Leadership In Turbulent Times: Women CEOs During COVID-19

There is gender affinity for key concepts and words

We also focused on organizing words based on their gender affinity, either tilting toward women or men for average use per transcript (see chart 1 at the beginning of this report). In the chart, we see the distribution for each word’s log-ratio, contrasted with its total frequency usage across the entire corpus. The more a term was used, the less affinity it had for gender. These include common English terms such as “the” and “of.” However, as we look at the more intermediate frequent words, we see differences arise. Some differences are due to gender representation in sectors like health, education, oil, and software. Some commonality occurs regarding crisis (covid-19) or general company growth and performance. However, some major differences arise between smaller variations, such as for the term “covid-19_crisis” (that more men used) over an “outbreak” (which more women used). Even in expressions of emotions, differences arise for terms such as “pessimistic” versus “enthusiasm.” For our dataset, for terms occurring over 100 times, 15.7% were women-associated (log ratio less than -0.5), 26.5% were associated with men (log ratio greater than 0.5), and 57.8% were shared (log ratio is between -0.5 and 0.5).

Women CEOs demonstrate a leadership style that values empathy, adaptability, accountability, and diversity

We then used the dataset of CEO words to see if they could tell us something about leadership style. We built a model to give us the top 30 contextually relevant words for each component (see chart 3). The components of leadership styles we considered were growth, adaptability, communication, flexibility, accountability, empathy, performance, diversity, transaction, transformation, and cultural. These are all grounded in leadership style and theory.

We found that women CEOs expressed language in the categories of empathy, adaptability, accountability, and diversity, for example. Men CEOs were more focused on transaction-related words (see chart 3). Mentions of words related to performance and growth appear to be gender-neutral in earnings calls. In the context of flexibility, we found women CEOs talking about agility, comfort, and financial flexibility (see table 4), while men CEOs were mentioning predictability, redundancy, stability, and scalability. For instance, in the leadership style of transformation, women CEOs spoke more about realignment, revitalization, simplification, modernization, and operational excellence, while men CEOs talked more about reorganization, integration, and cultural transformation.

Women CEOs spoke more about clients, while men CEOs talked more about metrics

Next, we wanted to test the premise that men CEOs are more prone to discussing metrics, which would show a performance-dominant leadership style. Indeed, we found that men CEOs more frequently used terms related to metrics like “EBITDA” and “profit.” Interestingly, women CEOs leaned toward terms related to customers (see chart 4).

Finally, we delved into terms related to family and labor, finding that words such as “state_of_emergency” and “unemployment” are highly associated with men, in contrast to concepts such as “paternity” and “maternity” for women (see chart 5).

Insights Regarding Women CEO Representation

We are very far from gender equality at the CEO level

Katie Darden, Financial Institutions Research Director, S&P Global Market Intelligence

Women CEOs continue to be significantly outnumbered globally by CEOs who are men. In fact, as of Jan. 25, 2021, only 5% of a sample of more than 8,000 CEOs from the S&P Global BMI were women (see table 2). This shows that there is still a wide gender gap in the highest corporate ranked position, in addition to the already known gender gap in corporate boards. In those countries where we observed more gender parity at corporate boards, there were deliberate policies implemented throughout the years. Women around the world continue facing increasing challenges, and companies will play a critical role in retaining and developing their talent and their capacity to assume leadership positions.

Countries with a more gender-diverse labor market do better at fostering women CEOs

Marion Amiot, Senior Economist, and Nicole Serino, Credit Market Research, S&P Global Ratings

The number of women CEOs does not come close to matching the number of men CEOs in any of the 61 countries covered by the S&P Global BMI companies whose CEO data we analyzed. This suggests that more progress is needed worldwide to close the gender gap at the leadership level. That said, some countries are ahead of others in terms of CEO diversity. Looking at the S&P Global BMI, Norway and Singapore score the highest with women representing 14% and 12% of women CEOs, respectively (see chart 6). By contrast, Japan and Brazil have the lowest share of women CEOs at only 0.8% and 0%, respectively (see chart 7).

The composition of the labor force can explain part of these differences. Countries exhibiting a relatively higher share of women CEOs also tend to have a more gender-balanced labor force, that is, similar rates of labor market participation for men and women (see chart 8). A simple linear regression suggests that the gender gap in the labor force explains about 13% of the cross-country variation in the CEO gender gap. To close that gap, policymaking could play a role. For example, countries with better access to child care are more likely to attract and retain a higher share of women in the workforce. Only then can larger numbers of women start a career, climb the ladder, and one day, perhaps, become CEO.

