Hiring women into senior leadership positions is associated with a reduction in gender stereotypes in organizational language

Study 1 aimed to test whether hiring a female CEO is related to an increase in the association of women with agency between the periods before and after hire (Fig. 1). To capture organizations’ outward-facing language, we extracted 1) DEF 14A filings (definitive proxy statements), 2) 10-K filings (annual reports), and 3) transcripts of investor calls for all S&P (Standard and Poor's) 500 organizations in a 10 y window between 2009 and 2018. We identified the subset of these organizations that had 1) a female CEO in the sample period (41 organizations), 2) text data available for at least 3 y prehiring and posthiring, and 3) a male CEO predecessor prior to the female hire. This selection procedure resulted in a total of 11 target organizations. For each of these 11 target organizations, we identified two propensity matched S&P 500 organizations that did not hire a female CEO, allowing us to account for changes in gendered language that were not related to the hiring of a woman as CEO. We also performed this same process for female CEOs who were replaced by male CEOs in the eligible period. This provided two additional target organizations and four propensity matches. In total, we had a main sample of 33 organizations and an additional sample of 6 organizations that we analyzed separately in SI Appendix.

We measure the change in the gendering of organizations’ language before and after hiring a female CEO using an unsupervised natural language processing algorithm [word2vec (27–29)] that estimates the semantic meaning of references to gender and references to agency. The intuition behind word2vec is reflected in the quote by the linguist John Rupert Firth, “You shall know a word by the company it keeps” (30). Seeing a word across many different contexts (sets of surrounding words) enables the algorithm to learn the word’s meaning from the contexts in which it is used. word2vec can be thought of as analogous to running factor analyses that uncover latent dimensions on which different words vary systematically. The number of these dimensions is specified in the model, but there is no a priori specification of how particular words will relate to or “load on” these dimensions. The output of the word2vec algorithm is a dictionary containing a vector for every word in the text, with a value for every dimension specified.

Due to the complexity of the space, no individual dimension of a word vector is interpretable. However, studying differences between vectors can yield remarkable insights regarding the semantic meaning of words and contribute to the understanding of gender stereotypes (31–33). In fact, the study of analogical reasoning using word vectors has highlighted the inherent gender biases in language, showing, for instance, that historically “man is to computer programmer as woman is to homemaker” (34, p. 1). In the present research, we study the distances between gendered words and agency words in vector space. This allows us to capture the strength of association between what it means to be a woman and what it means to be agentic in organizations’ language before and after the hire of a female CEO.

We estimated the female–agency association and male–agency association in the following four steps (see Methods and SI Appendix for further details): First, we developed and validated an agency dictionary of 100 words (e.g., capable, independent, and dominant) and identified 20 female-related words that could be matched with a direct male counterpart (e.g., she/he, woman/man, and her/his). Second, we used the official text of all 657 organizations that had entered the S&P 500 between 2009 and 2018 but were not in our set of 33 focal organizations (or the 6 other organizations), to pretrain a set of vectors capturing the semantic meaning for each of the words in the two dictionaries. This model was trained based on 12,974 SEC (US Securities and Exchange Commission) documents and 27,557 transcripts with a total of 1.15 billion words. Third, we updated the vectors based on the data of each of the 33 focal organizations separately for the periods before and after hiring the CEO (with an average number of 37 new documents containing 980,000 words per organization period). Finally, we calculated the average cosine similarity between the word vectors representing references to being female or male and the word vectors representing the meaning of agency, for each organization and prehiring/posthiring time period.

We first explored our hypothesis—that the hiring of a female CEO would be associated with a strengthening of the semantic relationship between female and agentic vectors in the target organizations relative to the propensity-matched organizations—in a descriptive way. Specifically, we visualized the grand mean-centered cosine similarity between our female and agency vectors, at an organization level, in the periods before and after the female CEOs were hired (Fig. 2). For example, comparing General Motors—which hired Mary T. Barra in 2014—and Ford Motor Company shows that while Ford saw a drop in the female–agency association after 2014, General Motors saw an increase. Providing initial support for our hypothesis, Fig. 2 shows that 8 of the 11 target organizations (72.7%) saw an increase in the association between the semantic meaning of female and agency words, while only 6 of the 22 control organizations did in the same time period (27.3%).

To test our hypothesis more formally, we used linear regression analysis, regressing either the female–agency association or the male–agency association in organizations’ language onto the binary coded time period (prehiring = 0, posthiring = 1) and the organization category (0 = propensity match, 1 = target), as well as their two-way interaction. Our main variable of interest, the interaction term (β₃), indicates the extent to which target organizations experienced a stronger increase in the association between gender words and agency words relative to their propensity-matched peers.

