The power of the people: labor unions and corporate social responsibility

Abstract

1. Introduction

A growing number of politicians have suggested that society would benefit if corporations became more socially minded. Bills such as the Reward Work Act (Baldwin 2019) and the Accountable Capitalism Act (Warren 2018, 2019) advocate that one way to encourage corporations to make decisions that benefit society at large, as opposed to just shareholders, is to give employees greater control of corporate decision-making. In this article, we examine the impact of employee unionization on environmental and social (E&S) scores.

Labor unions have two primary tools at their disposal to directly influence management’s decisions: collective bargaining and strikes. Employers usually cannot change working conditions independently because, under the National Labor Relations Act (NLRA), they have a duty to try to work out an agreement in good faith with the union within a reasonable amount of time.¹ For example, in 2018, 8,300 unionized Marriott housekeepers across nine cities went on strike after the hotel chain failed to make adequate progress renegotiating contracts that had expired over the previous several months. With the slogan “One job should be enough,” the unionized hotel workers demanded that Marriott provide greater job security to union members by discontinuing its Green Choice program. This program incentivized guests to opt out of daily housekeeping to help the environment, thereby dramatically cutting staff hours and ultimately reducing wages. The union also negotiated for Marriott to implement employee training programs if automation replaced jobs and demanded improved healthcare and parental leave policies. After two months of strikes, the local unions successfully bargained for improved training, healthcare, and parental leave policies. Marriott later discontinued the Green Choice program during the coronavirus disease 2019 (COVID-19) pandemic for reasons unrelated to union negotiations.

If unions cause firms to pursue socially responsible strategies that benefit society at large (i.e., external stakeholders), we would expect a positive relationship between unionization and E&S scores.² Alternatively, unionization may hinder firms’ E&S efforts for two reasons. First, there may be direct conflicts between employees and external stakeholders, as in the Green Choice program. In this section, there may be indirect conflicts as investing in employees reduces the resources available to invest in external stakeholders. Indeed, several studies have shown that unions decrease shareholder value (Lee and Alexandre 2012; Knepper 2020). They may also decrease external stakeholder value.

In this article, we construct ten yearly, firm-level E&S scores using the Thomson Reuters ASSET4 environmental, social, and governance (ESG) database from 2002 to 2021, which contains seventy environmental and seventy-eight social indicators. These indicators include four internal social scores that predominately benefit employees (Employment Quality, Training and Development, Diversity and Opportunity, and Health and Safety) and six external scores that predominately benefit external stakeholders (Emissions Reductions, Resource Reductions, Product Innovation, Community Involvement, Human Rights, and Product Responsibility). Higher scores always indicate greater E&S policies and outcomes. We compute the year-over-year percentage changes of these scores to measure the magnitude of yearly increases or decreases. Edmans (2013) advocates that ESG should be disaggregated precisely because there may be differences between employee and external social scores. Therefore, we focus our analysis on the disaggregated scores.

In our initial analysis, our measure of union strength is the establishment’s unionization rate, and we find that a higher unionization rate is associated with increased internal social and decreased external E&S score changes. We saturate the model with firm fixed effects to account for time-invariant, firm-level characteristics; industry–year fixed effects to adjust for the industry-level time trend of all variables, including industry-level differences in unionization rates; and a battery of time-varying firm-level controls. The findings from this initial analysis are consistent with Faleye Vikas, and Randall (2005), who show that when employees have more corporate decision-making power, they maximize their own value, rather than shareholder value.

A potential concern within our initial framework is that firm-level characteristics could drive changes in E&S scores and unionization rates simultaneously. To address endogeneity concerns within our initial setting, we bolster our identification by implementing a regression discontinuity design (RDD) framework using close union elections.

Within our RDD analysis, we limit our sample to firms that narrowly win (or lose) union elections by less than 20 percent. To establish the validity of our setting, we show that pre-election, unionized and non-unionized firms are similar in latent factors that may be correlated with unionization. We then show that there is no voter manipulation at the 50 percent voter threshold necessary for victory. By assuming a meaningful component of randomness in the outcome of these realized close elections, we can isolate the impact of firm unionization on our E&S variables of interest. Within the RDD framework, we confirm our baseline results and find that unionization leads to increased internal and decreased external E&S score changes. Our most conservative (linear) RDD estimates suggest that union election victories are associated with a 16.7 percent increase in diversity score changes and a 5.7 percent reduction in emission score changes. These estimates grow in magnitude when we examine alternative internal or external E&S score changes or when we apply quadratic polynomial estimates.

