2587580
Study program: MSc. Finance
Supervisor: Prof. dr. L.J.R. Scholtens
June 13, 2016
1. Introduction
This study analyzes the causal relationship between corporate social performance (CSP) and corporate financial performance (CFP). The study focuses on how the application of different methodologies and the use of different measures of CSP and CFP affect the results. Our study is a direct response to the ambiguous results from earlier work on the CSP-CFP relationship such as Makni et al. (2009), Nelling and Web (2008), Scholtens (2008) and Waddock and Graves (1997). Some authors have argued that this ambiguity stems from the great variety of methodologies applied (for example: Margolis and Walsh (2001)). Others argue that the different measures used to capture CSP and CFP affect outcomes (Dixon-Fowler et al., 2013; Orlitzky et al., 2003). This study will focus on finding the source of the inconsistent results across studies.
The study takes a systematic approach to analyze the causal relationship between CSP and CFP. This systematic approach entails that the relationship is analyzed by applying three methodologies and using three measures of CFP. In addition, the analysis is conducted for two measures of CSP. The three methodologies applied are Ordinary Least Squares (OLS) regressions with lagged values of the independent variable, OLS fixed effects regressions with lagged values of the independent variable and Vector Autoregressive (VAR) models in combination with Granger causation tests. The three measures of CFP are the market-to-book ratio, return on assets and market adjusted returns. The analysis is performed for two measures of CSP. First, the regressions will be estimated using aggregate measures of CSP provided by Thomson Reuters’ ASSET4. These aggregate measures however, have several limitations. These limitations will be addressed by repeating the analysis using a self-constructed outcome-based measure of corporate environmental performance (CEP).
This study finds that there is no strong evidence for a causal relationship between CSP and CFP. Secondly, we find that results are highly inconsistent across methodologies and measures of CFP when the aggregate measures of CSP are used. This inconsistency however, largely disappears when we repeat the analysis using our self-constructed measure of CEP. We therefore conclude that the way in which we measure CSP strongly affects the results on the causal relationship between CSP and CFP.
The remainder of this paper is organized as follows. First, there is a literature review on corporate social responsibility and the causal relationship between CSP and CFP. Based on this literature review, the hypotheses for this study are formulated. Section three discusses the data that is used. It will elaborate on the measures for CSP and CFP and will discuss the control variables. Section four presents the three methodologies applied in this study. Next, section five presents and discusses the main results. The paper ends with a conclusion and recommendations for future research.
2. Literature review
This literature review starts with a definition of corporate social responsibility and corporate social performance. In addition, it will discuss why companies engage in corporate social responsibility. Next, a theoretical framework on the causal CSP-CFP relationship is presented followed by a discussion on how the sign of this relationship changes when different measures of CFP are used. Lastly, the results of earlier work in the field will be discussed and we will explain how this study contributes to this literature.
2.1. A definition of corporate social responsibility and the reasons why firms engage in it
There are many different definitions for corporate social responsibility (see Dahlsrud (2008) for an overview). It is therefore important to indicate how corporate social responsibility is defined in this study. In this paper we use the definition of the Commission of European Communities (2001). They define corporate social responsibility as a concept whereby companies integrate social and
environmental concerns in their business operations and in their interaction with their stakeholders on a voluntary basis (p.6). Closely related to corporate social responsibility is the concept of corporate social performance. According to Pätäri et al. (2014), the level or practice of corporate social responsibility in a firm is reflected in its CFP. CFP thus measures how a company performs in the domain of corporate social responsibility.
Margolis and Walsh (2001) mention two reasons as to why firms engage in corporate social
responsibility. First, corporate social responsibility has become a trend over the past decades. Firms are held responsible by society for social problems they cause, but are also expected to contribute to problems which they do not directly cause. Bénabou and Tirole (2010) list several factors that contribute to this trend in corporate social responsibility. They mention that information on business’ practices has become easily accessible. This entails that stakeholders can more easily assess whether the firm acts in line with their interests. An additional factor is that the costs of certain externalities such as pollution and the awareness of these costs have risen significantly. According to Margolis and Walsh (2001), a second phenomenon that triggers firms to engage in corporate social responsibility is the fact that they take on more responsibilities in a global setting. Because of their globalized
social and environmental externalities are exerted in weakly regulated countries. This gives rise to the need of a systematic approach for managing these externalities. Corporate social responsibility can be seen as such an approach.
2.2. Theoretical framework on the CSP-CFP relationship
Over the past decades, many studies have focused on the relationship between CSP and CFP (see Friede et al. (2015) for an overview). The majority of these studies base their hypotheses on the theoretical framework of Preston and O’Bannon (1997). Preston and O’Bannon (1997) were one of the first to hypothesize and test for six possible relationships between CSP and CFP. They state that research concerned with the CSP-CFP relationship deals with two empirical questions. The first question relates to the sign of the relationship. The relationship may be either positive, negative or non-existent. The second question deals with causality. Is it CSP that causes CFP, or is it CFP that causes CSP? In addition, there is the possibility of a bidirectional relationship between the two concepts. Preston and O’Bannon (1997) develop six hypotheses on the CSP-CFP relationship. These hypotheses will be used in this study and are described next.
The first hypothesis is the social impact hypothesis. This hypothesis states that high levels of CSP lead to high levels of CFP. The social impact hypothesis is based on stakeholder theory as proposed by Freeman (1984). Freeman (1984) was one of the first to highlight the importance of the various stakeholders of the firm. Stakeholder theory argues that firms which only focus on the interests of shareholders will ultimately fail. A firm needs to invest in its relationships with key stakeholders to be financially successful (Baird et al., 2012). According to Waddock and Graves (1997), investments in corporate social responsibility contribute to improved relationships with a firm’s stakeholders and will ultimately lead to increased financial performance.
The second hypothesis is the trade-off hypothesis. This hypothesis states that high levels of CSP lead to low levels of CFP. The trade-off theory is based on the ideas of Friedman (1970) who argues that there are few measurable benefits of CSR while the related costs are substantial. According to Friedman (1970), managers invest in corporate social responsibility to pursue their own interests, which does not maximize shareholder wealth. Corporate social responsibility worsens the competitive position of the firm and reduces financial performance.
The third hypothesis is the available funds hypothesis. It suggests a positive causal relationship between CFP and CSP. The theory argues that investments in corporate social responsibility are costly, and therefore depend on the resources available (Preston and O’Bannon, 1997). Past
this reduces the resources that they have under control, thereby reducing the management’s power. In order to keep their control, the management over-invests in non-profitable projects. The available funds hypothesis can be a manifestation of this agency problem. McGuire et al. (1988) were among the first to formulate and empirically test the available funds hypothesis.
The fourth hypothesis is the managerial opportunism hypothesis. It suggests a negative causal relationship between CFP and CSP. Here, the main idea is that managers act to maximize their own private benefits at the cost of shareholders’ and stakeholders’ interests. The theory argues that after good financial performance managers like to reward themselves and spend less on corporate social responsibility. However, after bad financial performance managers will do the opposite. Management will spend more on corporate social responsibility to cover up or justify bad financial performance (Preston and O’Bannon, 1997).
An overview of the four hypotheses is presented in figure 1.
Figure 1
Hypotheses on the relationship between CSP and CFP
Preston and O’Bannon (1997) formulate two additional hypotheses. These hypotheses propose a bidirectional relationship between CSP and CFP. The virtuous circle hypothesis suggests a
The theoretical framework of Preston and O’Bannon (1997) is used to formulate the hypotheses of this study.
