2.1 Dutch occupational pension sector

Due to the quasi-mandatory status, the participation rate in the Netherlands is high: around 90 percent of the workforce participates in one or more occupational pension schemes. For some industries, mandatory participation exists, which implies that all companies – and therefore all employees – in such an industry are required to join an industry-wide pension fund. Besides industry-wide pension funds, there are also pension funds for the employees of a specific company (corporate pension funds) or a particular profession (professional group pension funds).

In the Netherlands, pension funds are legally independent, non-profit organizations whose task is to execute the pension scheme that representatives of employers and employees have negotiated as part of labor compensation. The board of trustees is responsible for managing the assets and administering the benefits and consists of employee representatives (labor unions), employer representatives, and external experts. This board formally sets the investment policy and strategic asset allocation, with the help of several advisory councils, consultants, and investment advisors. Most pension funds outsource the implementation of the investment policy to one or more asset management firms. Besides implementing the investment policy, asset management firms can also act as an advisor to the pension fund when developing the SI policy because they often possess more expertise on this topic. Specifically for engagement, pension funds sometimes use ESG service providers who conduct the engagement conversations regarding sustainable investing independently from the asset manager.

2.2 Legislation

A number of features in the legislation on Dutch occupational pension funds are relevant to sustainable investing. An important section of the Dutch Pension Act states that pension funds should invest their assets in the sole interest of their beneficiaries. This is the so-called prudent person rule.Footnote ¹⁰ The prudent person rule is an open norm and does not contain quantitative investment restrictions.Footnote ¹¹ To invest the assets in the best interest of beneficiaries, pension funds should take into account the beneficiaries' sustainability preferences. Moreover, Section 135 of the Dutch Pension Act states that pension funds should specify in their annual report how they incorporate ESG criteria in their investment policy.

In addition to Dutch legislation, European legislation is also relevant. IORP II states that pension funds can include ESG criteria in the prudent person rule as long as the application of ESG criteria does not harm the financial interests of the beneficiaries.Footnote ¹² Another IORP II requirement is the incorporation of ESG risks in risk management.Footnote ¹³ The only hard requirement regarding sustainable investing for Dutch pension funds is the prohibition of cluster munition investments, which has been in place since 2013.Footnote ¹⁴

Additional requirements are applicable as of March 2021, resulting from the European Sustainable Finance Disclosure Regulation (SFDR) regarding the provision of information on the sustainability of investments. Pension funds must explain to what extent they integrate ESG risks in their investment process. Moreover, they must indicate whether they take the adverse impacts of investment decisions on ESG factors into account.

2.3 Sustainable investment initiatives

Besides the legal requirements, there has also been a rapid increase in voluntary initiatives to stimulate sustainable investing. Such initiatives aim to enable and reinforce the development of a sustainable financial system by promoting ESG integration into investment decisions or transparent disclosures. For example, the UNEP FI comprised several international programs, including the well-known PRI. This program is an UN-supported initiative founded in 2006 by some of the largest institutional investors to stimulate the incorporation of ESG factors into investment practices. Other programs of the UNEP FI are the Principles for Responsible Banking (PRB), the Collective Commitment to Climate Action (CCCA), the Principles for Sustainable Insurance (PSI), and the Net-Zero Asset Owner Alliance (NZAOA). The best-known initiative in the Dutch pension fund sector is the IRBC initiative to identify, prioritize, and address ESG-related risks.Footnote ¹⁵ A group of pension funds signed it at the end of 2018. The initiative aims to bring the investment policy into line with the OECD Guidelines and UNGPs. Another national example is the commitment of a group of financial institutions to the climate goals of the Dutch government in 2019. They agreed to measure the CO₂ emissions of their investments and to publish their CO₂ reduction goals as of 2022.

In this paper, we focus on one particular SI initiative that many Dutch pension funds embraced: the IRBC initiative. The pension funds participating in this initiative made joint arrangements with NGOs, labor unions, and the government regarding integration into the investment policy, outsourcing, monitoring, and reporting. For example, they agreed that the SI policy should include an explanation of how sustainability is integrated into the various asset classes in which the pension fund invests. Moreover, the pension fund should disclose its approach toward voting and engagement and provide its stakeholders with information on which companies are excluded. This raises the question of whether pension funds that sign such an SI initiative enhance their SI policy more than non-signatories. We hypothesize that signatories enhance their SI policy more than non-signatories. The SI measures include the integration of sustainable investing into various asset classes and the implementation of different SI strategies. The measures will be described in more detail in Section 5.2.