We also show that the level of internal (external) E&S scores of the firms with winning union elections significantly increases (decreases) 1 year after the elections compared to the level one year prior to the elections. In contrast, the post-election E&S scores of the firms with losing union elections do not change significantly.

One potential reason for the reduction in external E&S scores is that firms may divert resources from external E&S investment toward other sectors, such as precautionary cash or firm investment. This is most likely to occur when firms face resource constraints since unconstrained firms should optimize over factors affecting employees and factors affecting stakeholders separately. We construct a measure widely used in the literature to proxy for firm financial constraints from Kaplan and Zingales (1997) and interact it with the main independent variable in the RDD regression. Our estimates show that when firms are more resource-constrained, this can amplify unions’ positive (negative) impact on some elements of firms’ internal (external) E&S metrics.

Furthermore, if unionization leads to changes in E&S scores, one might expect these effects to be more pronounced when the union has greater bargaining power. We empirically explore this conjecture and interact two different proxies for union strength with our main independent variables. We show that when unions have greater bargaining power, such as when firms are located in states without right-to-work laws or when local employment rates are high (low unemployment), our baseline results are amplified.

In supplemental analysis, we augment our E&S scores by analyzing the real effects of two additional outcome variables: worker injury rates and toxic gas emissions. If corporations are implementing internal safety policies, this would most directly benefit union members (i.e., by reducing worker injuries). While both union members and non-employees benefit from clean air, toxic gas emissions are closely related to external E&S categories. We find that unionization is associated with a decline in worker injuries and with an increase in certain types of toxic gas emissions. This finding complements our earlier results, which use E&S scores to show that unionization leads to increases in outcomes that benefit workers but decreases in outcomes that predominately benefit external stakeholders.

Our article contributes to both the unionization and corporate social responsibility (CSR) literature. The vast unionization literature has focused broadly on the effects of unionization on shareholders and union members. The theoretical literature tends to model negotiations between unions and firms using a two-party game in which the union’s objective is to maximize member utility and the manager attempts to maximize shareholder value (McDonald and Solow 1981; Clark 1990). Effects on external stakeholders, such as community members, are overlooked within these models. Empirically, several papers have shown that unions meaningfully impact firm decision-making. These papers establish that while unions provide a number of benefits to employees, they can also impose costs on firms.³

We also contribute to the CSR literature by exploring unionization as a potential determinant of CSR. The bulk of the existing CSR literature debates whether CSR activities are indicative of agency problems within a firm or whether they can create shareholder value. Some studies argue that investment in CSR can be a manifestation of agency problems since CSR may not be in the best interest of shareholders (Tirole 2001; Di Giuli and Leonard 2014; Masulis and Syed Walid 2015). A growing number of studies have found that CSR is a unique resource that can generate benefits or reduce costs (Dowell, Stuart, and Bernard 2000; Edmans, 2011; Flammer 2013; Dimson, Oğuzhan, and Xi 2015; Flammer 2015; Flammer and Jiao 2017; Flammer, Bryan, and Dylan 2019; Flammer and Aleksandra 2019; Edmans 2020). Although unions are not necessarily shareholders, they are still powerful internal stakeholders with the ability to influence corporate policies. Our framework allows us to infer causality and explore unionization as one of the determinants of CSR. In a paper publicly released at the same time, Ertugrul and Marciukaityte (2021) examine the impact of labor unions on aggregate CSR scores using data from the MSCI rating agency. They conclude that unionization rates lead to a decrease in aggregate CSR as well as the disaggregated categories that benefit communities, customers, and employees. In contrast, we implement a close election setting and more granular data, and we show that unionization leads to increases in E&S categories that predominately benefit employees. This suggests that unionization is an important way for employees to gain non-pecuniary benefits. To add credibility to our findings, we further our analysis by showing consistent findings when real outcomes not driven by rating agency methodology are examined.

Our article has significant policy implications and highlights the importance of recognizing the heterogeneity among stakeholders. Similar to shareholders who seek to maximize firm value, each stakeholder group also aims to maximize its own interests. Our findings indicate that empowering a specific stakeholder group, such as employees through unionization, does not necessarily lead to benefits for all stakeholders. Therefore, when policymakers advocate for strengthening the power of one stakeholder group, they must carefully consider the implications for all diverse stakeholders.