Hypothesis H1– the social impact hypothesis:
CSP precedes CFP, the relationship is positive (H1)
Hypothesis H2– the trade-off hypothesis:
CSP precedes CFP, the relationship is negative (H2)
Hypothesis H3 – the available funds hypothesis:
CFP precedes CSP, the relationship is positive
(H3)
Hypothesis H4– the managerial opportunism hypothesis:
CFP precedes CSP, the relationship is negative (H4)
In addition to the four unidirectional hypotheses, we will test for two bidirectional relationships:
Hypothesis H5– the virtuous circle hypothesis:
There is a positive bidirectional relationship between CSP and CFP (H5)
Hypothesis H6– the negative synergy hypothesis:
There is a negative bidirectional relationship between CSP and CFP (H6)
In all cases, the null hypothesis is that there is no relationship between the CSP and CFP.
2.3. Theoretical framework on the CSP-CFP relationship for different measures of CFP
for three different measures of CFP including the market-to-book ratio, return on assets and market adjusted returns. In this study, we will use these propositions as hypotheses and they are described next.
The first proposition is that the market-to-book (MTB) ratio of a socially responsible firm is always larger than the MTB ratio of an irresponsible firm. The smaller MTB ratio of irresponsible firms is due to their smaller investor base. The first reason for this smaller investor base is that investor’s
preferences are heterogeneous. Responsible investors do not want to invest in irresponsible companies leading to a lower demand and lower prices. According to El Ghoul et al. (2011) the smaller investor base of irresponsible companies can also be explained by informational asymmetry. They argue that information asymmetry is likely to be more severe for irresponsible firms since they typically disclose less information. The high information asymmetry causes the demand for irresponsible shares to go down, resulting in a lower share price. This concept is based on the behavioral assumption that investors are more likely to buy a stock when they know about it.
The second proposition is that the return on assets (ROA) of a socially responsible firm is always larger than the ROA of an irresponsible firm. Responsible firms will integrate the cost of social damage in their cost of capital. Irresponsible firms will not internalize the cost of social damage and therefore use a cost of capital that is too low. The result is that a responsible firm will use less capital than irresponsible firms. Less input of capital will result in reduced profits. Due to decreasing
marginal returns to capital however, capital will decrease more than profits. This means that ROA will be higher for responsible firms.
The third proposition states that it is ambiguous whether market adjusted returns are higher for
socially responsible firms or irresponsible firms. Since a higher MTB ratio lowers expected returns but higher ROA would increase expected returns, the ultimate effect for market adjusted returns is unclear.
This study will use the MTB ratio, ROA and market adjusted returns as measures for CFP. Based on the theoretical framework of Dam and Scholtens (2015) we set three additional hypotheses.
The relationship between CSP and MTB is positive (H7)
The relationship between CSP and ROA is positive (H8)
The relationship between CSP and market adjusted returns is ambiguous (H9)
2.4. Empirical findings on the CSP-CFP relationship
and find that 90 percent of these studies find evidence of a non-negative relationship between CSP and CFP. Margolis and Walsh (2001) summarize findings on the CSP-CFP relationship as well. They point out that the focus of most previous work has been on the sign of the relationship. According to them, studies that focus on causality are much scarcer. In response to this research gap, several studies have specifically focused on the causal relationship between CSP and CFP. Table A1 in appendix A gives an overview of these studies. The last column of the table shows the results of the studies which will be discussed next.
One of the first to address the causal relationship between CSP and CFP are Waddock and Graves (1997). They apply a basic OLS model including one lag of the independent variable. The analysis is performed for three accounting-based measures of CFP. Using this relatively simple approach, Waddock and Graves (1997) find evidence for the virtuous circle hypothesis. A decade later, various authors reexamine the CSP-CFP relationship. Table A1 shows that the results across those studies are highly inconsistent and not in line with the findings from the work of Waddock and Graves (1997). Using improved statistical techniques, Nelling and Webb (2008) find no evidence for the virtuous circle hypothesis. After applying various methodologies, they only find that stock returns positively affect CSP which is in line with the available funds hypothesis. Similarly, Makni et al. (2009) find no evidence of a virtuous circle either. They only find that CSP negatively affects stock returns, which supports the trade-off hypothesis. Scholtens (2008) finds that it is more likely that CFP causes CSP than the other way around. To the contrary, Pätäri et al. (2014) find no evidence for CFP causing CSP. We can thus conclude that results on the causal relationship between CSP and CFP are highly
inconsistent.
Table A1 shows that there is an additional factor that possibly accounts for the ambiguous results across studies. We observe that the studies analyze the causal CSP-CFP relationship by estimating regressions with different lag-lengths. Makni et al. (2009), Nelling and Webb (2008) and Waddock and Graves (1997) all use a lag-length of one. This means that they examine whether CSP/CFP at time t is related to CFP/CSP at time t-1. Pätäri et al. (2014) and Scholtens (2008) employ longer
lag-lengths. This entails that they also examine whether CSP/CFP at time t is related to CFP/CSP at time t-2, t-3,…,t-n. It is interesting to observe that only Pätäri et al. (2014) elaborate on why they choose to work with a specific lag-length. All the other studies fail to do this. We argue the chosen lag-length can be an additional factor that affects results on the causal CSP-CFP relationship.
We conclude that results of earlier work on the causal relationship between CSP and CFP are highly inconsistent across studies. Possible reasons for this ambiguity are that each study applies different methodologies, uses different measures of CSP/CFP and estimates regressions using different lag-lengths.
2.5. The contribution of this study to previous work
This study is a direct response to the ambiguous results from earlier studies on the CSP-CFP relationship. Our study focuses on finding the source of the ambiguity by examining how the application of different methodologies, the use of different measures for CSP/CFP and the use of different lag-lengths affect results. We will do this by systematically analyzing the CSP-CFP relationship. This systematic approach entails that we will apply three methodologies. These methodologies are standard OLS with lags of the independent variable, OLS with fixed effects with lags of the independent variable, and VAR modelling in combination with Granger causation tests. Secondly, the relationship will be analyzed using different measures of CFP. These measures are the MTB ratio, ROA and market adjusted returns. Thirdly, the regressions will be estimated for different lag-lengths. Lastly, the analysis is conducted for two different measures of CSP. First, the analysis is conducted for aggregate measures of CSP published by ASSET4. After this, the analysis is repeated using a self-constructed outcome based measure of CEP.
3. Data
3.1. Measurement of corporate social performance
This study uses ASSET4 ESG data provided by Thomson Reuters to measure CSP. ASSET4 ESG data is chosen for several reasons. The first reason is that ASSET4 is a global dataset. It covers more than 4000 companies for the time period of 2002-2015. A second reason is that ASSET4 ESG publishes overall ESG scores. The availability of overall ESG scores makes the data easy and efficient to work with. A third reason to use ASSET4 ESG data is that it is easily accessible via DataStream. ASSET4 ESG data has been used in studies on the CSP-CFP relationship, examples are Daszynska-Zygadlo et al. (2016) and Dorfleitner et al. (2013).
3.1. Choice of dimensions
The first part of this study uses aggregate measures of CSP to analyze the causal relationship between CSP and CFP. Section 3.1.1 will elaborate on the characteristics of this measure. In the second part of the study, we redo the analysis using a self-constructed outcome-based measure of CEP. Section 3.1.2 will discuss why we use this measure and how it is derived.