2.4 Sustainable investment motives

In this section, we discuss the motivation of pension funds to invest sustainably. Pension funds can have financial and moral objectives. Other possible motives are reputational risk and legislation.

The first motive for sustainable investing can be driven by financial objectives. As discussed briefly in the introduction, there is growing recognition that climate-related risks are a source of financial risk. Companies with a positive impact on society may be more likely to attract customers and employees and avoid potential environmental fines or regulatory intervention. These companies generate higher risk-adjusted returns if these benefits are not fully priced (Edmans and Kacperczyk, Reference Edmans and Kacperczyk2022; Edmans, Reference Edmans2023). As a result, pension funds can decide to invest sustainably based on financial objectives.

Moral objectives can drive the second motive for sustainable investing. It can be a result of the perceived moral obligation of a pension fund to contribute to a sustainable world or a reflection of the preferences of the beneficiaries of the pension fund. Research shows that most pension participants have strong preferences for sustainability even at the expense of lower financial performance (Delsen and Lehr, Reference Delsen and Lehr2019; Bauer et al., Reference Bauer, Ruof and Smeets2021). Especially in the context of the Dutch pension sector, in which beneficiaries are not able to switch their pension provider, pension funds have a strong responsibility to ensure that they act in the best interest of their beneficiaries.

A third motive is a concern about reputational risk. As mentioned in the introduction, there is societal and political pressure on pension funds, and several examples exist of protesters pushing pension funds to divest. Pension funds are aware that the material consequences of their investments cause lasting reputational damage. Peer pressure and benchmarking can also accelerate the activities of a pension fund in the SI domain. An example is the VBDO Benchmark for Responsible Investment by Pension Funds, which compares sustainable investing by the 50 largest pension funds in the Netherlands.Footnote ¹⁶

Finally, there are legal requirements regarding sustainable investing, as discussed in Section 2.2, which can stimulate (or in the future possibly force) sustainable investing by pension funds.

2.5 Sustainable investment strategies

In this section, we discuss different SI strategies. There are several strategies to realize sustainable investing. We distinguish the following strategies in this paper: divestment, ESG integration, screening, public engagement, and private engagement. It is noteworthy, however, that it is not always possible to distinguish clearly between these five investment strategies due to some overlap. The first strategy is divestment (or exclusion), in which a pension fund excludes companies or projects with a negative (environmental) impact from the investment portfolio. Many examples exist of pension funds that publicly declare their divestment from particular industries, such as the tobacco, nuclear weapons, and fossil fuels industries. In the Netherlands, some Dutch pension funds recently announced that they would stop investing in fossil fuel producers.Footnote ¹⁷ There is disagreement in the literature on the effectiveness of a divestment strategy. For instance, Choi et al. (Reference Choi, Gao, Jiang and Zhang2021) posit that divestment pushes companies to adopt climate-friendly policies and decrease carbon footprints, but Berk and van Binsbergen (Reference Berk and van Binsbergen2021) conclude that ESG divestiture strategies have little impact on the cost of capital and will likely have little impact in the future.

A second strategy is integrating ESG criteria into the investment process. The key objective of this strategy is to improve the risk-adjusted return of investments. When determining the strategic asset allocation, for instance, financial information is complemented by sustainability information. Since this strategy is quite broad, the exact implementation of this strategy may differ between pension funds. There is no single view in the literature on the impact of, for instance, climate risks on the risk-adjusted return of investments. Some papers provide evidence that carbon risk is starting to be priced in the market (e.g., Boermans and Galema, Reference Boermans and Galema2020; Bolton and Kacperczyk, Reference Bolton and Kacperczyk2021; Ilhan et al., Reference Ilhan, Sautner and Vilkov2021). Bolton and Kacperczyk (Reference Bolton and Kacperczyk2021) find higher returns for stocks with higher total CO₂ emissions. This evidence indicates that investors demand compensation for carbon emission risk. However, Sautner et al. (Reference Sautner, Van Lent, Vilkov and Zhang2023) do not find a positive risk premium for climate change exposure and Faccini et al. (Reference Faccini, Matin and Skiadopoulos2023) find that transition and physical risks that take longer to materialize are not yet priced. Integrating sustainability can be done by, for example, tilting portfolios toward certain Sustainable Development Goals or mandates with a small, selected number of highly sustainable companies.