2. Data

2.1 Corporate social responsibility data

Our measures of CSR come from the commercially available Thomson Reuters ASSET4 ESG database. Starting from 2002, Thomson Reuters has annually compiled information from annual reports, nongovernmental organizations, and news sources for large publicly traded companies to determine seventy environmental and seventy-eight social indicators of corporate policies. We focus our analysis on the E&S indicators that span 2002–2021.

Thomson Reuters constructs these indicators by compiling information pertaining to standardized questions such as, “Does the company show an initiative to reduce, reuse, recycle, substitute, phase, or compensate CO₂ equivalents in the production process?.” Questions can also have numerical answers, such as the total amount of waste produced in tons divided by net sales or revenue in US dollars. The seventy environmental indicators span three categories: Emissions Reduction (twenty-eight indicators), Product Innovation (twenty-five indicators), and Resource Reduction (seventeen indicators). The seventy-eight social indicators are broken down into seven categories: Diversity and Opportunity (ten indicators), Employment Quality (seventeen indicators), Health and Safety (nine indicators), Training and Development (ten indicators), Product Responsibility (ten indicators), Community (fourteen indicators), and Human Rights (eight indicators).  Appendix B contains further details on all indicators within each category.

We apply an equal weighting scheme to arrive at measures of a company’s E&S policies and outcomes by quantifying each of the provided indicators. Following Dyck et al. (2019), we translate each question into an indicator variable. Questions answered in a yes-or-no fashion receive a value of one if the firm’s answer has a positive effect on society or the environment and zero otherwise. For questions with numeric answers, such as the ratio of waste produced to net sales, we assign the firm a value of one (zero) if its policies are more positive (negative) compared to other firms in the database. For example, the indicator variable related to CO₂ emissions takes one when its value is below the sample median and zero when above the sample median; the indicator variable for environmental regulation violation takes one when the number of violations is zero and zero when it is positive.

After classifying all categories from the database, we employ an equal-weighting scheme to determine E&S scores. For each of the three environmental area categories and the seven social area categories, we compute an equal-weighted average of indicator variables associated with the area’s corresponding items. The values for these averages are between zero and one, like the indicator scores. To investigate changes in E&S scores, we compute the percentage changes for each category’s average. Table 2 shows the summary statistics for the percentage changes associated with each of the three environmental categories (E_Emission, E_Product, and E_Resource) and the seven social categories (S_Diversity, S_Employment, S_Health, S_Training, S_Product, S_Community, and S_Human).

2.2 Union data

To measure firm unionization, we rely on two primary data sources. Unionization rate data are from the Federal Mediation and Conciliation Service (FMCS) database, and union election data are from the National Labor Relations Board (NLRB).

Following previous literature, we proxy for union strength and bargaining power by examining firm–year unionization rates (Connolly Robert, Barry and Mark 1986; Bronars and Deere 1991; Matsa 2010; Chyz et al. 2013). Firms with labor union contracts are required to file notices of contract expiration with the agency under the NLRA. The FMCS database spans 2002–2021 and provides the start and expiration dates of the contract, the name of the employer, the size of the bargaining unit (i.e., the number of employees in the union), and the size of the business (i.e., the total number of employees). We use the employer’s name to match the FMCS data to Compustat. Then, following the literature, we calculate the firm–year unionization rate as the total size of the bargaining unit (number of union employees) as a fraction of the total number of firm employees for the 865 unionized firms in our sample. These firms represent 2,120 firm–year observations with E&S percentage change measures.

From there, we use a type of nearest neighbor matching algorithm to match each unionized firm with a nonunionized firm. For each unionized firm $f$ with $k$ nonmissing control variables between the starting and ending years (years $ts$ and $te$⁠) in our treatment sample, we compute the average of each control variable between years $ts$ and $te$ and construct a k-element vector $vu=u1,u2,…,uk$⁠. The control variables include the long-term debt scaled by lagged assets (LEVERAGE), the natural logarithm of total assets (SIZE), the net operating loss carryforward scaled by lagged assets (NOLCF), the change in net operating loss carryforward scaled by lagged assets (CHG_NOLCF), an indicator of whether the firm reports a loss in any of the last three fiscal years (LOSS), sales growth scaled by lagged assets (SGA), capital expenditures scaled by lagged assets (CAPEX), and Earnings Before Interest, Taxes, Depreciation, and Amortization (EBITDA) scaled by lagged total assets (EBITDA). All variables are defined in Table 1.