3.1.1. Aggregate ASSET4 scores (generic measures)
The first part of the study will make use of the aggregate ESG scores as published by ASSET4. These aggregate measures of CSP will hereafter be referred to as the generic measures of CSP. Table C1 in appendix C provides an overview of the taxonomy of the ASSET4 data. It shows that ASSET4 reports on four pillars which represent different dimensions of CSP. This study will use the aggregate scores on three of those pillars. Those three pillars are the environmental performance pillar, the social performance pillar and the corporate governance pillar. In addition, we will use an aggregate measure of CSP which is calculated by taking the average of these three pillars. Table C1 shows that ASSET4 also reports on an economic pillar. The economic pillar is not included in this study. The reason for this is that including this pillar increases the problem of endogeneity. Namely, ROA is part of the economic pillar and is also used in this study to measure CFP.
The first implication relates to the interpretation of the results in this study. The use of z-scores entails that results are interpreted differently than results of the majority of studies on the causal relationship between CSP and CFP. These studies use absolute measures of CSP such as the ratings by Kinder, Lydenberg and Domini (KLD) Research and Analytics. The absolute nature of these ratings allows these studies to investigate whether higher/lower levels of CSP relate to higher/lower levels of CFP and vice versa. To the contrary, the ASSET4 ratings have a relative nature. The ASSET4 z-score reflects the CSP of a specific company relative to the average CSP of all companies rated by ASSET4. This study therefore investigates whether relatively higher/lower levels of CSP relate to relatively higher/lower levels of CFP, and vice versa. In order to accomplish this, all measures of CFP and the employed control variables are z-scored.
The second implication stems from the normalization of the z-scores. ASSET4 normalizes its scores by calculating a so-called scaling divisor. This scaling divisor is used to make the z-scores fit into the range of zero to hundred. Equation C3 in appendix C shows how the scaling divisor is calculated. It shows that the scaling divisor takes on different values for different underlying data patterns. If the pattern of the underlying data varies from year to year, this means that the scaling divisor varies from year to year. In this case it is not possible to draw inferences from increasing or decreasing normalized z-scores over time. Namely, a change in a normalized z-score can simply reflect a change in the scaling divisor.
In addition to the implications discussed above, the use of ASSET4 pillar scores has some limitations. Firstly, ASSET4 bases its scores on self-reported data. The ratings are therefore based on the firm’s own definition and evaluation of CSP. To this date however, this concern applies to all ESG ratings. A second limitation of the ASSET4 ratings is that they are largely process-based. Process-based
measures of CSP focus on managerial principles and processes. They capture a firm’s efforts to address corporate social and environmental issues (Endrikat et al., 2014). Examples of process-based measures include data points that answer questions such as “Does the company have a policy
In addition to the specific limitations of the ASSET4 data, we will discuss some general problems that arise when working with ESG ratings. These problems directly link to the aggregation of data. First, the aggregation of data requires us to assume that this data is commensurable (Graafland et al., 2004). Commensurability entails that the aggregated values are comparable with each other. In the context of corporate social responsibility it is unlikely that this assumption holds. For example, employee satisfaction and the toxic emissions of a firm are clearly not commensurable. Secondly, the
aggregation of data requires a decision on which weight to apply to each aggregated component. This can be done by combining the preferences of various stakeholders. However, Condorcet’s paradox shows us that collective preferences can be non-transitive, even if the preferences of individuals are transitive (Silver, 1992). Thirdly, we should think about whether good scores can offset bad scores. Using a total ESG score implies a symmetrical treatment of good and bad scores. Mattingly and Berman (2006, p. 20), state that “positive and negative social action are both empirically and conceptually distinct constructs and should not be combined.”
Despite of the limitations of the ASSET4 ESG ratings, the first part of this study will use these
measures of CSP. By doing this, our study is in line with previous empirical literature on the CSP-CFP relationship. These studies also work with aggregated scores on predefined categories proposed by the rating agencies (for example Makni et al. (2009), Nelling and Web (2008), Pätäri et al. (2014),
Scholtens (2008) and Schreck (2011)).
3.1.2. Environmental performance data points (material measures)
In the second part of this study, the relationship between CSP and CFP is analyzed using a self-constructed outcome-based measure of corporate environmental performance (CEP). This measure of CEP will be referred to as the material measure of CEP. By constructing this measure, we will address two concerns related to the use of the generic measure of CSP. Firstly, by constructing a
self-composed measure of CEP we are no longer restricted to use the published normalized z-scores. This entails that the main concern in our previous analysis – the concern related to the scaling divisor – is omitted. Secondly, by using the self-constructed measure of CEP we are able to shift our attention away from process-based measures to outcome-based measures.
In order to construct the aggregate measure of CEP, the following methodology is applied. First, each individual data point is scaled according to equation 1:
𝑆𝑐𝑎𝑙𝑒𝑑 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡 = 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡
𝑡𝑜𝑡𝑎𝑙 𝑠𝑎𝑙𝑒𝑠 𝑖𝑛 𝑈. 𝑆. 𝑑𝑜𝑙𝑙𝑎𝑟𝑠 (1)
The data points CO2 and CO2 equivalent emissions, total energy consumption and water withdrawal are each divided by total sales in U.S. dollars of the specific company.
Next, the inverse is taken of the variables CO2 and CO2 equivalent emissions divided by sales, total energy consumption divided by sales and water withdrawal divided by sales, as shown by equation 2:
𝐼𝑛𝑣𝑒𝑟𝑠𝑒 𝑜𝑓 𝑠𝑐𝑎𝑙𝑒𝑑 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡 = 1
𝑠𝑐𝑎𝑙𝑒𝑑 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡 (2)
By taking the inverse of the relationship we ensure that for all variables the interpretation is similar. Namely, the higher the value of the variable, the better CEP.
Lastly, each variable is z-scored. The material measure of CEP is then calculated as the weighted average of these z-scores. This implies that we analyze whether relatively higher/lower levels of CEP relate to relatively higher/lower levels of CFP, and vice versa. In order to accomplish this, all other variables entering the regression equation will be z-scored.
3.2. Measurement of corporate financial performance (CFP)
An aim of this study is to analyze how different measures of CFP affect results on the CSP-CFP relationship. This study will use the MTB ratio, ROA and market adjusted returns as measures of CFP. These three measures are chosen since they are commonly used in empirical work on the causal relationship between CSP and CFP (for example: Makni et al. (2009), Nelling and Webb (2008), Pätäri et al. (2014), Schreck (2011) and Waddock and Graves (1997)). In addition, the use of these measures allows us to set hypotheses based on the theoretical framework of Dam and Scholtens (2015). This section will elaborate on how the MTB ratio, ROA and the market adjusted returns are calculated. In addition, we will discuss the characteristics of each measure.
The first measure of CFP is the market-to-book ratio. The MTB ratio of a company reflects a
The MTB ratio is calculated as:
𝑀𝑇𝐵 𝑟𝑎𝑡𝑖𝑜 = 𝑚𝑎𝑟𝑘𝑒𝑡 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑐𝑜𝑚𝑚𝑜𝑛 𝑒𝑞𝑢𝑖𝑡𝑦
𝑏𝑎𝑙𝑎𝑛𝑐𝑒 𝑠ℎ𝑒𝑒𝑡 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑐𝑜𝑚𝑚𝑜𝑛 𝑒𝑞𝑢𝑖𝑡𝑦 (3)
The second measure for CFP is return on assets. ROA indicates how profitable a company is relative to its total assets. ROA is an accounting-based measure. Accounting-based measures reflect internal decision-making capabilities and managerial performance; they capture the firm’s efficiency (Orlitzky et al., 2013). They have a short-term perspective and are backward looking measures of CFP. A limitation of accounting-based measures is that they are subject to the accounting policy of the firm and that they are sensitive to managerial manipulation (Cochran and Wood, 1984).
