The impact of intersectional racial and gender biases on minority female leadership over two centuries

Our basic conceptual model is based on the cross-disciplinary behavioural data science approach. Consider a 3-stage dynamic problem with the timeline depicted on Fig. 1. The leader goes through 3 stages in her development: Identification, Progression, and Achievement.

During the Identification stage, the “Nature” determines a set of priors, which it assigns to the future leader. The future leader considers the amount of risk she wants to take in her leadership career by deciding on the number of areas or fields she wants to enter. Since each entry is inherently costly, requiring the leader to take on multiple visible responsibilities, and, therefore, it is also risky: the higher is the number of entered fields, the higher is the risk tolerance capacity of the future leader. The number of entered fields is observable (e.g., it is easy to collect information on whether a particular leader decided to concentrate on one particular field such as law or decided to diversify into several fields by entering the field of law and, in parallel, engaging into activism and then politics). Yet, the decision to enter multiple fields may result from multiple factors such as suffering from the intersectional invisibility (i.e., being overlooked as a non-prototypical representative of a cultural group) due to gender and racial stereotyping¹⁴, as well as many latent factors such as individual psychological characteristics, family context, preferences, beliefs, developmental opportunities, etc.^20,21. It also may be the case that the future leaders consider the experience of previous leaders with similar cultural characteristics (i.e., they may have role models from their cultural group), yet, the degree of knowledge or awareness about these experiences as well as the way in which future leaders perceive these experiences are unobservable. Nevertheless, all these constructs may impact a leader’s decision-making process, which is why it is important to account for these possible effects in our empirical analysis.

In the Progression stage, the leader engages into leadership activity in organizational context through translating her set of leadership priorities by verbalizing her behavioral schemas. These schemas help the leader to set goals for herself as well as her followers and update the schemas once the goals are achieved. The verbalized schemas are measurable by considering the topics, on which leaders concentrate in their rhetoric as well as the speed with which these topics change over time. Apart from the gender and racial stereotyping, the leaders may be affected by a number of individual and contextual observable as well unobservable factors, which we are accounting for in our empirical analysis by controlling for individual effects as well as time effects.

Finally, in the Achievement stage, the leader gains prominence and is accepted by the broader society (not just by her immediate following within her organization) as a leader. At this stage, the observable construct is usually success level, which may have different measures from specific variables (such as the level of corporate profit, shareholder value, return on investment, etc.). It may also be as simple as a binary variable (where success is assumed to be 1 and failure is depicted as 0) or continuous (where success is distributed between 0 and 1). Similarly to previous stages, a number of important latent effects need to be taken into account in the empirical analysis. In our empirical strategy (described below) we assume that all leaders in our dataset are equally successful (in other words, we keep the leadership success level constant and assume it to be equal to 1 as this assumption most appropriately reflects our dataset).

To give a specific example of Black female leader, Fannie Lou Hamer’s trajectory from a Mississippi sharecropper to a nationally recognized leader epitomizes the three-stage leadership model²⁸. Initially, during the Identification stage, Hamer confronted racial injustices and risked her safety to register to vote, later diversifying into civil rights activism with the SNCC and CORE. In the Progression stage, her grassroots leadership style and impactful speeches, especially during the 1964 Democratic National Convention, illuminated the barriers faced by Black voters. Transitioning to the Achievement stage, Hamer shifted focus to economic disparities, co-founding the Freedom Farm Cooperative and advocating policies promoting Black economic upliftment, education, and health, solidifying her legacy in American civil rights history.

Considering the three stages described above, our basic behavioural data science model can be specified as follows. Consider an individual leader k, characterized by a set of cultural and socio-demographic priors. We concentrate on female leaders with different racial and ethnic backgrounds, hence, we assume that \(k\in \{i,j\}\), i.e., \(k=i\) if the female leader identifies herself with an ethnic minority and \(k=j\) if she identifies herself with the ethnic majority. The leader’s success (achievement) is depicted by \(s_k^T\), which depends on the leader’s risk tolerance capacity \(r_k^t\) as well as her behavioral schema updating measure \(\sigma _k\), capturing the relative importance of a particular topic of focus for the leader over time. This measure can be defined in many ways, but we use the average standard deviation of probabilistic measure of importance of all topics in leadership rhetoric over \(N\) number of time periods (years) as specified in subsequent sections. In other words, \(\sigma _k = \frac{1}{\eta } \sqrt{\left( \sum _{\tau =t+1}^{T-1} \left( e_k^{\tau }-\mu \right) ^2 \right) /N}\), where \(e_k^\tau\) is each probabilistic value of topic importance from the pool of important topics considered in leadership rhetoric, \(\mu\) is the pool mean, N is a number of time periods and \(\eta\) is the number of important topics, which represent verbalized behavioral schemas of the leaders. Therefore, the leadership success (achievement) is determined by a combination of a leader’s ability to take risk (i.e., given by her risk tolerance capacity) combined with her experiences (i.e., experientially informed and constantly updating behavioral schemas) and is given by: \(s_k^T=r_k^t+\sigma _k+\varepsilon\) where \(\varepsilon\) is a normally distributed noise parameter, which captures any possible stochastic shocks to the leadership trajectory.