From the full sample of Compustat firms, we select a subsample of nonunionized firms in the same industry, defined as the two-digit SIC code, of firm $f$⁠, compute each firm $i$’s averages for the $k$ control variables between years $ts$ and $te$⁠, and construct a k-element vector $vni=ni1,ni2,…,nik$⁠. For each firm $i$⁠, we calculate the standardized vector $svni=ni1u1,ni2u2,…,nikuk$ and compute the Euclidean distance between the identity vector (or all-ones vector) and $svni$⁠. The Euclidean distance between two vectors can be calculated by taking the difference between each corresponding element of the vectors, squaring them, summing up all the squared differences, and finally taking the square root of the sum. The distance between vectors X and Y can be mathematically expressed as $X-Y=∑i=1nxi-yi2$⁠. We then match firm $f$ with the nonunionized firm $j$ that has the smallest Euclidean distance and has not been matched with another unionized firm. Each year’s observation of firm $f$ is matched with the corresponding year’s observation of firm $j$⁠.

Using this matching algorithm, we match every treated firm’s firm–year observation with a nonunionized firm’s corresponding firm–year observation. Each treated firm is only matched with one nonunionized firm. Over the subsequent year, we examine the unionized and non-unionized firms for differences after the treated firm becomes unionized. In total, our sample has 4,240 firm–year observations.

To bolster our identification, we conduct an RDD analysis on close union elections that occur between 2002 and 2021. Our dataset originates in 2002, coinciding with the availability of the ASSET4 ESG data. Within this dataset, each record includes information such as the employer’s name, election date, election result, and petition type. There are three primary types of petitions: Proof of Representative, Revocation of Representatives, and Employer Application. The Revocation of Representatives and Employer Applications are used to decertify unions, while a Proof of Representative petition is filed when a union seeks to represent employees. Following the submission of a petition to the NLRB, a union can proceed with an election. For our union election analysis, we exclusively consider Proof of Representative petitions. Election data predating 2010 are sourced from Knepper (2020),⁴ and from 2010 onwards, we obtain our data directly from the NLRB website.⁵

After consolidating these two election datasets, we meticulously conduct data collection, processing, and manual matching of unionization records with the National Establishment Time-Series (NETS) database. Leveraging the corporate affiliation information for each establishment provided by NETS, we subsequently establish manual links between our union election sample and Compustat. Figure 1 shows the number of union elections per year as well as the average passage rate of unionization elections each year for firms with available ESG data. The sample used in the top figure has 663 elections obtained after step six in Panel B of Appendix Table A1. As the bottom figure shows, the average union passage rate is primarily between 40 percent and 60 percent each year.

Figure 1.

Union elections and passage rates by year. This figure plots the number of union elections by year (top) and the average passage rates by year (bottom). Union election results are from the NLRB and span 2002–2021. We show the elections where our E&S measures are available and there are at least ten voters.

Open in new tabDownload slide

2.3 Sample construction and summary statistics

We conduct our analysis using two separate datasets. The first sample, focused on unionization rates as a proxy for union bargaining power, combines data from the Thomson Reuters ASSET4 database, FMCS, and Compustat. After manually matching the FMCS data to Compustat, we have 15,650 firm–year observations from 1,893 firms from 2002 to 2021. Firms with missing controls (described in the following paragraphs) are dropped, leaving us with 5,934 records. After merging in the Thomson Reuters ASSET4 database, we have 2,120 firm–year observations. We augment our unionized sample with 2,120 matched non-unionized firm–year observations that contain ESG measures of equally weighted scores. Our merged dataset contains 4,240 firm–year observations.

Our second dataset combines the union election data from NLRB with the Thomson Reuters ASSET4 database and Compustat. We merge the NLRB data to Compustat to obtain a sample of 4,123 observations from 2002 to 2021 that spans 2,289 firm–years. Following Bradley, Incheol, and Xuan (2017), we keep only the first election if multiple elections occur in one firm within a given year. From there, we limit our sample to close elections where unions won or lost by less than 20 percent of votes. While this dramatically decreases our sample size, it is necessary for the RDD analysis. After matching the close election NLRB data to the Thomson Reuters ASSET4 database, we have 312 firm–year observations from 189 firms. Unions win 34.29 percent of the elections in our sample, which is comparable to the 36 percent victory rate reported in Bradley, Incheol, and Xuan (2017).