Taking a more sector-specific look at the CEO gender gap, data suggest that some industries seem to have more women CEOs than others. For example, in the S&P Global BMI, real estate and health care companies are around four times more likely to have a women CEO than a company in the energy sector (see chart 9).

Factors accounting for company performance during the COVID-19 crisis

To analyze whether CEO gender helps explain companies’ performance in the first year of the pandemic, we turned to stock market performance. We would have preferred a broader measure of performance, but since we are looking at a recent time period, we were constrained in our choice of indicators. Most companies had not yet released their annual results when the analysis was conducted. Thus, we measured stock market performance by looking at companies’ total returns between March 10, 2020, and Dec. 31, 2020. Equally, our choice of control variables for the model was constrained by the availability of data. We had information only about each company’s sector and country and the CEO’s age and experience in the company (that is, years in the position).

Using a quantile regression, we find that CEO gender doesn’t explain companies’ total returns performance. This is regardless of which of the five percentiles where the company's performance lies--from best to average to the worst--see table 3. This corroborates some of S&P Global’s previous findings in “When Women Lead, Firms Win,” and underscores that gender does not determine success. Putting these findings in the light of previous research would also suggest that diversity among a group of leaders is more likely to shape performance than the gender of one leader. What likely matters more than gender is a mix of both experience and views, a conclusion from our previous study, “The Changing Face of Tech.”

Economic sector and geographic location best explained company performance during the COVID-19 pandemic. Also, companies with CEOs with more experience in their current job (proxied by job tenure) performed better, highlighting that knowledge of the business may have helped. Our model also suggests that companies with relatively younger CEOs tend to have performed better. However, we believe age here is likely an indicator of the degree of digitalization of the companies. In other words, younger CEOs (i.e., younger than our sample average of 56) were more likely to work in those digitalized companies that fared better or even benefited from this crisis. Companies that were more digitalized and able to swiftly shift to work from home during lockdowns suffered less activity losses during the pandemic regardless of their geography or sector, which are the other factors we have controlled for in our model.

Looking Ahead

Daniela Brandazza, Senior Director and Analytical Manager, S&P Global Ratings, President of the WINS (Women’s Initiative for Networking and Success) at S&P Global; Katie Darden, Financial Institutions Research Director, S&P Global Market Intelligence

Much of the business community is warming to a stakeholder capitalism model that emphasizes a company’s creation of value in society and the economy, instead of a narrower focus on value created primarily for shareholders. Under this approach, employees are one of many stakeholders. A commitment to an increasingly diverse and inclusive workforce and leadership is part of that vision.

Even if companies do not consider employees as stakeholders, the consideration of their needs is obviously a good practice for attracting and retaining talent. Prior S&P Global research shows the importance of flexibility and adaptability as elements of corporate cultures and policies that employees view are best in class (“Something’s Gotta Give”). To the extent that CEOs demonstrate a commitment to such ideas through their public communications, it may benefit other leaders to follow their example. As our research shows, women CEOs may be leading the way in this regard even if they remain small in number.

The main conclusions of this global research about women's leadership shed light on their communication style during a moment of crisis, relying on data from relatively formatted communications (conference calls about corporate earnings). We see several opportunities for further research about women leaders in times of crisis and their role in creating social capital. Several studies have shown that organizations appoint proportionally more women to head up their organizations in times of crisis (Glass & Cook 2016; Barreto et al. 2009; Bruckmüller & Branscombe 2010). This bias, known as the glass cliff, is a point to be followed up by research that might, for example, revisit the data after the pandemic is over. Similarly, investigating the differences established here between the CEO communication styles of women and men leaders is relevant to exploring how this key element of leadership influences social capital (Day 2000). That is, how do CEOs promote a set of shared values to achieve a common purpose? Conducting a series of interviews with women CEOs who lead companies in the S&P Global BMI can help to do this. We are putting these two points on our research agenda.

Understanding women leaders as role models and the value they bring to our society is part of a changing vision of a world that needs to be more diverse, equitable, and sustainable. Our research indicates that countries and companies still need to do much more to ensure such a vision in the future. Foremost, they need more equitable participation in the workforce, which our research shows is related to higher shares of CEOs who are women. This is unsurprising but comes at a time when women are reportedly pulling back or out of the workforce in greater numbers to care for others in response to COVID-19-related needs for health care, child care, or education. This represents a huge economic loss and a missed opportunity when companies and countries ignore women’s potential. As we have shown, women CEOs, though small in number, have a broader notion of leadership that places greater emphasis on stakeholder leadership in an era when stakeholder capitalism is taking root. In this sense, women CEOs seem to be in the vanguard.