In understanding the leadership journey of Black women, it is imperative to consider the intersectionality of race and gender. Black women often face a unique set of challenges and experiences as they navigate leadership roles. This intersectionality does not simply mean they experience the additive challenges of being Black and being a woman, but rather, they face specific issues arising from the combination of these identities. For example, they might encounter racialized sexism or gendered racism in their roles. In our study, we specifically delved into these unique challenges and how they shape the leadership styles, strategies, and decisions of Black women.

We recognize that there is a significant difference between being a Black woman leader and being a Black woman leader who actively advocates for Black issues. While all Black women leaders bring their lived experiences into their leadership, not all choose or have the opportunity to center their leadership around advocacy for Black issues. Our research methodology differentiates between these two categories, ensuring we capture a holistic understanding of Black women’s leadership. We have ensured that our study both recognizes the diversity within Black women leaders and understands the specific nuances of those who are advocates for Black issues.

Justification of assumptions

Concentration on gender and ethnicity

Even though intersectional stereotyping in organizations may arise from multiple intersectional systems in the society (race, gender, socio-economic class, ability, age), we concentrate on the ethnic background and gender for several reasons. First, the impacts of gender and ethnic background are often overlooked compared to other combinations such as, e.g., gender and age, etc. As Crenshaw puts it using an example of Black female leaders, “Black women’s intersectional experiences of racism and sexism have been a central but forgotten dynamic in the unfolding of feminist and antiracist agendas…”²⁹. Second, the effects of gender and race are difficult to measure due to the fact that, unlike many other factors, both gender and race involve self-attribution. Unlike other characteristics, that are easy to measure objectively (e.g., age), an individual needs to associate herself with women and with a particular minority group in order for intersectionality analysis to be valid. Yet, our unique dataset allows us to focus on leaders, who identify themselves as females and as representatives of a particular racial and ethnical background.

Stability of gender and ethnical background identification

Our model assumes that gender as well as racial and ethnical identification remain constant throughout all three stages of the leader’s development. In practice, this may not always be the case as gender, racial, and ethnical identification may change over time. For example, a particular leader with a mixed ethnical background, e.g., Austrian and Japanese, may first identify as an Austrian because she was raised by her family as an Austrian. Yet, later is life she may decide to identify herself as Japanese. The leader may also undergo a trans gender transition through her life, e.g., first being known as a man and later as a woman. Consider, for example, the Wachowskis—leaders in the film making and authors of the Matrix movies—who were first known as men and are currently known as trans women. These processes are rare and complex and, most importantly, their intersectional effects are not very well understood in the literature, which is why we concentrate on women leaders with stable gender, race, and ethnicity identification over time.

Separability of risk tolerance determination and experiences

Our model implies that the leader’s risk tolerance capacity determination and behavioral schemas updating occur in different stages of a leader’s development (i.e., are time-separable). Specifically, the leader first makes a decision about how many fields she wants to enter (determines her risk tolerance capacity) and then proceeds to having experiences in these fields, updating her view of the word (behavioral schemas) and verbalizing these schemas as they emerge and update. This, however, may not always be the case. For example, the leader may initially decide to enter only one field (e.g., law) and through her experiences then decide to later enter another field (e.g., politics). The time separability is not an important assumption for our model: in principle, the Identification and Progression stages can be happening at the same time as long as we acknowledge that leaders’ risk tolerance capacity determination and behavioral schemas as products of experiences are distinct components of the leader’s development. Considering risk tolerance updating as a part of behavioral schemas updating would be an interesting extension of our model, which does not consider how the updating is happening (i.e., in order to incorporate risk tolerance updating into a model one would first need to explain the exact mechanism of behavioral schemas updating, which is outside the scope of this paper). We discuss this as a limitation of our model in the concluding section of this paper and provide several suggestions about how the updating process could be modeled in the future studies. Nevertheless, recent evidence from studies on female leadership and female ethnical minority leadership supports the validity of the separability assumption. Using qualitative interviews, these studies find that women leaders often make decisions to have multiple roles, take on many responsibilities, and diversify their efforts across several fields prior to engaging in leadership experiences in order to tackle gender and racial discrimination or to become more visible^13,14,30,31.