From Compustat, we include the control variables defined earlier, and all control variables are winsorized at the 1st and 99th percentile to curtail the influence of outliers.⁶Table 1 provides a detailed description of each variable. Appendix Table A1 presents the step-by-step process of the data matching and sample construction.

Table 2 presents the summary statistics for each variable of interest for each of our two synthesized datasets. Columns 1–6 show the observations, average, standard deviation, median, minimum, and maximum for the sample focusing on the unionization rate, while Columns 7–12 present summary statistics for the union election sample. In Panel A, we present the summary statistics for all dependent variables of interest quantifying a company’s E&S components. In Panel B, we present the summary statistics for our unionization rate (UNIONIZATION), which has a mean of 33.92 percent. For our sample of unionized firms, the mean is 67.8 percent. Chyz et al. (2013), who also use the FMCS data, have an average unionization rate of 73.35 percent, indicating that our sample is comparable. In Panels C and D, we show summary statistics for the variables that we use as interactions as well as for the control variables. To curtail the influence of outliers, all continuous variables are winsorized at the 1 percent and 99 percent levels.

3. Empirical strategy and results

We conduct two separate sets of analyses that provide consistent conclusions. We first present the main results using an ordinary least squares (OLS) approach with unionization rates as our independent variable of interest. This setting is advantageous because it includes more observations but suffers from the disadvantage of endogeneity. Subsequently, we also use a close election RDD setting, which allows us to compare the outcomes of firms that narrowly unionized to those that remained non-unionized. If the setting meets the criteria discussed in Section 3.2, it enables us to identify the causal impact of unionization on E&S score changes. By implementing these two settings and empirical strategies, we show that our results are consistent across both the larger sample provided by our examination of unionization rates and the cleaner setting provided by the close election setting.

3.1 Unionization rate and E&S score changes

For a given firm f and year t, the dependent variable $Depft+1$ measures the E&S score change over the next year. The variable $UNIONIZATIONft$ is the total size of the union bargaining unit (number of unionized employees) as a fraction of the total number of firm employees. All regressions include firm fixed effects, $Φf$⁠, industry × year fixed effects, $Φit$⁠, and the eight firm–year financial controls described in Table 1 and designated by $χft$⁠. We define an industry using the two-digit SIC code. Firm fixed effects account for time-invariant, firm-level characteristics and effectively allow us to compare the firm to itself. Industry × year fixed effects adjust for the industry-level time trend of all variables, including industry-level differences in unionization rates. We cluster standard errors at the firm level, and we present the results in Table 3.

As shown in Table 3, when firms have greater levels of unionized employees, there are statistically significant increases in four internal social score changes and decreases in three external environmental and three external social score changes. For example, a one standard deviation increase in the unionization rate leads to an increase in S_Health of 4.417 percent (=0.3937*0.1122) and a decrease in S_Human of 4.291 percent (=0.3937*0.1090) for the average sample firm. As shown later, within the RDD framework, these score changes grow in magnitude.

A primary concern among unions is improving internal working conditions for their members, which can be monetary or nonmonetary. Therefore, a union may value investment in E&S categories that more directly benefit its members more highly. This is consistent with the heterogeneous relationships between unionization rate and different types of E&S score changes. As shown in Table 3, the relationship is positive for the four internal social scores that predominately benefit employees, including Employment Quality, Training and Development, Diversity and Opportunity, and Health and Safety. The relationship is negative for the three environmental and three external social score categories that predominately benefit external stakeholders, including Emissions Reductions, Resource Reductions, Product Innovation, Community Involvement, Human Rights, and Product Responsibility.

The four internal E&S components all quantify the extent to which corporations have policies that directly benefit their employees. For example, the Health and Safety component quantifies the degree to which working conditions are safe for employees, such as the frequency and severity of workplace injuries as well as the existence of health and wellness plans. The Diversity and Opportunity category quantifies workplace employee diversity and opportunities, such as policies that value families, females, minorities, and work–life balance. The Training and Development category consists of various employee career development and training opportunities, such as policies favoring promotions from within, cooperation with schools, and programs to support employees in learning new skills. Employment quality consists of metrics that quantify whether firms have policies implementing fair employee compensation and working conditions. While there may be positive community spillovers from implementing these policies, we believe that employees most directly benefit from these policies. As such, we may expect unions, advocating for their members employed at the firm, to advocate for investment in such categories.