To increase the robustness of our findings, we used 27 different specifications (3 gender dictionaries, 3 agency dictionaries, and 3 dimensions of word vectors) and drew 1,000 bootstrapped samples for each specification from our 11 target and 22 propensity matched organizations, resulting in a total of 27,000 iterations. In Fig. 3, we present the standardized regression coefficients (betas) of the interaction term from the 27,000 model estimates, both for the female–agency and the male–agency effects (ordered by effect size). Providing support for our hypothesis, the findings show that the prehiring–posthiring change in female–agency association was significantly larger for organizations that hired a female CEO [average β₃ = 0.226, SD(β₃) = 0.076]. With the vast majority of standardized regression coefficients being positive (99.996%), the distribution of coefficients was significantly larger than zero (CI₉₅ = [0.2246, 0.2264], t = 489.96, df = 26,999, P < 0.001). We further found that the results were robust to drop-one analysis, showing that the effects were not driven by a single organization, and replicated without the use of bootstrapping (see SI Appendix for full details of these analyses).

In addition, we show that the increase in similarity between the semantic meanings of women and agency was not merely a product of the CEO (who is now a woman) being described performing agentic actions as part of their role. The event of a woman being hired as CEO means that public organizational documents are likely to refer to her and her actions as CEO. Hence, the observed effects could be driven by organizations describing the actions of their female CEOs, rather than strategically or incidentally associating their female leaders—or women more broadly—with agentic traits. We alleviated the concern that the observed changes could be limited to descriptions of CEOs’ behavior in a number of ways. First, we found that there was no negative effect on the male–agency association when female CEOs were hired [Fig. 3; average β₃ = 0.033, SD(β₃) = 0.071], which would be predicted if our results were purely driven by changes in the gender of the actor who is described performing the role of the CEO. Second, the effects replicated in direction and magnitude when controlling for the relative frequency of female references in our training documents [average β₃ = 0.220, SD(β₃) = 0.077, t = 470.2, df = 26,999, P < 0.001], showing that the contexts in which women are described explain our results, rather than the quantity of female references. Third, we found that the agentic terms driving the effect were predominantly adjectives and not action verbs (see Fig. 5). Finally, we aligned the semantic vector spaces for each organization between the prehiring and posthiring periods [using Procrustes analysis (35)] and examined word-level shifts, finding that the effects were driven by a congruence of the meaning of womanhood and agency, rather than a general change in semantic meanings or instability in the concept of gender (see SI Appendix for more information).

Similar to study 1, we used word embeddings to convert organizational documents into measures of gender stereotypes (women’s association with agency and communality). However, instead of focusing on the event of hiring a female CEO, we studied the relationship between the proportion of women on companies’ executive boards and the female–agency association longitudinally across 345 former and current S&P 500 companies. This approach allowed us to replicate the findings from study 1 in a way that 1) moves past a propensity-matched quasi-experimental design, 2) can disentangle the bidirectional relationship between female representation and gendered language, 3) shows that changes to gender stereotypes generalize beyond references to the CEO, and 4) highlights the benefits of female representation in senior leadership positions more broadly.

We created a longitudinal panel in the following steps: First, we randomly selected 345 of the 690 eligible organizations that entered the S&P 500 from 2009 to 2018 to pretrain a word2vec model using the same text data as study 1. Second, we updated word2vec models for the remaining 345 organizations using a 3-y sliding window (i.e., 2009 to 2011, 2010 to 2012, etc.), providing us with eight estimates per company. Third, we estimated the average cosine similarity between gendered words and agency/communality words for each model (akin to our approach described in study 1). Finally, we merged this panel with the average proportion of women on each organization’s board of directors for the relevant period as extracted from Institutional Shareholder Services (ISS) Director Data via Wharton Research Database Services (WRDS). After removing periods for which we either had no unique text documents or board data, we were left with 2,286 observations associated with 328 unique organizations.

We estimated panel vector autoregression [PVAR (36)] models to test how the proportion of women and the gendering of language affected each other without having to specify the direction of the effects a priori (see SI Appendix, Table S13 for a full model output). PVAR uses a model that contains both lagged endogenous variables and unobserved individual effects, allowing us to test for causal relationships with observational data (36, 37). However, the requirements for consistent estimation of the parameters of such a model are demanding as OLS estimates are not consistent (38). PVAR solves this problem by employing an instrumental variable strategy that uses lags of endogenous variables as exogenous instruments and using a general method of moments estimator (39, 40) with an appropriate model and moment selection procedure (41).

