Business Finance – Prepare a 4 page analysis Briefly describe the company Oiselle

Vol.:(0123456789)1 3

https://doi.org/10.1007/s10869-022-09819-x

Corporate Diversity Statements and Employees’ Online DEI Ratings: An Unsupervised Machine‑Learning Text‑Mining Analysis

Wei Wang1  · Julie V. Dinh1,2 · Kisha S. Jones3 · Siddharth Upadhyay3 · Jun Yang4

Accepted: 16 May 2022 © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract Following the deaths of many Black Americans in spring 2020, public consciousness rose around the societal mega-threat of racism. In response, many organizations released public statements to condemn racism and affirm their stance on diversity, equity, and inclusion (DEI). However, little is known about the specific thematic contents covered in such diversity state- ments and their implications on important organizational outcomes. Taking both inductive and deductive approaches, we conducted two studies to advance our understanding in this area. Study 1 employed structural topic modeling (STM)—an advanced unsupervised machine-learning text-mining technique—and comprehensively analyzed the latent semantic topics underlying the diversity statements publicly released by Fortune 1000 companies in late May and early June 2020. The results uncovered six underlying latent semantic topics: (1) general DEI terms, (2) supporting Black community, (3) acknowledg- ing Black community, (4) committing to diversifying the workforce, (5) miscellaneous words, and (6) titles and companies. Furthermore, drawing from the identity-blindness and identity-consciousness theoretical frameworks and leveraging millions of data points of employees’ DEI ratings retrieved from Glassdoor.com, Study 2 further tested and supported hypotheses that companies were more positively rated by their employees on organizational diversity and inclusion if they (1) released (vs. did not release) diversity statements and (2) emphasized identity-conscious (vs. identity-blind) topics in their diversity statements. Our findings shed light on important theoretical implications for the current research and offer practical recom- mendations for organizational scientists and practitioners in diversity management.

Keywords Diversity · Equity · Inclusion · DEI · Corporate statements · Text mining · Machine learning

On February 23, 2020, a 25-year-old unarmed Black man, Ahmaud Arbery, was shot and killed by two people claiming to make a “citizen’s arrest” (Fausset, 2020). Within the next 3 months, the murders of Breonna Taylor and George Floyd at the hands of police officers sent shockwaves through the USA (Cave et al., 2020; Oppel et al., 2021). These incidents exacerbated deeply entrenched racial tensions throughout the country. Nationwide protests and social media campaigns

demanded accountability of those responsible and gal- vanized people to acknowledge the existence of systemic racism against Black Americans and its devastating conse- quences (Cave et al., 2020).

Indeed, the senseless deaths of Black Americans at the hands of law enforcement have reignited calls and move- ments for social justice across the globe. During this time, scholars across disciplines have sought to make sense of these sweeping events and their potential consequences. To better conceptualize such phenomena, Leigh and Melwani (2019) extended the construct of social mega-events (Tilcsik & Marquis, 2013) and proposed the concept of a mega- threat: a negative, large-scale, diversity-related episode that receives significant media attention, which occurs when an individual or group is targeted, attacked, or harmed because of their social identity group, and that event is then highly publicized (Leigh & Melwani, 2019).

Although there exist many forms of mega-threats in recent history, the current research specifically focuses on

* Wei Wang [email protected]

1 The Department of Psychology, The CUNY Graduate Center, City University of New York, 365 Fifth Ave, New York, NY 10016, USA

2 Baruch College, City University of New York, New York, NY, USA

3 Florida International University, Miami, FL, USA 4 University of North Carolina at Greensboro, Greensboro,

NC, USA

/ Published online: 31 May 2022

Journal of Business and Psychology (2023) 38:45–61

1 3

the mega-threat of racism as it is a large-scale phenomenon involving intergroup and inter-racial behaviors that have direct important theoretical and practical implications for organizational research. Within this nascent area of the research of mega-threats of racism, scholars have predom- inantly focused on the intra-psychic and group effects of mega-threats. For example, group members, who identify with the targeted social group, may experience cognitions and emotions that change the relationship between their identities and behaviors (Leigh & Melwani, 2019), and Black Americans may suffer vicarious trauma from expo- sure to police violence (e.g., Anthym & Tuitt, 2019; Boykin et al., 2020). As such, these lines of research suggest that mega-threats have lasting effects on individuals and groups.