ROA is defined as:
𝑅𝑂𝐴 = 𝑛𝑒𝑡 𝑖𝑛𝑐𝑜𝑚𝑒
𝑡𝑜𝑡𝑎𝑙 𝑎𝑠𝑠𝑒𝑡𝑠 (4)
The last measure of CFP is the market adjusted returns. The market adjusted returns of a company are its stock returns corrected for the company’s level of systematic risk. Similarly as the MTB ratio, market adjusted returns are classified as market-based measures of CFP. They reflect how shareholders value the firm and have a long-term perspective.
Market adjusted returns are calculated for each company’s stock. As a first step, the total return index for all companies in the sample is retrieved from DataStream. In addition, the total return index for the MSCI world index is retrieved. The total return index is chosen since it reflects the total return on a stock or market index. Whilst price return only includes capital gains, a total return index also accounts for dividends and assumes that all cash distributions are reinvested. Moreover, a total return index corrects for stock splits. Therefore, the total return index is an appropriate and precise measure for a stock’s or market index’ performance. The MSCI world index is chosen as the market index since it represents a diversified set of global stocks. The alternative would be to choose a local market index, however these indices include less stocks and are therefore not as diversified as a world index.
Monthly log returns are calculated as:
𝑅𝑖,𝑡𝑢𝑛𝑎𝑑𝑗 = ln ( 𝑇𝑅𝐼𝑡
𝑇𝑅𝐼𝑡−1) (5)
where 𝑅𝑖,𝑡𝑢𝑛𝑎𝑑𝑗 is the unadjusted total monthly return for stock i at time t, 𝑇𝑅𝐼𝑡 is the value of the total return index at time t and 𝑇𝑅𝐼𝑡−1 is the value of the total return index at time t-1.
Beta is estimated using the Capital Asset Pricing Model (CAPM). We use the total monthly returns for the period 2002-2015. We choose this period since it matches the period for which the ASSET4 ratings are available. In case a company is not listed for the complete period of 2002-2015, the beta is estimated based on the data that is available.
The beta is estimated using the CAPM:
𝑅𝑖,𝑡𝑢𝑛𝑎𝑑𝑗 = 𝛼𝑖+ 𝛽𝑖× 𝑅𝑡𝑚𝑘𝑡 (6)
where 𝛼𝑖 is the Jensen’s alpha of stock i, 𝛽𝑖 is the beta for stock i and 𝑅𝑡𝑚𝑘𝑡 is the return on the market index at time t.
As a last step, the estimated beta 𝛽𝑖 is used to calculate the total yearly market adjusted returns: 𝑅𝑖,𝑡𝑚𝑘𝑡−𝑎𝑑𝑗= 𝑅𝑖,𝑡𝑢𝑛𝑎𝑑𝑗− 𝛽𝑖× 𝑅𝑡𝑚𝑘𝑡
(7)
where 𝑅𝑖,𝑡𝑚𝑘𝑡−𝑎𝑑𝑗 are the total market adjusted returns.
We choose not to include the risk-free rate. This study will focus on whether relatively higher/lower levels of CSP relate to relatively higher/lower levels of CFP, and vice versa. Since the risk-free rate is identical for all stocks, including the risk-free rate will not modify this relationship.
3.3. Control variables
There are several variables that moderate in the relationship between CSP and CFP. This study employs the control variables size, risk, industry and country. This section will elaborate on how each of these four control variables affect CSP and CFP.
ROA. When using market adjusted returns as a measure of CFP however, Fama and French (1992) show that small-firm stocks have historically outperformed large-firm stocks. This can mean that the size of a firm captures factors of risk which the CAPM model does not correctly control for. Based on the findings of Fama and French (1992), we expect a negative relationship between size and CFP. We measure firm size by the value of a company’s total assets in U.S. dollars.
The second control variable is risk. In this study, we capture risk by the use of two measures: the Altman z-score and the volatility of market adjusted returns. The Altman z-score is a measure for the bankruptcy risk of a firm (Altman, 1968). The higher the Altman z-score, the lower a firm’s
bankruptcy risk. In this study, we use the Altman z-score as a measure of risk when the MTB ratio and ROA are used as measures of CFP.
The Altman z-score is calculated as:
𝐴𝑙𝑡𝑚𝑎𝑛′𝑠 𝑧 − 𝑠𝑐𝑜𝑟𝑒 = 1.2𝑋 1+ 1.4𝑋2+ 3.3𝑋3+ 0.6𝑋4+ 1.0𝑋5 (8) where 𝑋1=𝑤𝑜𝑟𝑘𝑖𝑛𝑔 𝑐𝑎𝑝𝑖𝑡𝑎𝑙 𝑡𝑜𝑡𝑎𝑙 𝑎𝑠𝑠𝑒𝑡𝑠 𝑋2= 𝑟𝑒𝑡𝑎𝑖𝑛𝑒𝑑 𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑡𝑜𝑡𝑎𝑙 𝑎𝑠𝑠𝑒𝑡𝑠 𝑋3 =𝑒𝑎𝑟𝑛𝑖𝑛𝑔𝑠 𝑏𝑒𝑓𝑜𝑟𝑒 𝑖𝑛𝑡𝑒𝑟𝑒𝑠𝑡 𝑎𝑛𝑑 𝑡𝑎𝑥𝑒𝑠 𝑡𝑜𝑡𝑎𝑙 𝑎𝑠𝑠𝑒𝑡𝑠 𝑋4= 𝑚𝑎𝑟𝑘𝑒𝑡 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑒𝑞𝑢𝑖𝑡𝑦 𝑏𝑜𝑜𝑘 𝑣𝑎𝑙𝑢𝑒 𝑜𝑓 𝑒𝑞𝑢𝑖𝑡𝑦 𝑋5 = 𝑠𝑎𝑙𝑒𝑠 𝑡𝑜𝑡𝑎𝑙 𝑎𝑠𝑠𝑒𝑡𝑠
We expect that the Altman z-score is related to both CSP and CFP. First, we expect that CSP is positively related to a firm’s Altman z-score (CSP negatively relates to bankruptcy risk). We base this expectation on Godfrey et al. (2009). Godfrey et al. (2009) argue that firms who engage in corporate social responsibility have strong relationships with their stakeholders. Those strong relationships will create goodwill. In case a negative event happens, the generated goodwill will function as an insurance and will reduce the severity of sanctions (boycotts, badmouthing etc.) that stakeholders impose on the firm. Socially responsible firms are thus less risky, since goodwill protects them from costly
punishment by stakeholders. Secondly, we expect a positive relationship between the Altman z-score and CFP (a negative relationship between bankruptcy risk and CFP). This expectation naturally follows from how the Altman z-score is calculated. Equation 8 shows that good CFP results in a high Altman z-score.
both CSP and CFP. First, we expect that the volatility is negatively related to CSP. This expectation is based on the theory of Godfrey et al. (2009) as described in the previous paragraph. Secondly, we expect a positive relationship between volatility and market adjusted returns. The CAPM argues that there is a positive relationship between risk and return. Investors require a reward for the risks they face. The CAPM model assumes that idiosyncratic risk is not rewarded. In the case of limited risk sharing however, El Ghoul et al. (2011) argue that idiosyncratic risk actually is priced. We therefore expect a positive relationship between volatility and market adjusted returns.