A third strategy is screening. The key objective of screening is to improve the portfolio performance based on specific ESG criteria. Screening is the process of selecting investments based on these criteria. There are several screening approaches in practice. For example, under negative (or exclusionary) screening, certain sectors or companies that fail to meet specific ESG criteria are excluded.Footnote ¹⁸ In the case of positive screening, certain sectors or companies are selected based on their positive (or best-in-class) ESG performance relative to industry peers. With norm-based screening, companies that do not adhere to widely accepted norms of business conduct are excluded. Heinkel et al. (Reference Heinkel, Kraus and Zechner2001) and Gollier and Pouget (Reference Gollier and Pouget2014) predict with an equilibrium model that companies will be incentivized to implement reforms when a significant fraction of investors apply the same screening approach. Opposing conclusions exist in the empirical literature on the impact of screening approaches on asset prices. For example, Hong and Kacperczyk (Reference Hong and Kacperczyk2009) show that sin stocks have depressed prices relative to otherwise comparable stocks.

The fourth and fifth strategies are two types of engagement. Engagement is the process of shareholders influencing corporate decision-making. The objective of engagement is to encourage companies to adopt more sustainable practices. In this paper, we distinguish two types of engagement: public and private engagement. Investors can engage in active ownership strategies by voting on and sponsoring shareholder resolutions (public engagement) or by directly communicating with companies (private engagement) via meetings, calls, or letters. Diaz-Rainey et al. (Reference Diaz-Rainey, Griffin, Lont, Mateo-Márquez and Zamora-Ramírez2024) show that climate-related shareholder resolutions are associated with an increase in firms' environmental performance. Bauer et al. (Reference Bauer, Derwall and Tissen2023) provide evidence that firms targeted by successful material private ESG engagements significantly outperform their peers.

Many pension funds use a combination of different SI strategies with fundamentally different motivations. For example, a pension fund can exclude certain investments based on a moral objective while simultaneously electing a long-term value-seeking strategy through ESG integration (Hammond et al., Reference Hammond, Maurer and Mitchell2023). One of the five SI measures, the variety measure, counts the number of SI strategies each pension fund implements. Section 5.2 describes this measure in more detail. Before describing the measures in more detail in Section 5, Section 4 first explains how sustainable investing can be measured.

5.1 Documents

In the analysis, we use qualitative data from annual reports and statements of investment principles of 160 Dutch pension funds from 2016 to 2021. We employ web scraping to collect these documents efficiently (see Figure 1). However, some annual reports are unavailable online, in which case we collect them via DNB archives. In total, we process more than 1,000 documents. We calculate the five SI measures using the annual reports and by exploiting various NLP techniques. We use the statements of investment principles to extract a pension fund's beliefs regarding the risk–return relation of sustainable investments using a rule-based approach.

We consider the period from 2016 to 2021 because this period allows us to investigate the impact of the IRBC initiative that was initiated and signed by most pension funds at the end of 2018. Moreover, we expect that sustainable investing became a greater priority after the Paris Agreement in 2015. Boermans and Galema (Reference Boermans and Galema2019) find that before 2016 most pension funds did not start measuring or externally disclosing the carbon emissions of their investments, whereas they increasingly started to do so as of 2016.

This paper focuses on annual reports for several reasons. First, all pension funds publish an annual report each year, so we have a balanced panel of pension funds. Second, pension funds are legally required to specify in their annual report how they incorporate ESG criteria in their investment policy and ESG risks in their risk management (see Section 2.2). However, since there are no requirements governing how and in how much detail this should be done, there is no guarantee that the relevant statements in the annual report are a complete representation of the SI policy.