Despite these important findings and theoretical advance- ments, however, little research has systematically explored how organizations respond to the mega-threat of racism or examined the corresponding implications for organizational outcomes. Indeed, following the mega-threat of racism in 2020, many organizations released diversity statements (aka “DEI statements”), designed to denounce racism and affirm their stance on values of diversity, equity, and inclu- sion (DEI). Scholars have previously examined how these public-facing messages are presented, as well as how they impact their public image and particular groups (e.g., Kaiser et al., 2013; Leslie, 2019; Nishii et al., 2018). However, to our knowledge, none of these lines of inquiry have examined diversity statements as a response to acute societal events, such as the mega-threat of racism.

This gap in the literature deserves closer inspection for several reasons. Under normal circumstances, organizational diversity statements may fly under the radar, alongside other public-facing communications. However, mega-threats represent potential sea changes within society, which con- sequently amplify the importance of these diversity state- ments. Indeed, the mega-threat of racism can influence how organizations perceive and brand themselves, communi- cate with their stakeholders, and ultimately conduct human resource management and business. Such a situation thus compels organizations to develop their diversity statements carefully, given higher stakes than ever before. Therefore, we are interested in how these highly visible corporate messages are characterized and perceived during tumultuous times.

In addition, as these corporate missives provide a funda- mentally different type of organizational data with which to study long-standing issues of racism, thus, this wave of releasing corporate diversity statements has inherently cre- ated a novel opportunity for organizational research with societal impact. However, little is known about the specific content addressed by various diversity statements—particu- larly, how companies respond to a mega-threat of racism and consequently communicate their views on diversity vis-a-vis statements. More importantly, current literature has yet to

explore the potential effects of both releasing (vs. not releas- ing) a diversity statement and emphasizing certain particular topics within the statement. Indeed, when the mega-threat of racism looms large in the public consciousness, what happens when organizations publicly respond to, or fail to acknowledge, larger societal issues? As such, scientifically analyzing the quantity (release) and quality (content) of diversity statements can help advance both organizational research and practices.

Therefore, the goal of the current research is to system- atically examine corporate diversity statements as the most common form of organizational reactions to the mega-threat of racism, as well as the effects related to organizational out- comes, such as employees’ perceptions on the organizations. Specifically, we conduct two studies in the current research, taking an inductive and a deductive approach, respectively. In Study 1, we collected and analyzed a massive body of corporate diversity statements publicly released by Fortune 1000 companies in response to the George Floyd protests in late May and early June 2020. We were thus able to inves- tigate: What are the major topics/themes conveyed by these major corporations in their diversity statements responding to the mega-threat of racism? Then, in a follow-up study (Study 2), we drew from a well-established identity-con- sciousness theoretical framework and leveraged millions of data points of employees’ online ratings on diversity and inclusion. We hypothesized and tested whether or not com- panies that released (vs. did not release) a diversity state- ment tended to be more favorably rated by their employees on organizational diversity and inclusion and how emphasiz- ing different latent topics in a statement may differentially impact important organizational outcomes.

Across these studies, we make several key contributions to the diversity literature and theoretical advancement. First, we scientifically and systematically assess and identify the latent semantic topics underlying diversity statements, which enables us to better taxonomize how different organizations respond to the mega-threat of racism. This type of character- ization provides an empirically based foundation for future research and practice. Second, and relatedly, by novelly applying the identity-conscious (acknowledging group iden- tities) vs. identity-blind (minimization of intergroup differ- ences; Leslie et al., 2020; Plaut et al., 2014) framework to the topics that emerged, we advance our current understand- ing of how diversity statement composition may provoke differing reactions within stakeholders. Although the iden- tity-conscious vs. identity-blind theoretical framework has been widely used within the DEI literature to describe the nuances of and responses to diversity messaging, it has yet to be applied to systematically characterize different themes within organizations’ diversity statements. Thus, the cur- rent research expands the theoretical understanding of the dichotomy framework.