As expected, we found a significant positive effect of the lagged proportion of women on our measure of the female–agency association (b = 2.794, P = 0.039). A 1-SD shock to the proportion of women on an organization’s board of directors (roughly equivalent to appointing one additional woman) was associated with a 0.31-SD increase to the female–agency association across the following seven periods (see SI Appendix for further details). Importantly, as with the appointment of women as CEOs, the increase in the female–agency association in organizational language did not come at the expense of women becoming less associated with communality. The effect of lagged proportion of women on the board of directors on the female–communality association was, in fact, positive but nonsignificant (b = 1.517, P = 0.329). Further, we also found preliminary evidence for a relationship going in the opposite direction, with the effect of the lagged female–agency association on the proportion of women being positive and statistically significant (b = 0.006, P = 0.007). While this suggests that increases to the female–agency association may also precipitate future increased female representation, this effect was much smaller than that of the proportion of women on the female–agency association. Specifically, a 1-SD shock to the female–agency association was associated with a 1–percentage point increase to the proportion of women on the board. Yet the bidirectional nature of the effect highlights the potential for a virtuous cycle: the more women get hired into leadership positions, the more embracing organizational language becomes of women’s leadership competencies, and as gender stereotypes dissipate for women more broadly, the more likely organizations are to hire additional female leaders.

Overall, the findings from study 2 provide additional evidence for the relationship between female representation in leadership positions and a reduction of gender stereotypes in organizational language. We show that this relationship generalizes to a wider range of S&P 500 organizations. While we cannot make strong claims about whether the observed changes in language reflect a general shift in how organizations think about gender, study 2 shows that the mitigation of stereotypes extends beyond female CEOs to other women on other senior leadership roles. Although the changes could still be limited to the specific women appointed, board members are much less likely to be the sole subject of organizational documents, suggesting that the semantic changes may indeed generalize to women beyond CEOs and executive board members (see Discussion for more details). In addition, the findings highlight the bidirectional nature of the relationship between female representation and gender stereotypes.

The language we use to describe men and women speaks volumes and has consequences for stereotypes, career outcomes, and beyond. Here we aim to understand how semantic meaning changes as a function of appointing women as CEOs or to the board of directors, demonstrating that hiring a female CEO or board member is associated with shifts in the semantic relationship between what it means to be a woman and what it means to be a leader. Our findings contribute to the existing literature in three important ways. Theoretically, we contribute to research on gender equality and stereotype change. Over the past few decades, multimillion-dollar efforts have been devoted to breaking down gender norms and changing what it means to be a woman: from #LikeAGirl (42) to #ShesALady (43), there has been an effortful push to change the meaning of womanhood to incorporate agentic traits like confidence, power, and independence. Such attempts are meant to change stereotypes and increase women’s confidence and advancement into positions of power. Indeed, much empirical research has espoused the importance of stereotype change as a precursor to gender equality (2, 4). Yet, here we show that when it comes to stereotype change, female representation is not merely an end but also a means: simply hiring a female CEO or board member may shift the semantic association between what it means to be a woman and what it means to be a leader. Further, while previous research on backlash avoidance has demonstrated that individual behaviors (such as reframing agency as, or coupling agency with, communality) can mitigate backlash (44, 45), here we move beyond individual approaches and toward systemic ones. In our findings, increases to women’s agency are driven by the positive aspects of agency and occur without corresponding costs to communality. These findings suggest that when organizations are incentivized to portray their female leaders in positive ways, stereotypes can change in a way that is both systemic, removing the burden of stereotype change from the individual, and positive, in that it does not reflect a trade-off between appearing competent and likeable.

Methodologically, we contribute to the nascent literature of using word embeddings to quantify gender stereotypes and social change (32, 33, 46). Recent designs have used word embedding models trained on large corpora of historical text to chart how the understanding of concepts has changed over time. For example, Garg et al. (33) showed that women became more or less associated with different occupations (e.g., mechanic and housekeeper) over the course of the 20th century. In contrast, our study designs aim to isolate the effect of specific events (appointing women as CEOs and to the board of directors) on the meanings imbued in word embeddings at an organization level. This approach offers broader possibilities for understanding the relationships between the semantic meanings encoded in language and real-world outcomes and provides a template for further research studying word embedding models that are associated with specific organizations, institutions, and public figures.

Practically, this work contributes to a growing call and eager push for increasing women’s representation in positions of power, due to the many benefits this representation is expected to bring, both to organizations and to society. Women’s participation in management has been shown to empower subordinates (47), signal change (48), lead to better management of conflict (49), and improve decision-making and performance (50, 51). However, the benefits of hiring a female leader can cascade beyond these direct benefits conferred to their organizations. Female exemplars send signals to others: they empower other women (52), affect their career aspirations (53), and improve their performance in counterstereotypic fields (54). In the present research we identify an additional benefit to increasing female representation in positions of leadership: the potential for changing pernicious gender stereotypes that are encoded in language (4). Given the many hurdles women face along their career paths, changing such stereotypes is imperative in the push toward gender equality.