Empirical methodology and testable hypotheses

Even though our model could be applied to the individual leaders, we are particularly interested in racial and ethnic minority female leaders as a group. Instead of looking at the difference in leadership success (achievement) given leader’s characteristics, our empirical strategy is to hold the level of success constant and explore the differences between risk tolerance capacity and experiential behavioral schemas of majority and minority female leaders. We focus on women who all achieved societal endorsement in the sense that they became widely known and accepted by people within and outside their organizations and fields as leaders. For simplicity, if 0 would constitute the failure to achieve societal acceptance as a leader and 1 would constitute success, we assume that in our sample all female leaders achieved \(s_k^T = 1\) irrespective of their background. This means, that the success level of minority and majority leaders is the same, i.e., \(s_i^T=s_j^T\). This implies that \(r_i^t + \sigma _i + \varepsilon = r_j^t + \sigma _j + \varepsilon\) or, considering that \(\varepsilon\) is normally distributed and does not depend on whether the leader is or is not a representative of the racial or ethnic minority: \(r_i^t+\sigma _i=r_j^t+\sigma _j\). This allows us to formulate a number of testable hypotheses, as if \(\sigma _i<\sigma _j\), \(r_i^t\) should be greater than \(r_j^t\). Alternatively, if \(\sigma _i>\sigma _j\), \(r_i^t\) should be lower than \(r_j^t\). There is also a theoretical possibility that both risk tolerance capacity and experiences are the same for both majority and minority leaders, i.e. if \(\sigma _i=\sigma _j\), then \(r_i=r_j\). Since we observe that female minority leaders are disadvantaged compared to the female majority leaders¹³, we expect their leadership experiences to be a lot less positive than those of the female majority leaders, meaning that minority female leaders should update their behavioral schemas less often than majority female leaders. This implies that minority female leaders are forced to concentrate on the same priorities and verbalize the same behavioral schemas over multiple time periods, whereas majority female leaders are able to change their priorities often. Therefore, the standard deviation of the probabilistic measure of topic importance should be lower for the minority leaders than for the majority leaders. Hence, our first hypothesis could be formulated as follows.

Hypothesis 1

Verbalized behavioral schemas of minority female leaders should be less time-variant and less dispersed than those of the majority female leaders.

This means that we should observe that \(\sigma _i<\sigma _j\), i.e., the mean standard deviation of the topic probabilities (measured by the topic modelling in the leadership rhetoric) should be lower for the minority compared with the majority group. If minority leaders generally have less positive leadership experiences, this means that they should be compensating for the lack of positive progress in achieving their goals by taking more risk and concentrating on more fields than their majority counterparts. A recent longitudinal study approached 59 Black females twice over the course of 7 years and found through a series of qualitative interviews that these women tackled career challenges associated with intersectional invisibility by what we describe as a more risk-taking behavior¹⁴. Specifically, these women took on more responsibilities and engaged in larger number of visible leadership roles to progress in their careers. Importantly, the interview evidence suggests that these women first made the decisions about the number of roles and responsibilities and then proceeded to their experiences. Another study focused on minority female leaders in Pakistan, the United Kingdom as well as Brazil, who revealed through a set of qualitative interviews that they needed to deliberately plan to take more responsibility and engage in more roles before they could proceed to having leadership experiences³². Further studies have confirmed that the necessity for higher risk taking becomes apparent from an early age as future minority leaders are encouraged to respond to more opportunities by their families³⁰ as well as opt for more visible and diverse set of responsibilities throughout the school year³¹. Therefore, our second hypothesis can be formulated as follows.

Hypothesis 2

Minority female leaders tend to take more risk with their leadership careers than their majority counterparts.

Considering the lack of improvement in female minority leadership over time, our model also allows us to formulate the third hypothesis:

Hypothesis 3

Since experiences improve for majority, but not for minority female leaders over time, minority leaders (as a group) should take progressively more risk than majority leaders.

Data mining and analysis

One of the main constructs in our behavioural data science model is the construct of a behavioral schema, which allows the leaders to comprehend, internalize, and understand their goals and priorities. These priorities need to be communicated to others and if they are achieved, the leader develops a new set of behavioral schemas, which, in turn, need to be communicated again. The extant psychological and leadership literature stresses the importance of effectively expressing and communicating behavioral schemas to others: leaders often verbalize schemas as strategic directions, values, and vision, to achieve development of organizations, and empowerment of followers^33,34,35. Hence, leadership rhetoric is suggested as a good proxy to measure leaders’ dynamic behavioral schemas^36,37,38,39. We concentrate on female leaders’ rhetoric and communication.