The third control variable is industry. Industry affects both the CSP and the CFP of a firm. First, industries differ with regard to their environmental and social impact (Endrikat, 2014). Moreover, industries face different regulations and are subject to different levels of pressure from their stakeholders. We therefore expect that industry is a determinant of CSP. Secondly, we expect that industry affects CFP. Graves and Waddock (1994) show that industries differ with regard to their level of R&D investment. R&D investments can lead to specific technological capabilities which result in long-term value for the firm. Through this channel, R&D investments positively affect CFP. In addition, R&D investments also affect CSP since they can result in innovative approaches to environmental and social issues (Endrikat, 2014). This study uses the four-digit Standard Industrial Classification (SIC) code to assign each company to an industry. Industry is controlled for by including nine dummy variables.
The last control variable is country. First, we expect that country is a determinant of CSP since companies in different countries face different levels of pressure from their stakeholders. In addition, in each country a company has to adhere to different regulation. Secondly, we expect that country affects a firm’s CFP. For example, because macro-economic conditions vary across countries. We control for country by including country dummy variables in the regression equations.
This section elaborated on the different measures for CSP and CFP used in this study. Moreover, the control variables were discussed. The next section will present the three methodologies that are applied in this study.
4. Methodology
4.1. Methodology 1: OLS including lagged values of the independent variable
The first methodology is an Ordinary Least Squares (OLS) regression that includes lagged values of the independent variable. It is the most basic methodology for testing causality and is used in both early work such as Waddock and Graves (1997) as more recent work such as Nelling and Webb (2009). Here, the dependent variable is regressed on an n number of lags of the independent variable. The control variables size, risk, industry and country are included. The regression is specified as:
𝐶𝑆𝑃𝑖,𝑡 = 𝛼 + ∑ 𝛽𝑗𝐶𝐹𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + 𝛾𝑆𝐼𝑍𝐸𝑖,𝑡+ 𝛿𝑅𝐼𝑆𝐾𝑖,𝑡+ 𝜖𝐼𝑁𝐷𝑖+ 𝜁𝐶𝑂𝑈𝑖+ 𝑢𝑖,𝑡 (R1) and: 𝐶𝐹𝑃𝑖,𝑡= 𝛼 + ∑ 𝜂𝑗𝐶𝑆𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + 𝜃𝑆𝐼𝑍𝐸𝑖,𝑡+ 𝜗𝑅𝐼𝑆𝐾𝑖,𝑡+ 𝜄𝐼𝑁𝐷𝑖+ 𝜅𝐶𝑂𝑈𝑖+ 𝑢𝑖,𝑡 (R2)
where in our case, 𝐶𝑆𝑃𝑖,𝑡 is one of the CSP measures for company i at time t (generic measures or material measures), 𝐹𝑃𝑖,𝑡−1 is one of the measures of CFP for company i, at time t-1 (MTB, ROA or market adjusted returns), 𝑆𝐼𝑍𝐸𝑖,𝑡 is defined as total assets in U.S. dollars for company i at time t, 𝑅𝐼𝑆𝐾𝑖,𝑡 is the risk measure for company i at time t (Altman z-score for the CFP measures ROA and MTB, and volatility of market adjusted returns for the CFP measure market adjusted returns), 𝐼𝑁𝐷𝑖 represents the set of nine dummy variables that control for industry, 𝐶𝑂𝑈𝑖 represents the set of dummy variables that control for country. 𝑢𝑖,𝑡 is the error term for company i at time t.
4.2. Methodology 2: OLS fixed effects including lagged values of the independent variable
To control for unobservable characteristics that differ between firms, an OLS regression with cross-sectional fixed effects will be estimated. Due to the fact that the fixed effects approach takes account for cross-sectional differences, the terms that do not vary over time are cancelled out (industry and country). OLS fixed effects regressions have mostly been employed in more recent work on the CSP-CFP relationship such as in studies by Nelling and Webb (2009).
The model is represented as:
𝐶𝑆𝑃𝑖,𝑡 = 𝛼 + ∑ 𝜆𝑗𝐶𝐹𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + 𝜇𝑆𝐼𝑍𝐸𝑖,𝑡+ 𝜈𝑅𝐼𝑆𝐾𝑖,𝑡+ 𝜀𝑖+ 𝑢𝑖,𝑡 (R3) and: 𝐶𝐹𝑃𝑖,𝑡= 𝛼 + ∑ 𝜉𝑗𝐶𝑆𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + 𝜊𝑆𝐼𝑍𝐸𝑖,𝑡+ 𝜋𝑅𝐼𝑆𝐾𝑖,𝑡+ 𝜀𝑖+ 𝑢𝑖,𝑡 (R4)
4.3. Testing for causality
Both methodologies discussed require the development of a rule that determines whether causality is present or not. When we conduct the analysis for a lag-length of one this is rather simple. In this case, causality is present when the coefficient on this single lag is significantly different from zero. When employing other lag-lengths however, we require a rule that specifies when causality is present. Since the majority of earlier work uses a lag-length of one, this issue has not been addressed in previous literature. In this study, we define causality as follows. For a lag-length of three we require two out of three lags to be significantly different from zero. For a lag-length of six we require three out of six lags to be significantly different from zero. (Section 5.1.2 and section 5.2.2 elaborate on why we choose to work with lag-lengths of three and six.)
4.4. Methodology 3: VAR modelling
When analyzing the causal relationship between CSP and CFP, we need to take into account the concept of endogeneity. Endogeneity means that the value of both the dependent variable and the independent variable are determined in the system (Brooks, 2008) The independent variable is thus not truly independent, as its value depends on values of the dependent variable. With regard to the CSP-CFP relationship, endogeneity entails that neither CSP nor CSP-CFP are independent variables. It is likely that CSP affects CFP, but that CFP also affects CSP.
Standard OLS regression and OLS regression with fixed effects do not account for endogeneity. Both methods assume that the independent variable is exogenous. A way to incorporate endogeneity in the analysis is to use Vector Autoregressive (VAR) models. VAR models allow two equations to be estimated simultaneously. There are some advantages as well as disadvantages involved when
working with VAR models. The obvious advantage is that no restrictions need to be made with regard to which variable is labeled as either exogenous or endogenous. Both variables can be simultaneously determined (Brooks, 2008). The major disadvantage of VAR models is that they can become a-theoretical. It can be challenging to interpret the coefficients when many lags are included.
This study will use a system of two equations to analyze the CSP-CFP relationship. The VAR regression that will be estimated is:
𝐶𝑆𝑃𝑖,𝑡 = 𝛼 + ∑ 𝜛𝑗𝐶𝑆𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + ∑ 𝜌𝑗𝐶𝐹𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + 𝜚𝐼𝑁𝐷𝑖+ 𝜎𝐶𝑂𝑈𝑖+ 𝑢1,𝑡 (R5) 𝐶𝐹𝑃𝑖,𝑡= 𝛼 + ∑ 𝜍𝑗𝐶𝐹𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + ∑ 𝜏𝑗𝐶𝑆𝑃𝑖,𝑡−𝑗 𝑛 𝑗=1 + 𝜐𝐼𝑁𝐷𝑖+ 𝜑𝐶𝑂𝑈𝑖+ 𝑢2,𝑡 (R6)
4.4.1. Granger causation
In order to test for causality using the VAR methodology, we use Granger causality tests. The idea behind Granger causality is that a “variable X is said to Granger-cause variable Y if the past values of X help to explain Y even after the impact of Y’s lagged values are taken into account and vice versa” (Pätäri et al., 2014, p. 145). If there is Granger causation running only from X to Y, we speak of unidirectional causation. If there is Granger causation running from both X to Y and vice versa, we speak of bidirectional causation (Brooks, 2008). Granger causation tests can be conducted on a system of equations such as equation R5 and R6. Applied to this study, CFP is said to Granger cause CSP if the coefficients on all the lagged values of CFP as a group are significantly different from zero. Granger causation tests are thus block significance tests on all lagged values of the dependent variable. It is important to note that Granger causality reflects a correlation between past values of one variable and the current value of the other variable. One cannot use Granger causality to predict whether a rise in X increases or decreases Y (Pätäri et al., 2014).