Because we focus on disclosures of sustainable investing by pension funds in annual reports, there is concern about potential greenwashing or window-dressing. In the corporate finance literature, there is evidence that companies report mainly positive or general information about sustainable investing and that disclosures therefore suffer from greenwashing (Kim and Lyon, Reference Kim and Lyon2015; Marquis et al., Reference Marquis, Toffel and Zhou2016; Fabrizio and Kim, Reference Fabrizio and Kim2019; Bingler et al., Reference Bingler, Kraus, Leippold and Webersinke2022). Greenwashing or window-dressing incentives could potentially occur in pension funds' annual reports, although the institutional setting of pension funds differs from companies. Dutch pension funds are non-profit organizations but can have other incentives to focus on sustainable investing, for example, because of beneficiaries' preferences for sustainability (Bauer et al., Reference Bauer, Ruof and Smeets2021).

5.2 SI measures

As discussed in the previous section, the SI measures are calculated using annual reports. We construct five measures to measure a pension fund's SI policy along two dimensions. First, we measure the awareness of sustainable investing. We use three measures to quantify awareness: intensity, spectrum, and specificity. Second, we track the implementation of sustainable investing by constructing two additional measures: variety and scope. In this section, we describe the construction of these five measures one by one.

5.2.1 Intensity

Intensity quantifies the attention a pension fund pays to sustainable investing by calculating the proportion of the annual report devoted to sustainable investing. As discussed in Section 4, we create a dataset with all SI-related sentences using a combined rule-based and classification approach. Using this dataset, the intensity measure for each pension fund i in year t is calculated as follows

(1)$${\rm Intensity}_{i, t} = \displaystyle{{\# \;\hbox{SI-related sentences}_{i, t} \over \# \;\hbox{sentences annual report}_{i, t}}}.$$

5.2.2 Spectrum

Spectrum determines how many SI topics are discussed in the annual report in a year. We construct a spectrum of SI topics by applying the BERTopic tool to the dataset of all SI-related sentences. Our dataset of SI-related sentences consists of more than 40,000 sentences. The BERTopic tool generates 27 relevant topics. Table 3 presents these topics. The topics consist of, amongst others, different SI initiatives (e.g., PRI, IRBC), SI strategies (e.g., exclusions, engagement), and excluded firms (e.g., coal mines, weapons manufacturers). Figure 2 shows for a selection of topics five related words in order of their c-TF-IDF score.Footnote ²⁴ This score represents the importance of a word in the sentence. For example, a sentence on energy transition frequently contains the words energy, renewable, and solar. Figure 3 shows for a selection of topics how many pension funds discuss this topic over time. It shows that attention paid to the green bond topic has increased significantly over time: in 2016, only four pension funds discussed this topic, whereas 40 pension funds discussed it in 2021. Moreover, the graph shows that pension funds discussed the IRBC initiative the most in 2018, which makes sense since the IRBC initiative started in 2018. The spectrum measure is equal to the number of topics pension fund i discusses in the annual report of year t

(2)$${\rm Spectrum}_{i, t} = \# {\rm \;SI\ topics}_{i, t}.$$

Table 3.

SI topics

Overview of the SI topics generated with the BERTopic tool applied to the dataset with all SI-related sentences.

Figure 2.

Topic word scores.

Notes: This figure presents the word scores of the five most relevant words for a selection of topics in the BERTopic model.

Figure 3.

Development of topics over time.

Notes: This graph shows the number of pension funds which discuss a certain topic in a certain year based on the BERTopic output.

5.2.3 Specificity

Specificity quantifies the number of statements regarding sustainable investing that contain details of actions specific to the pension fund, detailed performance information, or tangible and verifiable targets set by the pension fund. A pension fund's statement is non-specific if it only contains generalized descriptions that can apply to each pension fund or general and non-verifiable goals without explaining how to achieve them. A pension fund's statement is also non-specific if it contains a description of SI legislation without explaining how the pension fund is implementing it. Our approach is similar to those of Subramanian et al. (Reference Subramanian, Cohn and Baldwin2019), who consider political speeches, and Bingler et al. (Reference Bingler, Kraus, Leippold and Webersinke2022), who analyze climate-related disclosures of companies. We use a classification approach to determine which SI-related paragraphs are specific and which are not. Table 4 presents some examples of labeled paragraphs, and Table 1 shows the performance of the trained language model. The specificity measure for each pension fund i in year t equals

(3)$${\rm Specificity}_{i, t} = \# \; \hbox{specific SI-related paragraphs}_{i, t}.$$

Table 4.