There are several key limitations of the present research that should be addressed in future work. First, to what extent is the observed effect causal? Although we cannot speak definitively to causality (as we cannot randomly assign female CEOs to organizations), we attempted to support causal inference in two ways. In study 1’s quasi-experimental design, the comparison of target firms to propensity matches controlled for both time trends and relevant firm-level factors. We found no evidence of changes to the female–agency association in the prehire period (SI Appendix, Fig. S15), suggesting that an unobservable factor would have to cause hiring before it began to affect language. Further, study 2 employs PVAR, which is a cross-lagged panel model that aims to isolate how two or more variables are causally associated with each other over time. Research has suggested that cross-lagged panel models may be able to support causal inference by accounting for confounds and reverse causation (55). The causal interpretation of PVAR models can be controversial: in the absence of an appropriate lag testing procedure, causal inference is inappropriate (36), and if the residuals in the specification are correlated, the coefficient estimates may not be consistent (37). Although these do not appear to be problems in our analyses, future research should seek to further strengthen the case for the causality of this relationship by pursuing additional methods of analysis. For a review of PVAR models, see Canova and Ciccarelli (56).

Second, to what extent do the effects generalize beyond the appointed female leaders and impact perceptions of women more broadly? Our theoretical reasoning suggests that female representation in senior leadership positions changes gender stereotypes by providing female role models, who become associated with leadership-relevant, agentic qualities. The changes to how the appointed women are perceived are then expected to transfer to the wider category—conferring greater agency to women in general. While we observe a shift in the gendering of language when women are appointed to senior leadership positions, we cannot definitively say whether this change extends beyond the characterization of the female CEOs and board members. However, our findings provide tentative evidence that this spillover effect is likely. First, as SI Appendix, Fig. S15 shows, the association between women and agentic characteristics continuously increases in the posthire period. If the effects were merely driven by organizational documents referencing the new female CEO, one would expect the effect to be strongest at the point of hiring. The continuous upward trend suggests that as the female leader fulfills her role, perceptions of women as being agentic may extend to women more broadly. In addition, the findings of study 2 are unlikely to solely be explained by direct references to board members, given that they are not typically discussed extensively during investor calls or in 10-K documents. However, future research should investigate this proposition more directly by trying to disentangle references to specific women leaders versus women in general, by studying the temporal dynamics of how perceptions of women change after the appointment of female leaders, and by expanding the data sources to organizational documents such as emails and minutes from meetings.

Third, at what level of seniority do the benefits of female representation cease to be present? Our findings show that appointing women as CEOs or executive board members is associated with a reduction in gender stereotypes in organizational language. However, it remains unclear whether we would expect similar effects when increasing the number of female middle managers, for example. At lower tiers of the organizational hierarchy, both the visibility of female leadership and the incentives for organizations to signal the competence (and warmth) of their female workforce might be significantly reduced compared to senior leadership positions. Thus, any changes to stereotypes may be either negligible or more ambivalent in terms of their valence and the extent to which they circumvent the typical warmth–competence trade-offs faced by women.

Fourth, to what extent do the results generalize to other organizations? Most notably, the sample of organizations that hired a female CEO in study 1 was relatively small (n = 11). This could mean that idiosyncratic features of these organizations might distort our results. Importantly, the convergent evidence of study 2 largely mitigates this concern, finding the same result across 328 organizations. However, our analyses are still exclusively focused on organizations from the S&P 500, which tracks large organizations listed in US stock exchanges. The effects of hiring women to senior leadership position may be different in other cultures (57) and in smaller organizations in which leaders are more directly embedded in the organizational culture. Similarly, the effect of appointing women to senior leadership positions in other domains—such as academia or the government—may produce different effects as a function of the particular context.

Finally, our inferences were based on use of the word2vec algorithm, which has recently been shown to produce word embedding estimates that might not always be stable across situations (58, 59). In order to address this potential shortcoming, we replicated our results across different initializations, vector sizes, and hyperparameter combinations (SI Appendix, Figs. S4 and S5). While the robustness of our effects increases our confidence in the validity of the results, future research should apply other natural language processing algorithms such as Global Vectors for Word Representation (GloVe) (60) to corroborate our findings.

Taken together, our findings suggest that female representation in leadership positions can induce language change, such that the semantic meaning of being a woman and being agentic become more similar. While previous work had suggested that changing the gendering of language could provide a fruitful avenue for changing gender stereotypes and discrimination in the workplace, the question of how to change language remained largely unanswered. Our results suggest that hiring female leaders might not only combat discrimination by making the top seem possible for other women but might also indirectly do so by changing our understanding of what it means to be a female leader.

We thank Michal Kosinski, Ruth Appel, Heinrich Peters, Jack Soll, Sanaz Mobasseri, Mike Norton, Hemant Kakkar, Daniel Keum, and Adam Galinsky for valuable feedback. We thank Aneesh Rai for providing data regarding the board composition of S&P 500 organizations. M.A.L. presented parts of this research at the 2021 Society for Personality and Social Psychology meeting and the 2021 Annual Meeting of the Academy of Management.