4.4.2. Stationarity
All variables that enter a VAR equation need to be stationary (Brooks, 2014). A variable is stationary when the distribution of its values does not change as time progresses. This entails that the probability that value y falls within a particular interval remains the same over time. A variable is said to be stationary when it satisfies three conditions for t = 1 to infinity. These conditions are (Brooks, 2014):
The mean of the variable is constant over time. (C1)
The variance of the variable is constant over time. (C2)
The covariance between two time periods depends only on the lag between two time periods, but not on the actual value of time at which the covariance is computed.
(C3)
cause their results to vary (Pätäri et al., 2014). The null hypothesis for all four tests is that the variable contains a unit root.
4.4.3. Lag-length
When working with VAR models, we need to decide on which lag-length to use. The most common approach to determine the appropriate lag-length for a VAR model is to make use of information criteria (Brooks, 2014). Information criteria embody two factors. The first factor is a term which relates to the residual sum of squares of the regression. The second factor reflects a penalty for the loss of degrees of freedom from adding extra parameters to the regression. The optimal lag-length is the lag-length for which the information criterion is minimized. Adding an extra lag to the model only decreases the information criterion when the decline in the residual sum of squares outweighs the penalty related to the loss in degrees of freedom (Brooks, 2014).
There are various information criteria. The three that are most frequently used are Aikaike’s
information criterion (AIC), Schwarz’s information criterion (SC) and the Hannan-Quinn information criterion (Brooks, 2014). The different information criteria generally minimize for different lag-lengths. All three information criteria have their own advantages and disadvantages, and no single criterion is superior to the others (Brooks, 2014). This study will test for the appropriate lag-length analyzing all three information criteria.
4.5. Proposed methodology compared to the methodologies applied in previous work in the field
The methodologies presented above have been used in previous studies on the CSP-CFP relationship. What makes this study unique however, is that it takes a systematic approach in applying these methodologies. This systematic approach entails that we will analyze the relationship using three methodologies, three measures of CFP and three different lag-lengths .In addition, we analyze the relationship for both the generic measure of CSP and the material measure of CEP.
5. Results and discussion
The causal relationship between CSP and CFP is analyzed using two measures of CSP: the generic measures and the material measures. Section 5.1 starts with a discussion of the results from the analysis that uses the generic measures of CSP. Section 5.2 discusses the results from the analysis that uses the material measure of CEP. For both measures of CSP, the analysis is conducted by applying the three methodologies and by using the three measures of CFP. In addition, the regressions will be estimated for different lag-lengths. It is worth noting that this chapter presents and discusses the main results of this study. The calculations that underlie these results can be found in appendices F to K.
5.1. Results of the analysis that uses the generic measures of CSP
This part of the analysis uses the generic measures of CSP. These measures are the total ESG score (CSPTOT), the environmental pillar score (ENV), the social pillar score (SOC) and the governance pillar score (GOV).
5.1.1. Stationarity
Before we conduct the regression analysis we test whether all variables in the regression equation are stationary. The dependent, independent and control variables are tested for stationarity and table D1 in the appendix shows that there is no evidence that one of the variables contains a unit root. For all variables and under all tests, the null hypothesis stating that the variable has a unit root is rejected. We conclude that all variables can be treated as stationary. This means that the stationarity condition as discussed in section 4.4.2 is satisfied and that we can proceed with using the variables in the regression analysis.
5.1.2. Lag-length
As discussed in section 4.4.3, it is important to specify the regression equation correctly. Therefore, we have to decide which lag-length to use. In order to do so, the VAR model as represented by equation R5 and R6 is used as the reference model. Lags in the OLS and in the OLS fixed effects regressions are adjusted accordingly. Table E1 in the appendix presents the appropriate lag-lengths as determined by the three information criteria. Table E1 shows that the information criteria are
CSP and CFP such as Nelling and Webb (2009), Makni et al. (2009) and Waddock and Graves (1997). These studies all employ lag-lengths of one.
Results of the analysis using a lag-length of six are presented in the main text. The reason to present the results for this lag-length is that the information criteria most often indicate that a lag-length of six is appropriate. In addition, table E2 shows that a lag-length of six is most appropriate for the material measure of CSP as well. Presenting results using a lag-length of six in both analyses allows us to easily compare results. The results of the regressions with a lag-length of one and three are presented in appendix F and G respectively.
5.1.3. The fit of the model
In this section we will analyze the fit of the model. A total of 72 regressions are estimated. The results are presented in appendix H. First, we discuss the adjusted R-squared values. Secondly, we elaborate on the coefficients for the control variables size and risk.
Table H1, H2 and H3 in the appendix present the results of applying the OLS methodology, the OLS methodology with fixed effects and the VAR methodology respectively. In general, we observe that the value of adjusted R-squared is higher for the models that employ CSP as the dependent variable than for the models that employ CFP as the dependent variable. This means that the independent variables in the first model have more explanatory power than the independent variables in the latter model. In addition, we find that both the OLS model with fixed effects and the VAR models have a better fit with the data than the standard OLS model. For the OLS model with fixed effects this is due to the fixed effect component explaining a large part of the variation in the dependent variable. This means that unobserved differences across firms determine the variation in either CSP or CFP. For the VAR model, the lagged values of the dependent variable have high explanatory power for the
variation in the dependent variable. This entails that past CSP explains current CSP, and that past CFP explains CFP.
explanatory power to the model. Size and risk are both determinants of a company’s CSP and are therefore appropriate to include as control variable.
In the models where CFP is the dependent variable, size in general has a negative coefficient. The negative coefficient is mostly significant when CFP is measured by market adjusted returns. This supports the argument of Fama and French (1992) that smaller firms outperform larger firms when financial performance is measured by stock returns. For the other measures of CFP, the coefficient on size does not have significant explanatory power. This means that size does not explain any variation in the MTB ratio or in ROA. The coefficients on risk show a distinct pattern. When the Altman z-score is employed as a measure of risk, the coefficient is generally positive. This is in line with our
expectation that a higher Altman z-score (lower risk) relates to higher CFP. When the volatility of stock returns is taken as a measure of risk the coefficient is significantly negative. This result is not in line with the argument of El Ghoul et al. (2011), who argue that idiosyncratic risk is positively priced when there is limited risk-sharing.
5.1.4. Overview of results for a lag-length of six
This section presents an overview of the results for the analysis that uses a lag-length of six. Overview table 1 presents a summary of the results for the 72 regressions estimated. Results are shown for the different methodologies applied and the different measures of CFP. For each methodology, table 1 presents the amount of significant lags and whether causality is present. In the cases where causality is present, the table also reports on the sign of the causal relationship. Table 1 will be discussed as follows. First, we compare the results for the different methodologies. Secondly, we analyze the results for the different measures of CFP.