Our strategy is to concentrate on successful female leaders, who not only self-select into the leadership roles, but also achieve societal recognition as leaders. We use an inclusion into a leadership repository of high achievers as a proxy of success, which implies that women represented in these repositories are endorsed not only by their direct followers in organizations, but also by the broader society as being “worthy” of being included in the repository. We obtained data from two repositories: the Iowa State University Archives of Women’s Political Speech and the Gifts of Speech. Both repositories are hosted by universities and inclusion into the repositories is subject to expert review. The Iowa State University Archives of Women’s Political Speech is hosted by the Iowa State University, a public university in Ames, Iowa, and collects specimen of speech in English from prominent women who through their whole careers or at some point in their careers were leaders in activism, politics, civil service, public life, or assumed other leadership roles of power and public importance. The Gifts of Speech is hosted by Sweet Briar College, a private women’s college in Sweet Briar, Virgnia, and collects speech specimen in English from prominent female leaders representing a wide variety of fields. Though both repositories aim to suggest role models to the future generation of (female) leaders and a college of experts is consulted before including a particular individual into each of the repositories, the Iowa State repository includes women who are widely accepted as leaders by the American public (for example, prominent female American business women, lawyers, politicians, activists, journalists, etc., are included in the repository), the Gifts of Speech repository appears to apply additional success criteria such as winning an important global award such as the Nobel Prize or an well-known specialized prize such as Fields Medal, etc.

Our focus on Black women leaders who are widely recognized by the general public was chosen for its accessibility and relatability. However, we acknowledge the potential limitations inherent in focusing solely on these leaders. Historical and contemporary structures may determine which Black women are allowed or chosen to be in the limelight. For instance, certain personality traits, leadership styles, or even appearances might be deemed more “palatable” or “acceptable” by mainstream society, thereby allowing some Black women to rise to prominence over others. In our analysis, we aim to understand the systemic, societal, and cultural barriers and enablers that have shaped the public perception and acceptance of Black women leaders.

By combining the speech specimens from the two repositories, we obtain a unique dataset of female leadership speech. Yet, there are several important aspects about the dataset, which should be noted and which we take into account when conducting our analysis. Initially, we have mined all specimen of text available from both repositories and obtained 3,207 specimens of text in total. The only mining criteria we apply is that the latest date of the text specimen should be December 31, 2019. We deliberately avoided collecting text specimens from 2020 and 2022 due to the prevalence of COVID-19-themed speeches in those two years. While analyzing data from 2020 and 2022 is of interest, it would be more appropriate to consider these data in a separate investigation. The female speeches in the text format were collected using scripts coded in Python 3.7.7. Of 757 women in our total database, 608 (80.3 percent) were American (see Supplementary Materials for raw data). Considering our concentration on the comparison of majority and minority leaders, we concentrated on 608 female leaders from the US. Apart from collecting specimens of leadership speech, the Iowa State repository also contains specimens of political advertisement. Political advertisements are samples of text representing transcripts of televised advertisements used by politicians as a part of their election campaign. These text specimens are not suitable for our analysis as they are produced for the purpose of winning the election, have the goal of attracting attention to a particular individual rather than describe this individual’s agenda, and often contain direct speech from other individuals. Therefore, we excluded all specimens of political advertisement obtained from the Iowa State repository and concentrated on speeches only.

For inclusion into each repository, the speech has to go through the review scrutiny by an editorial committee. As a result, there are a number of speeches which are selected for both repositories. We removed the duplicated speeches, which appeared in both repositories and obtained unique 2181 specimens of speech from 608 women leaders from the US. Speech data was added to the database.

Speeches hold a pivotal role in the realm of leadership⁴⁰. They not only provide a platform for leaders to convey their vision, strategies, and goals but also help in building rapport with their audience. The form of a speech, encompassing its structure, language, and delivery, reflects the leader’s intent, preparedness, and approach to their subject. It becomes a mirror to the leader’s mindset, revealing nuances about their priorities, concerns, and aspirations. Functionally, speeches serve multiple purposes for leaders. They act as tools of motivation, education, and persuasion. Leaders use speeches to inspire teams, educate stakeholders about shifts in strategy or market dynamics, and persuade audiences to align with their viewpoint or vision. The content of a speech can influence public opinion, drive organizational change, or even reshape industry perspectives. Considering our dataset of female leadership speech, it is important to note that the speeches collated provide a rich tapestry of insights into the leadership styles, communication strategies, and priorities of the women leaders represented. This collection, while vast and diverse, serves as a unique reflection of the evolving dynamics of female leadership in the US, especially when observed through the lens of its form and function.