46 Journal of Business and Psychology (2023) 38:45–61

1 3

Third, we examine how diversity statements may impact organizational stakeholders, specifically employees, in the immediate aftermath of a mega-threat. During a mega-threat of racism, these communications (or lack thereof) can shape employee perceptions, which may then influence their deci- sions to select into and remain engaged with their organi- zation (Schneider, 1987). To this point, Glassdoor.com’s recent addition of “diversity and inclusion” as a new metric of organizational satisfaction reflects its growing impor- tance in the workforce. This study, therefore, examines how organizational treatment of diversity may result in measur- able impact through large-scale assessment of unsolicited employee reactions. Thus, we provide scientific evidence in the diversity literature regarding important organizational effects of diversity management and advance the research in this area.

Finally, by investigating how employee ratings may reflect if and how organizations address diversity in their public- facing messages, our research clarifies which diversity messaging topics are positively associated with stakeholder perspectives (e.g., employees’ ratings) and, thereby, offers practical recommendations for organizational scientists and practitioners on designing effective DEI communications.

Study 1

In our first study, we utilized unsupervised machine-learning models to text mine and analyze how organizations com- municate about DEI topics in response to the mega-threat of racism. Given that this work is motivated by serious social and societal issues, we begin our literature review by out- lining the real-life and evidence-based foundations for this research by highlighting the organizational motivations for releasing diversity statements in responding to the mega- threats of racism; then, we discuss the importance of under- standing the text topics underlying the diversity statements is critical for organizational research.

Releasing Diversity Statements as an Organizational Response

The mega-threat of racism in current-day society has become unignorable, with many organizations experiencing mount- ing pressure to respond appropriately (Gupta & Briscoe, 2020) and confirm that they share their stakeholders’ val- ues. As a result, organizations have become increasingly motivated to respond through DEI initiatives, particularly by publicly releasing diversity statements—a type of official corporate document that emphasizes diversity-related prac- tices, such as equal opportunity employment and/or values (Leslie, 2019). Such statements often go beyond mention- ing affirmative action policies and speak more about how

diversity is valued and managed in the organization, tending to result in more positive attitudes among women and/or racial/ethnic minorities (e.g., Highhouse et al., 2009; McKay & Avery, 2006; Williams & Bauer, 1994). We believe that organizations’ motivations to release diversity statements may be well understood through the lenses of impression management theory (Highhouse et al., 2009) and signaling theory (Connelly et al., 2011; Spence, 1973).

Organizations have a vested interest in cultivating a positive social image. Reputation has been thought to enhance many favorable organizational outcomes, including performance (e.g., Cable & Graham, 2000; Dowling, 2002; Fombrun, 1996; Fombrun & Shanley, 1990; Roberts & Dowling, 2002). To this end, impression management theory posits that organizations may engage in strategies in order to earn approval and respect (Highhouse et al., 2009). These tactics, which include advertising, public relations, and social responsiveness (Fombrun & Shanley, 1990), may be used to demonstrate that organizations value more than financial profit. One approach to earning general approval and respect is to release public statements condemning and/or supporting events in the public consciousness. According to signaling theory (Connelly et al., 2011; Spence, 1973), a company must send out a signal to resolve any information asymmetry between itself and its stakeholders. That is, the company (signaler) can provide necessary perspectives in their communications to the public (receivers). Once the receivers process and respond to the signal, they send potentially affectively charged feedback back to the signaler. In the case of the mega-threat of racism, companies can release a statement to make the stakeholders aware of the information they are unlikely to access on their own —that is, the company’s stance on DEI issues. Corporate signals, such as those conveyed by releasing diversity statements, may be met with positive benefits, including increased customer loyalty and employee commitment (Riordan et  al., 1997). Moreover, acknowledgment of the mega-threat of racism may demonstrate a company’s investment in important societal questions and may therefore position organizations as responsible and responsive to mega-threats.