The remaining six external E&S scores provide benefits to a broader group of stakeholders who are not necessarily employees. For example, firm reductions in emissions and the use of natural resources benefit society at large. Similarly, actively donating to and participating in community events, implementing human rights policies, and creating high-quality products benefit all stakeholders, regardless of whether they work for the firm. While union members still benefit from these external categories, we believe that the benefits derived from internal corporate policies provide more direct benefits.

For all internal E&S score changes (Columns 1–4), higher levels of unionization are associated with higher levels of E&S categories that benefit members. In contrast, the effect is negative for all external E&S measures (Columns 5–10). These results suggest that unions maximize member value and that increasing internal E&S is an additional channel through which unions derive value for their members. However, this increase in firm internal E&S may come at the expense of stakeholders external to the firm and union.

The NLRA mandates that employers must negotiate with unions before changing any working conditions within a reasonable amount of time, even working conditions that may ultimately benefit employees.⁷ Thus, we may expect management and unions to quickly initiate a round of negotiations following unionization, focusing on issues of primary importance, which is why we focus our analysis on year t + 1. However, there may be subsequent rounds of negotiations between unions and managers over time. Thus, it is possible that the effects of unionizations may manifest over longer horizons. As shown in Supplementary Appendix Table OA2, the positive (negative) relationship between unionization and internal (external) E&S score changes persist in years t + 2 and t + 3 for all E&S metrics reported in Table 3.

3.2 Union election and E&S score changes

One concern with our empirical strategy presented in Section 3.1 is endogeneity. While we have included time-varying controls and fixed effects, there may be unobservable time-varying factors correlated with unionization and E&S score changes. To address this identification concern, we substantiate our findings by using a cleaner RDD framework within a close union election setting. By employing this framework and setting, we can establish causality and compare E&S score changes of firms that narrowly passed unionization to those that narrowly failed to do so.

While there is no single path for an establishment to become unionized, as discussed in DiNardo and Lee (2004), a common route is through a secret ballot NLRB election. Once a group of workers decides to form a union, they contact a labor union and ask for assistance in organizing a “card drive.” In this drive, the union collects “authorization cards” from workers petitioning the NLRB to hold an election within a given time frame, typically six months. If enough cards are collected (a minimum of 30 percent) and the NLRB rules that the workers seeking union representation have a “community of interest,” the NLRB will facilitate a secret ballot election. If a simple majority of voters casts their votes in favor of the union, the NLRB certifies the union as the sole authorized representative of the workers in the bargaining unit.

As a requirement for RDD analysis, we first limit our sample to close election settings where the union won or lost by a narrow margin; this restriction yields a very small sample size. In the following subsection, we present compelling evidence supporting the efficacy of the close election framework employed in previous literature (DiNardo and Lee 2004; Lee and Alexandre 2012; Bradley, Incheol, and Xuan 2017; Huang et al. 2017; Qiu and Shen 2017; Campello et al. 2018; Knepper 2020; Kini et al. 2022).

3.2.1 RDD framework validation

To implement an RDD framework, we need to show “local” exogenous variation in unionization that is generated by union elections that pass or fail by a small margin of votes around the 50 percent threshold. To the extent that there is some randomness in the outcome of elections, we can establish a causal relationship between firms that barely became unionized and those that did not. An additional advantage of this RDD setting is that we do not have to include observable firm covariates in our analysis to obtain identification (Lee and Thomas 2010).

Another requirement for implementing the RDD framework is showing that both employers and voters cannot perfectly manipulate the outcome variable near the known cutoff (Lee and Thomas 2010). If we can satisfy this assumption, the variation in firms that unionize is as good as in those from a randomized experiment. Figure 1 shows the passage rate of unionization elections each year for firms with available E&S scores; the passage rate is primarily between 40 percent and 60 percent in our sample period.

Following Bradley, Incheol, and Xuan (2017), we further test this assumption in Figure 2, which plots the distribution of pro-union vote shares, shown on the x-axis, into forty equally spaced vote share bins. Figure 2 indicates that the distribution of vote shares is relatively smooth and, importantly, that there is no sharp discontinuity around the 50 percent vote threshold. This distribution suggests that voters and employers are not able to self-sort and manipulate voting outcomes.