First, we compare results for the three methodologies. We observe that results across methodologies are inconsistent. Consistency is found in only nine of the 24 relationships. In eight of these nine cases, there is consistency regarding the fact that there is no relationship between CSP and CFP. For the ROASOC relationship all methodologies indicate the existence of a positive causal relationship, which is in line with available funds hypothesis. Results across methodologies differ for the other 15 relationships. Of special interest are some highly inconsistent cases. In these cases, all three
with the findings of Nelling and Webb (2008). Nelling and Webb (2008) use different methodologies to investigate the causal relationship between CSP and CFP. Similarly as this study, they find that outcomes are inconsistent across the different methodologies they apply. Because results vary across methodologies, we are not able to find consistent evidence for hypothesis H1 to H4.
Table 1 Overview of results
This table presents a summary of the results for all estimated regressions. Results are shown separately for each methodology. The three methodologies are OLS, OLS with fixed effects and VAR modelling. Causality is indicated as present when 3 out of 6 lags are significant when applying the first two methodologies. Granger causation is used to test for causality when the VAR methodology is applied.
OLS OLS FIXED EFFECTS VAR
# of significant lags Causality Sign of significant lags # of significant lags Causality Sign of significant lags Granger causality Sign MTB → CSPTOT 0 No - 2 No - No - CSPTOT → MTB 1 No - 1 No - No -
ROA → CSPTOT 4 Yes Positive 6 Yes Mixed Yes Positive
CSPTOT → ROA 1 No - 1 No - No -
RET → CSPTOT 6 Yes Negative 4 Yes Negative Yes Positive
CSPTOT → RET 2 No - 2 No - Yes Negative
MTB → ENV 0 No - 2 No - No -
ENV → MTB 2 No - 1 No - No -
ROA → ENV 0 No - 2 No - Yes Positive
ENV → ROA 1 No - 3 Yes Negative No -
RET → ENV 6 Yes Negative 3 Yes Negative Yes Positive
ENV → RET 1 No - 2 No - Yes Negative
MTB → SOC 2 No - 0 No - No -
SOC → MTB 0 No - 0 No - No -
ROA → SOC 6 Yes Positive 4 Yes Positive Yes Positive
SOC → ROA 2 No - 2 No - No -
RET → SOC 6 Yes Negative 2 No - Yes Positive
SOC → RET 1 No - 1 No - Yes Negative
MTB → GOV 0 No - 0 No - Yes Positive
GOV → MTB 0 No - 2 No - Yes Negative
ROA → GOV 5 Yes Positive 0 No - Yes Positive
GOV → ROA 1 No - 1 No - Yes Mixed
RET → GOV 6 Yes Negative 6 Yes Mixed Yes Positive
Secondly, we analyze the results for the different measures of financial performance: the MTB ratio, ROA and the market adjusted returns. Table 1 shows that there is no clear pattern with regard to how different measures of CFP relate to CSP. We observe that in the majority of cases there is no evidence of a causal relationship when MTB is used as the measure of CFP. Only in the case where MTB is related to the governance dimension of CSP, one methodology finds a causal relationship. This finding is not in line with hypothesis H7 which predicts a positive relationship between the MTB ratio and CSP. In models where ROA is the measure of CFP, the methodologies indicate a causal relationship more often. Evidence of causality is especially found in relationships where ROA is the independent variable. The sign of the causal relationships is either mixed or positive, which is partly in line with hypothesis H8 which argues that ROA and CSP are positively related. When ROA is the dependent variable there are fewer significant relationships. Here, the sign of the significant relationships is either mixed or negative, which contradicts hypothesis H8 of this study. Lastly, the market adjusted returns are analyzed. Evidence of causality is present for several relationships and especially in cases where RET is the independent variable. For the models RETCSPTOT and RETENV we find a negative relationship under both OLS and OLS with fixed effects. When applying the VAR methodology however, the relationship is positive. In cases where RET is the dependent variable, causality is mostly found when applying the VAR methodology. The sign of the relationship is either mixed or negative. We conclude that there is no distinguishable pattern with regard to how different measures of CFP relate to CSP. We find no consistent support for hypotheses H7 to H9 of this study. This finding is in line with the outcomes of previous work in the field. For example, many studies have used ROA to measure CFP (such as: Makni et al. (2014), Nelling and Webb (2008), Pätäri et al. (2014) and
Waddock and Graves (1997)). Results of these studies however, do not provide consistent evidence as to how ROA is related to CSP.
The same analysis is performed for the lag-lengths one and three. Results of these analyses can be found in appendix F and G. Table F4 and table G4 are the overview tables for length one and length six respectively. We can make several remarks on the findings for the analysis regarding lag-length one and three. First, we observe inconsistency in the results across the methodologies applied. In the extreme cases, results from the application of one methodology indicate a positive causal relationship while another methodology indicates a negative causal relationship. Secondly, we cannot find a clear pattern regarding how the CSP-CFP relationship changes for different measures of CFP. The two findings are in line with our findings from the analysis that uses a lag-length of six.
H4 of this study. The second conclusion entails that we do not find evidence for hypotheses H7 to H9 either.
5.1.5. Results related to the hypotheses of this study
In this section we combine results across the different dimensions of CSP, the three measures of CFP and across the three methodologies. Table 2 relates these combined results to hypotheses H1 to H6. The table presents how many regressions provide significant outcomes in support of the hypotheses. Each hypothesis is tested by estimating 108 regressions.
Table 2
Results related to hypotheses H1 to H4
This table summarizes the results of the analyses across the different dimensions of CSP, the different measures of CFP and across the three methodologies. It list hypotheses H1 to H4 and indicates by how many significant regressions these hypotheses are supported. The column ‘total’ shows how many regressions of the 108 regressions estimated are significant. Results are shown for the analyses employing a lag-length of one, three and six respectively.
Hypothesis Expected relationship Number of confirmed relationships
One lag Three lags Six lags Total
H1: Social impact CSP → + CFP 7 5 0 12 H2: Trade-off CSP → - CFP 18 3 6 27 H3: Available funds CFP → + CSP 8 10 13 31 H4: Managerial opportunism CFP → - CSP 14 10 6 30 H5: Virtuous circle CSP ↔ + CFP 6 3 0 9 H6: Negative synergy CSP ↔ - CFP 12 0 1 13
First, table 2 shows that results are inconsistent across the different lag-lengths employed. For example, some evidence in support of the trade-off hypothesis is found when we use a lag-length of one (18 significant relationships). However, this evidence almost completely disappears when we use a lag-length of three (3 significant relationships). This finding is in line with our expectation that the choice for lag-length can affect the results on the CSP-CFP relationship. Secondly, we observe that the majority of regressions do not support the existence of a causal relationship between CSP and CFP. This finding is in line with the findings of Makni et al. (2009) and Nelling and Webb (2008), who do not find evidence of a causal relationship either. The fact that we do not find strong evidence for a causal relationship between CSP and CFP can mean that our hypotheses do not correctly capture the CSP-CFP relationship.
In the cases where table 2 shows evidence of a causal relationship, the trade-off hypothesis (H2), the available funds hypothesis (H3) and the managerial opportunism hypotheses (H4) are mostly
with the ideas of Friedman (1970), who argues that the costs of corporate social responsibility are substantial while it has no benefits. The evidence for the available funds hypothesis suggests that managers use slack resources to invest in corporate social responsibility. Lastly, the evidence for the managerial opportunism supports the theory that after bad financial performance managers want to cover up the damage by investing in corporate social responsibility.