From the repository speech data, we obtained the name of individual, date of speech, place of speech, exact text of the speech, and occupation of the individual at the time of the speech, type of organization where an individual worked at the time of the speech. These data were also merged with publicly available information about each speaker. This information included race, nationality, as well as profession or professions. Considering that engaging into a new profession requires risk taking, the number of professions obtained and mastered by each female leader was taken as a proxy of their individual risk parameter. These publicly available data were mined from the Wikipedia pages as well as from the official website of each woman leader (where available).

Nationality information was available for all 608 women, only 4 of whom had dual nationalities (US plus one other country). In terms of ethnical and racial composition, of the 608 women in our sample, 462 identified themselves as White, 87—as Black or African American, 10-Asian, 18-Hispanic or Latino , and 10 had other single race background. In our sample, 21 women had either mixed background or unknown background. Specifically, we could not find background information about 10 women, and 11 women were identified having several racial or ethnical backgrounds simultaneously, which made it difficult to place them in any one group. Information about the racial background was compiled from three main sources, which were cross-checked to form a unified classification: (i) a women leader’s own words in speeches, (ii) Wikipedia pages, as well as (iii) relevant lists (e.g., list of African American politicians, etc.). Black or African American group represents the largest minority group in our sample (see Fig. 2).

Figure 2 shows minority female leaders in our database. Speeches from majority and minority leaders covered the period from 1828 to 2019. We obtained women leaders’ occupation information (main area of expertise), which was cross-checked using several sources, including the relevant speech repository pages about female leaders (Iowa State and Gifts of Speech repositories) and Wikipedia biography pages. In addition, we also collected data about all areas in which a particular female leader was a recognized expert (i.e., information about a particular leader’s profession). Hence, for each female leader in our database we had the main occupation and all areas where she was considered an expert. The speeches captured in the sample from this group spanned a period from 1851 to 2019 (i.e., pre-COVID19). As a result, we captured a diverse group of Black or African American women leaders, who represented a wide range of areas from activism and feminism to science and technology. Figure 2 showcases this heterogeneity.

Women in our database had different number of speech specimens per person. The average number of speeches per female leader was equal to 3.32 with the median of 1 and a standard deviation of 8.05. Over 70 percent of women in our sample had 1 speech specimen. There were 8 female leaders in the database with the number of specimens greater that 18: Carly Florina with 23 specimens, Madelaine Albright with 26 specimens, Elizabeth Dole with 27 specimens, Joni Ernst with 33 specimens, Carrie Chapman Catt with 45 specimens, Michelle Obama with 54 specimens, Elizabeth Warren with 64 specimens, and Hillary Rodham Clinton with 181 specimens. Importantly, many speeches from these female leaders were recorded in the same calendar year. For example, Hillary Clinton is the largest outlier in terms of number of speeches in our sample. Even though her speeches span a time period from 1969 to 2019, 64 of 181 of her speeches are from 2016 when she was running a presidential election campaign in the US. Similar pattern is observed for other outliers: for example, most speech specimens from Madeleine Albright (15 of 26) are from 1997 when she became the Secretary of State. In order to mitigate issues arising from multiple speeches per leader, our analysis takes into account the fact that some text specimens are from the same person (i.e., assumes that specimens from the same female leader are correlated). Furthermore, our text analysis is dynamic in the sense that it first considers the average trend in text from each individual in a particular year and only then calculates the overall behavioral schema trend for the whole year.

Our analysis is initiated with the utilization of a dataset comprising female leadership rhetoric for conducting a topic modelling exercise. Our corpus, designed specifically for this topic modelling, encompasses 2,181 speech samples attributed to 608 female leaders, collated between the years 1828 and 2019. This corpus includes over 3 million words. Following a pruning process that eliminates punctuation, frequently used function words, single occurrence words, and university names as well as organizational names, the total word count of the corpus stands at 1,000,401.

The final corpus contains both individual and temporal effects, necessitating consideration of the correlation of speech samples from the same leaders as well as the potential correlation of topics across time periods, with our unit of time being one year. Our approach addresses individual effects via the topic modelling process and temporal effects at a post-topic-modelling stage. This process is required since existing topic modelling methods are incapable of addressing both effects simultaneously. For instance, while Latent Dirichlet Allocation (LDA)-based Correlated Topic Model (CTM) procedures permit modelling of topic correlations, they are unable to model correlations between text samples from the same author. Similarly, Dynamic LDA permits the capture of topic evolution within a sequentially organised document corpus but cannot be jointly estimated with Correlated LDA due to its use of different distribution models.