Appendix I: How We Constructed Our CEO And Transcripts Databases

Katie Darden, Financial Institutions Research Director, S&P Global Market Intelligence

S&P Global conducted this analysis based on S&P Global Market Intelligence professionals data, market data, and transcripts data for the constituents of the S&P Global BMI.

Using the S&P Capital IQ and S&P Global Market Intelligence platforms, we screened for constituents of the S&P Global BMI and compiled CEO data for those companies on Jan. 25, 2021. This resulted in a list of more than 11,200 individuals whom we categorized by gender.

A similar analysis, for which data was compiled on Feb. 8, 2020, provided a point of comparison for the Jan. 25, 2021, data. The resulting more limited comparative sample included CEOs for nearly 8,500 companies worldwide. These companies are S&P Global BMI constituents for which CEO data was available through S&P Global Market Intelligence as of the compilation dates and for whose CEOs we were able to apply a gender category in both years.

We consider that these dates--Feb. 8, 2020, and Jan. 25, 2021--serve adequately as bookends to the first unsettling year of the COVID-19 crisis. The first date preceded by slightly more than one month the date when the World Health Organization declared the COVID-19 outbreak to be a pandemic. While we cannot consider the second date to represent the end of the pandemic, which continues around the world, it came as mass vaccination efforts had begun.

We identified the gender of the individuals covered in the analysis based on several factors, including honorifics, personal pronouns, and given names. The categorization methodology prioritized gender identification based on honorifics, followed by pronouns. If neither of these markers were available in individuals' biographical data, we used first, or given, names for gender categorization. First name-based gender determinations were largely determined using Genderchecker.com and Genderize.io databases. We looked for possible indicators of gender neutrality and gender identities other than men and women during the classification process.

Assumptions and limitations

As part of the categorization process, we made certain assumptions about gender identity based on these markers. For example, we categorized a CEO as a man if that person’s professional biographical description used the honorific title “Mr.” or the personal pronoun “he,” absent contradictory indicators. We looked for possible indicators of nonbinary gender identities, such as the use of personal pronouns other than “he” and “she,” as well as honorific titles, such as “Mx.,” that might indicate a gender identity other than “man” or “woman.” Even so, we recognize that our assumptions that certain pronouns, honorifics, and given names correspond to certain gender identities may be flawed. It is our hope that the benefits of our research for furthering diversity outweigh potential errors in this classification process.

For the natural language processing-enabled transcript analysis described in more detail below, we identified nearly 5,000 companies from our larger Jan. 25, 2021, list that met all the following criteria:

• CEOs from the Jan. 25, 2021, dataset were hired before March 2020, and
• Transcripts with CEO commentary were available through S&P Global Market Intelligence.

Appendix II: How We Analyzed Earnings Call Transcripts

William Watson, Data Scientist, S&P Global Ratings

Preprocessing

For preprocessing, we applied the spaCy tokenizer pipeline per transcript, then a Gensim phrase transformer to merge significant bigrams and phrases as a single token. Here, a token was considered a single unit, which could be a word (example: employee) or a bigram (example: independent_contractor) (Mikolov et al., 2013). In addition, standard English connection words (such as: of, for, on, and) were included for phrases (examples: bank_of_america, working_from_home). This allowed us to reduce the ambiguity among individual words that compose larger concepts.

Sentiment and Emotion analysis

We aimed to identify and compare overall sentiment (positive or negative) as well as eight emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust). To achieve this, we used the NRC Word-Emotion Association Lexicon (EmoLex) (Mohammad & Turney, 2010, 2013) that contains over 14,182 unigrams (single word units) and 13,901 binary associations. Note that a word can have multiple associations or no associations at all. The number of words associated with each concept in the lexicon are:

• Sentiment: negative: 3,324, positive: 2,312
• Emotion: fear: 1,476, anger: 1,247, trust: 1,231, sadness: 1,191, disgust: 1,058, anticipation: 839, joy: 689, and surprise: 534

Our objective was to capture both sentiment and emotion from transcripts of earnings calls to show how women and men CEOs led during the COVID-19 crisis. For each transcript, the emotion and sentiment scores were computed based on the EmoLex, akin to a frequency count of words associated with a particular emotion or sentiment. These were then normalized by the number of tokens in the transcript. Our analysis looked at average word use in transcripts published on March 9, 2020 (a few days before the official WHO declaration of the pandemic on March 11, 2020). We also carried out a time-series analysis of the evolution of sentiment and emotion in those transcripts from Jan. 1, 2019, to Dec. 31, 2020, to capture usage before the pandemic, so we could make more informed conclusions about the effect of the COVID-19 crisis on word choice in the transcripts.