Identifying the Underlying Text Topics in Corporate Diversity Statements

Racism is a sensitive and complex issue that can mean many things to different parties; perspectives can be colored by unique experiences, relationships, salient group identity, political ideologies, the ability to empathize, and a multitude of other factors (Emerson & Murphy, 2014; Purdie-Vaughns et al., 2008). Given the sensitivity and misunderstanding around racial issues, as well as DEI broadly, conceptualiza- tions of and statements about DEI may differ substantially from one organization to another. Therefore, it is helpful to

47Journal of Business and Psychology (2023) 38:45–61

1 3

understand how organizations are conceptualizing DEI in their publicly released statements in response to the mega-threat of racism. Although there are several aforementioned theories that explain why organizations might release diversity state- ments, questions remain regarding the language and framing companies use to make these DEI communications. That is, it is essential to investigate not only if companies release state- ments, but also how they address such a multifaceted and com- plicated topic in their statements.

Indeed, the nuances of diversity statements play a critical role in the overall reception of the message and the reputa- tion of the organization. Generally speaking, critics of social justice movements have responded to diversity initiatives with skepticism and even backlash. For example, color-blind approaches, which ignore differences between groups, often draw negative reactions (e.g., Cheng et al., 2019). Relatedly, diversity communications may treat racism in a vague manner in order to make unpleasant realities more palatable. During the most acute days of the mega-threat, diversity statements would vary in how they acknowledged racism, ranging from blatantly and explicitly to, at least, inadvertently and vaguely. In our research, we expand on this thread by examining to what extent diversity messages specifically address the groups targeted by the mega-threat of racism.

Indeed, the murders of 2020 made clear that Black Ameri- cans are especially marginalized and vulnerable to acts of vio- lence and aggression—but how explicitly did organizations call attention to this? Did firms discuss these racism-related issues in broad strokes (e.g., as generic terms) or was lead- ership more explicit about their role in the maintenance of these systems? Were organizational statements identity-blind or identity-conscious (i.e., explicitly naming and discussing the discrimination faced largely by Black Americans)? Essentially, how, qualitatively, did organizations navigate and embody these socially sensitive topics in the public arena? These questions, and others, point to the ambiguity surrounding DEI framing, highlighting the lack of scholarship and empirical understanding in this sensitive time period. Therefore, as an initial step in this arena, we aim to text-mining analyze corpo- rate operationalizations of DEI by exploring:

Research Question 1: What are the major latent topics/ themes underlying the corporate diversity statements pub- licly released in response to the mega-threat of racism (e.g., George Floyd protests)?

Study 1 Method

Collecting Corporate Diversity Statements

Our research team manually searched and collected diversity statements from the Fortune 1000 companies.

Every year, in approximately July or August, Fortune Magazine ranks and publishes the largest US companies by revenue, known as the Fortune 1000. We focused on the Fortune 1000 metric as it included all major compa- nies across all industries. For the current study, we first obtained a complete list of 2020 Fortune 1000 companies from https:// fortu ne. com. The list ranked the companies from 1st to 1000th and provided important company information.