Labeled sentences with regard to specificity

This table presents some examples (translated from Dutch to English) of specific paragraphs (label 1) and non-specific paragraphs (label 0).

5.2.4 Variety

The variety measure quantifies the SI implementation by counting the number of SI strategies implemented by each pension fund. As discussed in Section 2.5, we distinguish the following five SI strategies: divestment, ESG integration, screening, public engagement, and private engagement. We apply a rule-based approach to the dataset with SI-related sentences using a dictionary with keywords and combinations of keywords for each strategy. In this way, we determine which strategies are implemented by pension fund i in year t. The variety measure equals

(4)$${\rm Variety}_{i, t} = \# \;\hbox{implemented SI strategies}_{i, t}.$$

5.2.5 Scope

Finally, scope quantifies the fraction of the asset portfolio that is covered by the pension fund's SI policy. Pension funds invest in various asset classes, but may apply the SI policy only in specific asset classes. We consider the asset classes specified in the OECD guidance for institutional investors (OECD, 2017): equity, corporate bonds, government bonds, real estate, infrastructure, and private equity. We add mortgages as an additional asset class because Dutch pension funds invest a significant fraction of their assets in mortgages.Footnote ²⁵ We apply a rule-based approach to the dataset with SI-related sentences using a dictionary with keywords and combinations of keywords for each asset category. In this way, we determine which asset classes are covered by the SI policy of pension fund i in year t. We combine this information with asset allocation data of pension funds (see Section 5.4). This yields the scope measure

(5)$${\rm Scope}_{i, t} = \displaystyle{{\sum\limits_k {c_{i, k, t}} W_{i, k, t} \over \sum\limits_k {W_{i, k, t}} }},$$

where c _i,k,t takes value one if asset category k is covered by the SI policy of pension fund i in year t and zero otherwise. W _i,k,t is the amount pension fund i invests in asset category k in year t.

We exploit a novel textual analysis approach to construct these five SI measures that quantify the awareness and implementation of sustainable investing. By measuring the sustainable investment policy along two dimensions and considering five SI measures, we robustly measure sustainable investing by pension funds, and we reduce the risk of subjectivity. We track the implementation of sustainable investing by constructing two measures that quantify the implementation of sustainable investing at a meta level: the variety measure counts the number of SI strategies implemented and the scope measure is the fraction of the portfolio included in the SI policy. However, these two measures do not quantify the actual quality of the implementation of sustainable investing. While some measures are concrete and objective (e.g., the intensity measure), others are more abstract and somewhat subjective (e.g., the specificity measure). Sometimes, the term ‘measure’ refers to concrete or objective attributes, and the term ‘metric’ refers to abstract or somewhat subjective attributes. For consistency purposes, we use the term ‘measure’ only.

5.3 Data on the SI initiative

The SI measures, described in the previous section, are used to investigate the impact of signing an SI initiative on the awareness and implementation of sustainable investing. This paper focuses on the IRBC initiative because it is the best-known SI initiative in the Dutch pension sector.Footnote ²⁶ The initiative started with a declaration of intent signed by 40 pension funds in March 2017 (see Figure 4). Subsequently, 73 pension funds signed the initiative at the end of 2018, and several others signed later. The number of pension funds in our sample that signed the initiative in 2018 is 60 instead of 73, because some pension funds left the initiative and some ceased to exist.Footnote ²⁷ In our sample, six pension funds signed the initiative in 2019, two in 2020, and two in 2021.Footnote ²⁸

Figure 4.

Timeline SI initiative.

Notes: This figure shows a timeline with relevant dates for the IRBC initiative and the sample period of this study.