In the following table, we list the top 15 words in our analysis for each emotion and by gender. We see that both genders expressed the toll the pandemic was taking in their use of words for the emotions of anger, disgust, fear, sadness, and negative. For example, “pandemic,” “disease,” and “crisis” are some of the most frequent words for both genders in these categories. Small differences appear for the most frequent words, indicating uniform usage between gender for this frequency band. We also display the top 10 words with the most skewed log-ratio scores per gender for each emotion, i.e., the most gendered terms. Note that the words themselves have an association with an emotion, and are not just synonyms for said emotion. For example, the word “cash” has associations with anger and joy, reflecting different contextual meanings for the same word.

Word embeddings for contextually relevant terms and concepts

To quantify and operationalize leadership styles and concepts for CEOs, we first needed to extract what could be relevant terms or phrases for each concept. There are many ways to describe the same event. For example, COVID-19 can also be referred to as the coronavirus, the COVID outbreak, the corona pandemic, or C-19. However, these are functionally the same concept and should be included as one “topic.”

To build each topic, we created a Word2vec model using all words in each transcript (Mikolov et al., 2013). Using the Skip-Gram algorithm, this model learns associations among words within a context window of 5 tokens. This results in a high-dimensional set of word vectors, where vectors that are close together in the embedding space have similar contextual meanings and distant vector pairings have different meanings. The trained word embedding provides a meaningful way to suggest the context or perspective around a given word. For our purpose, we query the model for a concept token, for example, “empathy,” and retrieve a list of similar words such as “compassion,” “teamwork,” “humility,” and “kindness.” This allows us to aggregate a more robust topic. By resolving the different ways CEOs refer to the same topic, this method provides a more holistic view of what they are discussing.

Normalized log-ratio for gender affinity

With our set of relevant terms and phrases for each topic, we then measured affinity for each token against our classes: men and women. For this analysis, we used the following algorithm from the field of computational linguistics (Hardie, 2014a, 2014b):

• Calculate word frequency per transcript
• Normalize the frequency by the transcript token length
• Average the normalized frequency per gender
• Compute the log-ratio between the average for women versus men

This procedure provides several benefits. First, word use is normalized per transcript to accommodate words used exclusively by a single person, as the lack of usage elsewhere in the class will penalize the average. Second, by averaging transcript frequencies, we reduce the bias that would stem from unbalanced vocabulary sizes (there are more men in our dataset, so the variety and amount of words are significantly higher). Third, we smooth the ratio comparison to avoid outliers from non-occurrence by a factor of 1e-6.

Log-ratios show the magnitude and direction for a gender affinity. In our example, since the numerator is men, log-ratios that are positive simply indicate words that men use more, and negative log-ratios indicate words that women use more. This choice was arbitrary. The metric is symmetric if flipped. To reduce the effect of outliers, we aggregate the top 30 contextually relevant words queried from the embedding model in our calculation.

Limitations

We recognize there are certain limitations regarding transcript data availability for the transcript analysis presented in the previous pages, which significantly reduces the sample of companies from the S&P Global BMI. As a result, our transcript data is not entirely representative of the S&P Global BMI and our sample has a different geographic and sectoral distribution. For example, 50% of the companies extracted in the sample for the transcript analysis are based in the U.S. (see chart 10), while the S&P Global BMI is a more global index. The data used for this analysis also displays a lower share of women CEOs overall and a slightly different sectoral distribution (see chart 11). We recognize that there is a sector bias in the distribution of women and men in the dataset: Women CEOs transcripts are almost twice as likely to come from health care companies than in the men’s sample. Men CEOs are more likely to work in financial, industrials, and technology companies.

The way CEO phrases were classified might also have limitations. We worked on phrases to discover associations, to be agnostic to CEOs themselves. Although the methodology aims to reduce the possibility of unconscious bias influencing how each statement was classified, we recognize there might be some limitations about how the classifications or associations were made.

References

Avolio, B., Walumbwa, F. & Weber, T. 2009. “Leadership: Current Theories, Research, and Future Directions.” The Annual Review of Psychology. 60, 421-449.

Barreto, M., Ryan, M. K., & Schmitt, M. T. (Eds.) 2009. Psychology of women book series. “The glass ceiling in the 21st century: Understanding barriers to gender equality.” American Psychological Association.