Using the Fortune 1000 list, we undertook extensive searches for diversity statements or open letters released by each company that addressed racial injustice. The goal of our search was to be as comprehensive as possible, with search strategies emphasizing the companies’ state- ments that were motivated by the killing of George Floyd. Thus, all the collected statements were released by organ- izations after May 2020, in the context of the mega-threat of racism. To complete the search, two graduate research assistants were trained to separately search for state- ments by each company, one by one, on the list. They first searched by using the following terms, and then all pos- sible combinations thereof: “racial equity,” “diversity,” “inclusion,” “George Floyd,” “open letter,” “newsroom,” “news press,” and “press release.” These were crossed with each company name, CEO name, and public rela- tions department. The graduate research assistants per- formed searches using tools and websites that are publicly accessible, including Google, Bing, Yahoo, Twitter, and the company’s own websites, particularly on the com- pany’s press release webpage. If a company released fol- low-up statements after their first one, we only included the first statement in order to preserve consistency. See an example of the statement released by Walmart’s CEO Doug McMillon in Appendix 1. After each graduate assis- tant completed the search, they compared the lists and resolved any potentially different results. This process resulted in 511 statements from Fortune 1000 companies (see Table 1 for a summary).

Analytical Strategy: Topic Modeling

Topic modeling—a type of statistical model aiming to understand the hidden topics underlying a collection of documents—is a relatively new method developed in the machine learning (ML) and natural language processing (NLP) areas. Although it appears complex, the statistical logic behind topic modeling is rather straightforward. It cal- culates the probability of how different words occur together in a document, and, based on the probability of word co- occurrence, classifies words into different groups, which are then labeled as topics. For example, in a collection of documents, one may find that “doctor” and “nurse” more

48 Journal of Business and Psychology (2023) 38:45–61

1 3

frequently appear in a medical document, and “dog” and “cat” appear more often in a document about animals.1 Thus, in the collection of these documents, we may find two topics emerge: the medical topic and the animal topic.

In practice, topic modeling is guided by two general prin- ciples (Silge & Robinson, 2017): (1) every topic is a mixture of words—e.g., the medical topic includes the words doctor and nurse; and (2) every document is a mixture of topics— e.g., a document is 85% medical topic and 15% animal topic, while another document is 30% medical topic and 70% ani- mal topic. As such, topic modeling categorizes words into different groups to form topics and calculates the probabili- ties of each topic in a document. Specifically, in the current study, topic modeling was used to statistically identify word groups underlying all the diversity documents released by

the Fortune 1000 companies and also calculate how much each company’s diversity statement emphasized each topic.

Although various methods have been developed for topic modeling, in the current study, we used Structural Topic Models (STM; Roberts et al., 2014),2 an advanced text- mining technique that has been widely used in social sci- ences such as management and political science. Similar to the classic topic modeling methods such as Latent Dirichlet Allocation (LDA; Blei et al., 2003) and Correlated Topic Model (CTM; Blei & Lafferty, 2007), the STM method takes an unsupervised machine-learning approach to identify and organize latent topics based on the semantic structure in a textual corpus. In other words, the STM model statistically classifies similar words together to form a latent semantic topic—similar to the factor analysis (FA) method that clas- sifies similar items together to form a latent factor and also estimates the probability that a document is associated with

Table 1 Summary of the 2020 Fortune 1000 companies included in the study

Note..aIn million USD

Sector N Representative companies Avg. rank Avg. N of employee Revenuea Growth CEO race

Non-White White

Aerospace and defense 5 Boeing, Lockheed Martin 141.40 89,860 $41,450.22 3.22% 0 5 Apparel 13 Nike, Ralph Lauren 532.15 27,765 $8,425.35 3.16% 1 12 Business services 31 Visa, ManpowerGroup 478.74 33,921 $8,320.95 7.22% 2 29 Chemicals 12 PPG, Ecolab, Chemours 505.25 15,980 $7,124.58 0.36% 1 11 Energy 39 Chevron, Duke Energy 387.79 11,000 $14,803.62 -0.34% 2 37 Engineering and

construction 8 AECOM, Granite

Construction 522.75 26,838 $8,113.24 6.34% 0 8

Financials 96 Berkshire, JPMorgan, MetLife

382.72 30,332 $23,017.57 10.07% 6 90

Food and drug stores 5 Walgreens, Kroger, Publix 250.40 196,009 $64,271.48 -5.94% 0 5 Food beverage and tobacco 21 PepsiCo, Tyson, Coca-cola 327.71 40,851 $17,526.06 4.14% 1 20 Health care 42 Johnson & Johnson, Pfizer 366.67 54,295 $31,179.11 11.40% 6 36 Hotels, restaurants, and