Figure 2.

Distribution of votes. This figure plots the histogram of the distribution of the number of elections with the percentage of votes in favor of unionizing in our sample across forty equally spaced bins with a width of 2.5 percent. Union election results are from the NLRB from 2002 to 2021. We show the elections where our E&S measures are available and there are at least ten voters.

Open in new tabDownload slide

Additionally, we follow the two-step procedure outlined in McCrary (2008) and provide a formal test for discontinuity of the density in Figure 3. The x-axis plots the percentage of pro-unionization votes, and the y-axis shows the density estimates. The dots depict the density, and the lines represent the fitted density function of the number of votes with a 95 percent confidence interval. The density plot is generally smooth and does not show evidence of a discontinuity around the 50 percent cutoff, providing corroborating evidence that there is no precise manipulation by voters at the indicated threshold. Furthermore, in the Frandsen Brigham (2017) test for valid RDD design for discrete running variables, the P-value is zero, suggesting that there is a random assignment of firms to the close-win and close-loss groups and reveals no evidence of voter manipulation around the majority threshold. These tests indicate that ex-ante, the outcomes of these elections are not predictable based on E&S scores.

Figure 3.

Density of union vote shares. This figure plots the density of union vote shares following the procedure in McCrary (2008). The x-axis is the percentage of votes favoring unionization. The solid line represents the fitted density function of the forcing variable (number of votes) with a 95 percent confidence interval around the fitted line. Union election results are from the NLRB from 2002 to 2021. We show the elections where our E&S measures are available and there are at least ten voters.

Open in new tabDownload slide

Another key assumption of the RDD is that there should not be a discontinuity in other covariates correlated with firm E&S outcomes at the cut-off point. Table 4 compares the covariates of firms that unions won or lost by a margin of less than 20 percent in year t, the year of the union elections.⁸ In Section 3.2.3, we use the methodology of Calonico, Matias, and Titiunik (2014) to justify the bandwidth choice. We examine the same eight observable covariates that have been used as previous controls and show that all covariates are similar between firms that barely unionize and those that do not unionize. More importantly, we do not observe that the ten E&S score changes are significantly different across these two groups of firms in the union election year.

3.2.2 RDD results

Our RDD results confirm that unionized firms exhibit lower E&S score changes. First, we visually check our relations around the cutoff in Figure 4. Similar to Figures 2 and 3, our x-axis represents the share of pro-union votes, and the y-axis is our dependent variable of interest, which is each of the ten E&S score changes. We distribute vote shares into equally spaced bins, each with a width of 2.5 percent. The zero percent on the horizontal axis indicates a zero-winning margin, which is equivalent to the threshold of 50 percent pro-union votes. The dots depict the average score examined, and the solid line represents the fitted quadratic polynomial estimate with a 95 percent confidence interval around the fitted value. There is a sharp discontinuity around the 50 percent threshold: once pro-union votes cross the 50 percent threshold, the internal social categories rise substantially, and the external E&S categories drop substantially. For eight of the ten figures, including all six figures for the external sub-scores, the confidence intervals between the left and right side of the 50 percent threshold do not overlap. However, the confidence intervals slightly overlap in two of the ten figures, S_Community and E_Emission. Taken together, the graphs in Figure 4 suggest that unionization leads to higher internal and lower external E&S score changes.

Figure 4.

Continued

Open in new tabDownload slide

Figure 4.

Unionization regression discontinuity plots. This figure plots the fitted quadratic polynomial estimate with a 95 percent confidence interval around the fitted value. The x-axis is the percentage of votes favoring unionization; the next year’s percentage change of equal-weighted E&S indicators is on the y-axis, as S_Employment (Panel A), S_Health (Panel B), S_Training (Panel C), S_Diversity (Panel D), S_Product (Panel E), S_Human (Panel F), S_Community (Panel G), E_Resource (Panel H), E_Product (Panel I), or E_Emission (Panel J). The dots depict the average score in each of sixteen equally spaced bins with a width of 2.5 percent. Union election results are from the NLRB from 2002 to 2021, and E&S scores are from the Thomson Reuters ASSET4 ESG database.

Open in new tabDownload slide