5.4 Pension fund characteristics

In addition to the public annual reports and SI initiative data, the analysis is based on a proprietary dataset from the prudential supervisor of pension funds, De Nederlandsche Bank (DNB), containing information on occupational pension funds in the Netherlands. All pension funds are obliged to report this information to DNB. This dataset has been used before, by for example, Bikker et al. (Reference Bikker, Broeders, Hollanders and Ponds2012), Boermans and Galema (Reference Boermans and Galema2019), and Broeders et al. (Reference Broeders, Jansen and Werker2021). We use a balanced panel of 160 occupational pension funds that reflects almost the entire population of Defined Benefit (DB) pension funds in the Netherlands from 2016 to 2021.Footnote ^29, Footnote ³⁰ Pension funds must report general statistics, such as funding ratio, assets under management, liability duration, and the type of pension fund (corporate, industry-wide, or professional group pension fund). Moreover, pension funds report information on their stakeholders. They report information on the board of trustees, including the gender, age, and tenure of each trustee. Finally, pension funds report information on their actual asset allocation, that is, how much a pension fund invests in each asset category. This information is used to calculate the scope measure.

6.1 Pension fund sample overview

Table 5 shows the statistics of pension fund and board of trustees characteristics for the balanced panel of 160 pension funds. This table shows that the average funding ratio equals 112 percent and the average liability duration is 20.3 years.Footnote ³¹ The average size of total assets under management is about 9 billion euros. The sample contains a small number of very large pension funds, hence the skewness in the distribution. Some 66 percent of the pension funds in our panel are corporate pension funds, 28 percent are industry-wide pension funds, and 6 percent are professional group pension funds. Further, 64 percent of the pension funds invest at least part of their assets actively. Only 18 percent of the pension funds have a positive belief regarding the risk–return relation of sustainable investments, that is, sustainable investing pays off after correcting for risk. The other pension funds either have a more neutral position or do not report their beliefs about the risk–return relation of sustainable investing in their statement of investment principles. The board of trustees' size varies between 3 and 16 trustees, and the average age of an individual trustee is almost 56. The average fraction of female trustees is 21 percent, but there are also pension fund boards with no female trustees. There is a wide dispersion in the average tenure of trustees, which varies between 1 and 20 years.

Table 5.

Statistics on pension fund and board of trustees characteristics

Panel A presents information on pension funds' characteristics and panel B on the boards of trustees for the 160 pension funds in our sample. The mean and standard deviation are measured across pension funds and over time for each variable.

Table 6 shows some statistics on annual reports. We analyzed 938 annual reports from 160 pension funds from 2016 to 2021.Footnote ³² The average report consists of about 2,000 sentences and 840 paragraphs, but there is substantial variation between pension funds.Footnote ³³ There is also a wide dispersion in the number of SI-related sentences. In some annual reports, sustainable investing is not discussed at all, whereas one report contains more than 750 sentences related to sustainable investing.

Table 6.

Statistics on annual reports

This table presents information on pension funds' annual reports. The mean and standard deviation are measured across pension funds and over time for each variable.

For a first inspection of the SI measures data, we plot the distribution of the SI measures in Figure 5. In this figure, each plot visualizes the distribution of a particular measure in a specific year. Two stylized facts stand out. First, the figure shows that for each measure the median, visualized by the dotted vertical black line, increases over time. Second, the value of the scope measure equals zero for a significant number of pension funds. Although this number decreases over time, 40 pension funds still do not report which asset classes are covered by their SI policy in 2021.

Figure 5.

Distribution of SI measures.

Notes: This figure presents the distribution of the different SI measures for all years. Each plot visualizes the distribution of a particular measure in a specific year. The dotted black vertical line in each plot represents the median.

To better understand the implementation of sustainable investing, Figures 6 and 7 provide more information on the data underlying the variety and the scope measure. Figure 6 shows the fraction of pension funds that implemented an SI strategy over time. Divestment is the most popular strategy. This can be explained by the legal requirement introduced in 2013 that pension funds are not allowed to invest in cluster munitions. As a result, most pension funds are forced to implement a divestment strategy for these specific investments. ESG integration shows the biggest relative increase over time, while screening, public engagement, and private engagement have also grown steadily over time.

Figure 6.

Variety measure over time.

Notes: The percentages in this figure represent the fraction of pension funds that implemented a certain SI strategy over time.