leisure 19 Starbucks, McDonalds,

Hilton 493.89 108,895 $8,555.39 10.26% 0 19

Household products 15 P&G, Kimberly-Clark 568.27 25,119 $10,580.60 1.43% 1 14 Industrials 20 Caterpillar, Deere 445.25 40,188 $14,368.98 5.83% 0 20 Materials 13 International Paper,

Westrock 484.15 24,167 $8,164.09 3.32% 0 13

Media 11 Disney, DBS, Discovery 560.45 30,891 $12,927.91 11.05% 0 11 Motor vehicles and parts 6 Ford, GM, Tesla 238.67 100,296 $58,135.25 -2.93% 0 6 Retailing 47 Walmart, Amazon.com 411.89 128,807 $34,419.81 0.97% 3 44 Technology 68 Apple, Alphabet, Microsoft 471.21 38,608 $21,127.10 7.27% 20 48 Telecommunications 7 AT&T, Verizon, DISH 265.29 92,432 $66,415.01 0.04% 0 7 Transportation 21 UPS, FedEx, Delta Airlines 423.10 65,697 $18,337.08 2.02% 0 21 Wholesalers 12 US Foods, Synnex 318.50 37,176 $15,289.88 14.27% 0 12 Sum (mean) 511 – – (58,054) ($23,454.92) (4.44%) 43 468

1 Please note that the categorization is not mutually exclusive, as some words can be cross loaded on multiple categories. In other words, a word can be simultaneously related to more than one topic. For this “doctor”/ “dog” example, words such as “food” can be mean- ingfully associated with both medical and animal topics.

2 More technical details can be found via the website http:// www. struc tural topic model. com/

49Journal of Business and Psychology (2023) 38:45–61

1 3

a certain topic (a.k.a., topic prevalence). However, compar- ing LDA and CTM, the major advantage and innovation of the STM method is its ability to further model the relation- ships between the topic prevalence and document-level vari- ables (a.k.a., metadata; e.g., company size and CEO race). Please see such analyses in the Supplemental Materials.

The STM modeling involved four computational steps, and all the analyses were performed in R Statistical Pro- gramming version 4.1.0 (R Core Team, 2021). The first two steps were inputting the text data and document metadata, preparing and pre-processing the data, removing stop words (e.g., “a,” “the,” etc.) and punctuations, stemming words (e.g., converting words “diverse,” “diversely,” “diversity,” “diversified,” “diversification” to the stem “divers”; convert- ing words “inclusion,” “inclusive,” “inclusiveness” to the stem “inclus”; etc.). Some words and documents were also removed in this pre-processing step because of extremely low frequency. For example, infrequent words that only appeared in one document were dropped for the subsequent analyses. A document with less than ten words was removed as well. This pre-processing step resulted in 469 documents/ companies for the final text modeling analyses.

After preparing and pre-processing the text data, the third step was estimating STM models. In this step, we included metadata in the model, which included the company level variables such as industry sector, Fortune 1000 rank, the number of employees (company size), revenue growth from the previous fiscal year, CEO race and gender, and corporate political orientation. To normalize highly skewed variables and improve model convergence, the number of employees was logarithm-transformed and normal standardized, the rank variable was Z-scored, and the revenue growth was cube root transformed. The fourth and last step was evaluat- ing and selecting models, in which we first ran models with a various number of topics ranging from 2 to 15. Then, follow- ing the guide by Roberts et al. (2014), we selected a model with the best model fit based on the criteria of semantic coherence and exclusivity. Semantic coherence is concerned with the maximum probability of a set of words in a given topic co-occurring together (Mimno et al., 2011). Exclu- sivity balances word frequency across topics based on the FREX metric—the weighted harmonic mean of