Bell, S. 2020. “The stock market boost from having more women in management European groups with higher numbers of female leaders benefit from stronger share prices.” Financial Times, Nov. 9, 2020.

Bruckmüller, S. and Branscombe, N.R. (2010). “The glass cliff: When and why women are selected as leaders in crisis contexts.” British Journal of Social Psychology, 49: 433-451.

Chamorro-Premuzic, T. & Wittenberg-Cox, A. 2020. “Will the Pandemic Reshape Notions of Female Leadership?” Harvard Business Review. June 26.

Chin, J. L. 2014. “Women and Leadership.” The Oxford Handbook of Leadership and Organizations, edited by David V. Day, Oxford Library of Psychology.

Day, D. V. (2000). “Leadership development: A review in context.” Leadership Quarterly, 11, 581-613.

Eagly, A. H. & Johnson, B. T. 1990. “Gender and Leadership Style: A Meta-Analysis.” Psychological Bulletin, Vol. 108, No. 2, 233-256.

Eagly, A. H., & Carli, L. L. 2003. “The Female Leadership Advantage: An Evaluation of the Evidence.” Leadership Quarterly, 14: 807-834.

Fine, M. G. 2007. “Strategic planning: Gender, collaborative leadership, and organizational change.” In J. L. Chin, B. Lott, J. K. Rice, & J. Sanchez-Hucles (Eds.). 2007. Women and leadership: Transforming visions and diverse voices, Malden, MA: Blackwell.

Foels, R., Driskell, J. E., Mullen, B., & Salas, E. 2000. “The effects of democratic leadership on group member satisfaction: An integration.” Small Group Research, 31, 676–701.

Glass, C. & Cook, A. 2016. “Leading at the top: Understanding women's challenges above the glass ceiling.” The Leadership Quarterly, 27, 51-63.

Mason, E.S. 1991. “Gender differences in job satisfaction.” The Journal of Social Psychology, 135(2), 143-151.

Morin, G. 2016. “The making of a leader and of leadership: analysis of transformation processes in three large organisations,” doctoral thesis, University of Panthéon-Assas Paris II.

Saint-Michel, S. 2010. “Le genre et le leadership. L’importance d’introduire les traits de personnalité des leaders.” Revue internationale de psychosociologie et de gestion des comportements organisationnels, 40(XVI), 181-201.

Saint-Michel, S., & Wielhorski, N. 2011. “Style de leadership, LMX et implication organisationnelle des salariés : le genre du leader a-t-il un impact?” Revue @grh, 1(1), 13–38.

Sandberg, D. 2019. “When Women Lead, Firms Win.” S&P Global, Oct. 16.

Wood, J. 2001. Gendered Lives: Communication, Gender, and Culture (4th Edition). Wadsworth Thomson Learning, Belmont, CA.

Hardie, A. 2014a. “Statistical identification of keywords, lockwords and collocations as a two-step procedure,” paper presented at ICAME 35 Corpus Linguistics, Context and Culture, 49. Nottingham, U.K.

Hardie, A. 2014b. “Log Ratio – an informal introduction.” ESRC Centre for Corpus Approaches to Social Science (CASS). Accessed April 27, 2020, at: http://cass.lancs.ac.uk/?p=1133.

Mikolov, T., Chen, K., Corrado, G., & Dean, J. 2013. “Efficient Estimation of Word Representations in Vector Space.” Proceedings of Workshop at ICLR.

Mikolov, T., Sutskever, I., Chen, K. Corrado, G. & Dean, J. 2013. “Distributed Representations of Words and Phrases and their Compositionality.” Proceedings of NIPS.

Mikolov, T., Sutskever, I., Chen, K., Corrado, G. & Dean, J. 2013. “Distributed Representations of Words and Phrases and their Compositionality.” Proceedings of NIPS.

Mikolov, T., Yih, W., & Zweig, G. 2013. “Linguistic Regularities in Continuous Space Word Representations.” Proceedings of NAACL HLT.

Mohammad, S. & Turney, P. 2013. “Crowdsourcing a Word-Emotion Association Lexicon.” Computational Intelligence, 29 (3), 436-465.

Mohammad, S. & Turney, P. 2010. “Emotions Evoked by Common Words and Phrases: Using Mechanical Turk to Create an Emotion Lexicon.” Proceedings of the NAACL-HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, June, Los Angeles, California.

Sullivan, A. et al. 2020. “The Changing Face Of Tech.” S&P Global, March 5.

Zenger, J., & Folkman, J. 2020. “Research: Women are better leaders during a crisis.” Harvard Business Review, Dec. 30.