Figure 7.

Scope measure over time.

Notes: The percentages in this figure represent the fraction of pension funds that covered a certain asset category with their SI policy over time.

Similarly, Figure 7 shows the fraction of pension funds that covered a specific asset class with their SI policy. All asset classes show a significant increase over time. The most popular asset category covered by the SI policy is equity. This observation can be explained by the fact that the SI policy can cover this asset category in various ways. A pension fund can implement exclusion, screening, and ESG integration based on ESG ratings of listed equity. While the coverage of ESG ratings for listed equity is high, ESG ratings are often not available for alternative asset classes. Moreover, the SI policy can cover equity by influencing the decisions of companies in the equity portfolio (engagement). As the green bond market grew fivefold between 2016 and 2021, it has become easier for pension funds to include fixed income (government bonds and corporate bonds) in their SI strategy.Footnote ³⁴

6.2 Empirical model

In this section, we present empirical models for the three hypotheses discussed in Section 3. To test the first hypothesis, we use the following pooled OLS model

(6)$$ESG_{i, t} = \alpha + {\beta }^{\prime}\cdot {\bf X}_{i, t} + \theta _t + {\rm \epsilon }_{i, t}, \;$$

where ESG _i,t is one of the five SI measures of pension fund i in year t and ${\bf X}_{i, t}$ contains several explanatory variables. θ _t is a set of year dummies to control for year-specific conditions and ϵ _i,t is the error term. ${\bf X}_{i, t}$ contains both pension fund characteristics and board of trustees characteristics that might impact the SI measures. The pension fund characteristics include the size of the pension fund, represented by the natural logarithm of the total assets under management, the funding ratio, and the liability duration, which is the average time to maturity of the pension liabilities. Further, two dummies for professional group pension funds and corporate pension funds represent the type of pension fund. Industry-wide pension funds are the omitted category. Finally, we include a dummy for active investing. This dummy variable equals one if the pension fund invests at least part of its assets actively and zero otherwise. The board characteristics include the average age of the board of trustees, the fraction of female trustees, and the average tenure of the board of trustees. Moreover, we include a dummy variable that represents the belief regarding the risk–return relation of sustainable investing. This dummy variable equals one if the pension fund expects a positive impact of sustainable investing on the risk–return relation and zero otherwise.

For the second hypothesis, we use a probit model to analyze the effect of pension fund and board characteristics on the probability of signing an SI initiative

(7)$$P[ SIGN_i = 1] = \Phi [ \alpha + \beta \cdot {\bf X}_i + {\rm \epsilon }_i] , \;$$

where SIGN _i takes value zero for non-signatories and value one for signatories and ${\bf X}_i$ contains several explanatory variables (pension fund and board of trustees characteristics) that explain whether or not a pension fund signs the IRBC initiative. ϵ _i is the error term. We do not use the whole panel dataset for this hypothesis, because we use a cross-sectional probit model. For pension funds that signed the IRBC initiative in 2018 or did not sign the IRBC initiative, we use the explanatory variables of 2018. For pension funds that signed the IRBC initiative in a later year, we use the explanatory variables of the year of signing.

To test the third hypothesis concerning the impact of signing the IRBC initiative on the development of SI measures over time, we use a difference-in-differences (diff-in-diff) specification to estimate the differential effect of signing the IRBC (treatment) on the SI policy measures. We apply the diff-in-diff specification with staggered treatments on the panel of IRBC signatories and non-signatories to evaluate the between-group differences of the change in SI measures over time.Footnote ³⁵ The diff-in-diff model is specified as follows:

(8)$$ESG_{i, t} = \alpha + \gamma \cdot SIGN_i + \delta \cdot IRBC_{i, t} + {\beta }^{\prime}\cdot {\bf X}_{i, t} + \theta _t + {\rm \epsilon }_{i, t}, \;$$

where ESG _i,t is one of the five SI measures of pension fund i in year t and SIGN _i takes value zero for non-signatories and value one for signatories. IRBC _i,t takes value zero for non-signatories and signatories before signing and value one for signatories after signing, and ${\bf X}_{i, t}$ contains several explanatory variables. Finally, θ _t is a set of year dummies to control for year-specific conditions and ϵ _i,t is the error term of pension fund i in year t. We are interested in the coefficient δ, which measures the effect of signing the IRBC initiative on the SI measure. A positive coefficient δ indicates that, on average, the difference between the SI measure of IRBC signatories and non-signatories has increased after the signatory year.