Table 3 relates the results to hypotheses H7 to H9. Results are combined across methodologies and dimensions of CSP. The table lists hypotheses H7 to H9 and indicates how many times evidence is found for one of these hypotheses. Each hypothesis is tested by estimating 72 regressions.
Table 3
Results related to hypotheses H7 to H9
This table summarizes the results of the analyses across the different dimensions of CSP and across the three methodologies. It lists hypotheses H7 to H9 and shows how many regressions indicate evidence of either a positive, negative or mixed sign for the CSP-CFP relationship. The column 'total' shows how many regressions of the 72 regressions estimated indicate either a positive, negative or mixed sign for the CSP-CFP relationship. Results are shown for the analyses employing a lag-length of one, three and six respectively.
Hypothesis Number of confirmed relationships
One lag Three lags Six lags Total
+ +/- - + +/- - + +/- - + +/- -
H7: CSP and MTB + 1 0 12 1 0 2 1 0 1 3 0 15
H8: CSP and ROA + 10 0 4 8 2 2 8 2 1 26 4 7
H9: CSP and RET +/- 4 0 16 6 3 8 4 3 10 14 6 34
Table 3 shows that the results are inconsistent across the different lag-lengths employed. For example, for a lag-length of one we find that 12 regressions support a negative relationship between CSP and MTB. For the lag-length of six however, we find that this number of significant regressions decreases to one. This finding again supports our expectation that the use of different lag-lengths affect results. Secondly, we find that the majority of regressions indicate that there is no relationship between CSP and CFP. This can mean that our hypotheses H7 to H9 do not adequately capture the relationship between CSP and different measures of CFP.
have a negative sign. This does not support hypothesis H9, which suggests that the relationship should be ambiguous. According to the framework of Dam and Scholtens (2015) the finding of negative returns means that the effect of a higher market-to-book ratio is stronger than the effect of higher return on assets. This line of reasoning however, does not correspond with our findings on the CSP-MTB and CSP-ROA relationships. We find that higher CSP relates to lower market-to-book ratios which implies higher expected returns. In addition to this, we find that CSP relates to higher ROA which also implies higher returns. When combining these two effects we would expect positive market adjusted returns.
From this section we conclude that (1) results are inconsistent for the different lag-lengths employed and (2) that the majority of the estimated regressions find that there is no significant relationship between CSP and CFP. In case we find evidence of a causal relationship, this evidence is mostly in support of the trade-off hypothesis, the available funds hypothesis or the managerial opportunism hypothesis. Moreover, we find that the sign of the significant relationships between CSP and MTB is mostly negative, for the significant relationships between CSP and ROA the sign is mostly positive and for the significant relationships between CSP and RET the sign is mostly negative.
5.2. Results of the analysis that uses the material measures of CEP
The main findings from the previous analysis are that results are inconsistent across methodologies and across the different lag-lengths employed. Moreover, using the generic measures of CSP we cannot distinguish a clear pattern with regard to how CSP is related to different measures of CFP. We are interested whether we find similar results when using the material measure of CEP. The material measure addresses two limitations of the generic measure, as discussed in section 3.1.1. Firstly, the material measure addresses the concern related to the normalization of the z-scores. Since ASSET4 normalizes its z-scores through the use of a scaling divisor, changes in CSP scores can be caused by a change in the value of the scaling divisor. Secondly, the material measure addresses the concern that the generic measure is largely process-based.
This part of the analysis will make use of the self-constructed material measures of CEP (CSPENV). In addition to the self-constructed measure for CEP, we estimate regressions for the underlying variables from which this score is constructed. These variables are CO2 divided by sales (CO2), energy divided by sales (energy) and water use divided by sales (water).
5.2.1. Stationarity
induce stationarity this variable is first-differenced. Panel B of table D2 shows that when we take the first-difference of size, the variable meets the stationarity condition.
Despite the fact that the tests do not provide evidence for the presence of a unit root, various VAR systems do not pass the stability condition. The estimated VAR is stable if all roots have modulus less than one and lie inside the unit circle. Table D3 in the appendix shows that for some systems of equations there are unit roots that lie outside the unit circle. (Results are not presented for the aggregate measure of CEP since the VAR equations including this variable all pass the stability condition.) To stabilize the VAR equations that do not pass the stability test, we take first-differences of all variables. Panel B of table D2 presents the results of the unit root tests on first-differenced data. There is no evidence that one of the series contains a unit root. Table D3 shows that first-differencing the data stabilizes various VARs that were characterized as unstable before. The only VAR that remains unstable is the VAR that includes the variable recycled waste divided by total waste. We exclude this variable from the analysis since any inferences that we will make based on this variable are unreliable.
The fact that we take first-differences of the data makes the interpretation of the results different from the interpretation in the previous analysis presented in section 5.1. Here we analyzed whether
relatively higher/lower levels of CSP relate to relatively higher/lower levels CFP and vice versa. For the variables CO2 divided by sales, energy divided by sales and water divided by sales we now analyze whether changes in CEP relate to changes in CFP and vice versa.
5.2.2. Lag-length
The information criteria are analyzed to determine the appropriate lag-length for the analysis. The VAR model represented by equation R5 and R6 is used as the reference model. Lag-lengths for the other methodologies are adjusted accordingly. Table E2 in the appendix shows that the information criteria most often indicate that the analysis should be conducted for an equation with a lag-length of six. In addition, the analysis is conducted for a lag-length of one and three. By doing this we ensure comparability to the previous analysis. Results of the analysis containing six lags are presented in the main text.
5.2.3. The fit of the model
In this section we will analyze the fit of the model. A total of 72 regressions are estimated. The results are presented in appendix K. First, we analyze the adjusted R-squared values. Next, we will elaborate on the coefficients for the control variables size and risk.
observe large differences with respect to the values of adjusted R-squared. For the standard OLS model which uses CSP as the dependent variable, we observe values of adjusted R-squared close to zero, ranging from -0.002 to 0.006. The fit of the OLS model with CFP as the dependent variable is not much better. The values of adjusted R-squared are more dispersed and range from -0.005 to 0.394. A negative value of adjusted R-squared indicates that the model fits the data poorly. In addition, it can indicate the inclusion of too many explanatory variables (Brooks, 2008). We observe that the values of adjusted R-squared on average increase when the OLS fixed effects and VAR methodologies are applied. Similarly as in the previous analysis, for the OLS fixed effects model this is caused by the fixed effects component explaining a lot of variation in the dependent variable. This means that unobserved differences across firms determine both CEP and CFP. For the VAR model, lagged values of the dependent variable explain a large part of the variation in the dependent variable. This means that past CSP relates to current CSP, and that past CFP relates to current CFP.
Secondly, we assess the coefficients for the control variables size and risk. When CSP is the dependent variable, we observe that the variable size takes on positive values in the majority of cases. A positive coefficient supports the theory of Endrikat et al. (2014) that large firms are to be associated with higher CSP. The coefficient is significant for roughly half of the regression equations when applying the standard OLS methodology. However, the significance of the variable size disappears completely when applying the fixed effects methodology. This means that in this model, size does not explain any variation in CSP. Next, we analyze the coefficient for risk. We observe that in models where CSP is the dependent variable, the coefficient for risk takes on both positive and negative values which are overall not significant. The fact that the coefficients are not significant means that risk is not related to CSP. This does not support the theory of Godfrey et al. (2009), that corporate social responsibility mitigates risk.