To rule out spurious correlation, we control for the following endogeneity concerns: selection bias, omitted variable bias, reverse causality, and measurement error. Below we discuss these endogeneity concerns one by one.

First, selection bias arises in our sample because signing the IRBC is voluntary. Pension funds that have already been enhancing their SI policy in the past or are planning to do so are more likely to sign the IRBC initiative. Therefore, pension funds that signed the IRBC may not be representative and could differ systematically in their main characteristics compared to pension funds that did not sign the IRBC. A simple comparison of IRBC signatories and non-signatories is thus not feasible. Since we cannot analyze pension funds in two conditions (signatory and non-signatory) simultaneously, we use a matching methodology. Matching aims to equate the distribution of covariates in the treated (signatories) and control (non-signatories) groups (Stuart, Reference Stuart2010). While several matching methods exist, one of the most common methods is r:1 nearest neighbor matching (Rubin, Reference Rubin1973). Nearest neighbor matching matches control units to the treated group. For each treated unit i nearest neighbor matching selects the r control units with the smallest distance from i. We conduct a 3:1 nearest neighbor matching with probit regression-based propensity scores using the pension fund and board characteristics as matching variables.Footnote ³⁶ The method matches pension funds in the control group (non-signatories) to the treated group (signatories) with the smallest distance, discarding non-matched pension funds. We use nearest neighbor matching with replacement allowing the same control fund to be matched multiple times. The propensity score is used as the similarity measure between pension funds and is defined as the probability of signing the IRBC initiative given the observed pension fund and board characteristics. Subsequently, we weight the regression in equation (8) with these propensity scores.

Second, we address omitted variable bias in two ways. We include explanatory variables (${\bf X}_{i, t}$ in equation (8)) in the diff-in-diff regression to control for the effect these variables have on the SI measures. Potentially, some variables not included in the model impact the pension fund's SI policy and correlate with the explanatory variables in ${\bf X}_{i, t}$. As a result, the estimates of the model are potentially biased. Therefore, we include fixed effects as additional explanatory variables in the model to control for omitted variable bias based on the assumption that the omitted variables are constant over time. We add pension fund fixed effects κ _i to the model in equation (8) and analyze this model as a robustness check

(9)$$ESG_{i, t} = \alpha + \delta \cdot IRBC_{i, t} + {\beta }^{\prime}\cdot {\bf X}_{i, t} + \theta _t + \kappa _i + {\rm \epsilon }_{i, t}.$$

In this model, ${\bf X}_{i, t}$ contains fewer explanatory variables compared to equation (8) because the time-invariant variables are excluded from ${\bf X}_{i, t}$. The pension fund fixed effects capture the effect of these variables.

Third, we use the diff-in-diff specification to overcome reverse causality concerns. An essential requirement for a diff-in-diff specification concerns the parallel trend assumption. This assumption requires that the difference between the treatment group (signatories) and control group (non-signatories) is constant before the treatment. Although there is no statistical test for this assumption, visual inspection is useful. Figure 8 presents for each SI measure the mean value for both the treated group (signatories) and control group (non-signatories). This figure shows that signatories had higher values for all SI measures compared to non-signatories. Generally, the trends of signatories and non-signatories before the IRBC initiative are similar for each SI measure in line with the parallel trend assumption. Note that the figure shows the mean values for all signatories and non-signatories. Because we apply matching to reduce selection bias, not all signatories and non-signatories are included in the diff-in-diff model.

Figure 8.

SI measures of signatories and non-signatories over time.

Notes: This figure presents for each SI measure the mean value for both the treated group (signatories) and control group (non-signatories).

Fourth, we reduce measurement error concerns by measuring awareness and implementation of sustainable investing with different SI measures. Moreover, we will construct indices that combine multiple individual SI measures and analyze the impact of signing on these indices as an extra robustness check.