WHEN A.I. TELLS YOU TO SEGREGATE: A DIFFERENT KIND OF FOUNDER’S DILEMMA

March 1, 2019
and Founder, HereWe and PhD post-doc at Stanford

Surprising findings about diversity in tech companies and what it could mean for you.

Imagine you’re building the most intelligent, A.I.-powered, career manager. Your vision is to connect every single person with her or his ideal next job across companies, repeatedly. Because you’re building for the long-term, you can’t just focus on getting people jobs—you have to assure that what you build maximizes each person’s chances for success at the company where you place them. You’re working through the sea of data that you’ve amassed, and you discover that what a person looks like, their gender or their skin-color, significantly affects their trajectory once at the company. What do you do?

To improve our matches, we spent over a year collecting thousands of bits of information about every tech company that exists in the US, from Seed to Post-IPO. Our data comprise more than 13,000 tech companies, and generate almost as many questions. We found surprising correlations in the demographics of who founded the company, who was hired, and who was leading. The most surprising discovery? Tech companies of all sizes and at any stage suffer from what we call Divertia, or the inertia to hire generally the same “faces” that they started out with. Whether at 8 employees or at 8,000, the demographic composition of companies tends to stay as it was.

This is where we find ourselves at stellares.ai, and it has caused us endless debate. Our platform is already “diverse-aware”. We anonymize profiles and strip away information to employers like gender and name, all of which cause bias in the early stages of the hiring process. But we don’t just show every resume to every employer. We facilitate a match. While we want to place people where they will succeed, we are very aware of the tech industry’s gaps in diversity, and we don’t want to further cement the status quo

A question about the financial implication of a more or less diverse company naturally arises here. Research finds that there is one. Given the breadth and specificity of our data, we are exploring a separate deep-dive into this as well.

In the meantime, we can’t change what we can’t measure and talk about, so we decided to share our findings. We hope that, in the future, our algorithms and matching models won't need to take gender and skin color as statistically meaningful parameters for matching people with companies.

Until then, here’s the reality we live in:

To achieve 50-50 gender-parity in the leadership team, ALL of the founders need to be female, statistically speaking.

Equal representation can be a one-way street, in more ways than one. Of the 13,000 tech companies, when all of the founders are male, fewer than a fifth of the Leadership (defined as VP and above) are females. On the flip side, if all of the founders are female, then the executive suite is 50% male. And it’s a sliding scale. The higher the proportion of female founders, the closer to gender-parity the leadership layer is. This doesn’t mean there are no all-male teams with 50-50 gender balance in their leadership, but these are rare.

Moreover, the gender-balancing effect of more women on founding teams only holds for gender. When we looked in the data, the proportion of female founders did not significantly influence the proportion of racial/ethnic diversity in the leadership of the company.

There really isn’t a statistical reason for this disparity in leadership. All male-founded teams should be able to achieve as much gender-balance as all female-founded teams.

Impact of the founding genders’ mix on the gender mix of their leadership teams

tl;dr: If you're a woman, go work for a founding team that includes at least another woman. The chances that you will get to lead there are much higher. Alas, only 17% of all venture-backed startups have at least one female founder.

WYSIWYG—when HR has more women or people of color, the company does too.

What You See Is What You Get. HR, or People Ops, is a microcosm of what you can expect to see in the rest of the company. We were surprised to find this strong of a correlation. When the proportion of females in HR increased, the proportion of females in leadership roles in the company also increased.

Comparing the % of women in HR to the % of women at other levels of tech companies

And when the proportion of people of color in HR increased, the overall proportion of people of color across the company (in the leadership team, and everywhere else) also increased.

Comparing the % of people of color in HR, to the % of people of color at other levels of tech companies

In this case, similarity bias may actually help some companies. While we can’t imply causation, there is a pattern of “similarity bias” both in HR demographics and in the case of the founders’ genders. Depending on the goals of the company, this may actually help.

If the founders attended a ‘top’ university, you will see 37% more people of color in the leadership team

The alma mater of the founder/s, matters. Founders who had attended top universities (defined as top 25 according to U.S. News school ranking) had leadership teams (VP+) that were more diverse in terms of race.

However, this impact did not necessarily extend to other areas of the company—for instance, the engineering team. The alma mater of the founders only significantly impacted the racial diversity of the leadership team.

Why it is that attending a top school leads to more diversity in management teams, but not in engineering teams? Is it that founders from top universities are more concerned about the optics and doing this as a formality? Or is it that they believe that their management decisions will be more thorough this way, but don’t see the benefit of a racially diverse engineering team? Are top universities themselves simply more diverse environments than other universities? Data suggests this is true, but also the location of the university matters. The why is still TBD, but there’s clearly an opportunity to dive deeper.

Relationship between founder/s alma mater and the racial composition of their leadership teams

Transcripts aren’t required, but knowledge is power. If you are a person of color interested in moving up, their diploma might affect your chances.

If you’re female and an engineer, work for a Baby Boomer. But if you’re a person of color, work for a Millennial.

Age is more than just a number. We looked at the age of all founders today and found that those who are younger have built more racially-diverse Leadership and Engineering teams—more so than those who are middle-aged.

But founders who are Baby Boomers (at least in their mid-50s) have built more gender-diverse Engineering teams than their younger counterparts

What could be behind this? Is it that older founders are more likely to have had a daughter and give women a chance? Or maybe there are more women in their networks given that the rates of women who studied Computer Science were significantly higher in the 80-90s. Similarly, is it that younger founders find themselves in more racially diverse work environments to begin with (ie cities vs. suburbs, top universities on the coasts, etc)? All speculation of course, but still a pattern worth thinking twice about.

Racial and Gender composition of the Leadership or Engineering teams based on the Founders’ current age

The number of candles on your cake can be telling. Moving up in a given company might be more or less likely, depending on when the founders were born.

Early-stage companies are just as homogenous as the big ones — and they stay that way.

No, the company’s diversity problem won’t “get better when we’re bigger”. We were hoping for better news. Regardless of the stage of the company today, whether Seed, Series A, B, all the way to post-IPO, rates of racial diversity do not significantly vary.

And, despite the broader conversation in the mainstream media about the lack of diversity in large technology companies, companies that were founded more recently aren’t actually more diverse along gender and race measures than older ones.

Composition of tech companies based on current stage of financing

Overall gender and racial composition of tech companies based on their founding date

We can’t cut Founders slack at the start. This finding, that the proportions don’t significantly change if the company is older or larger, can be a hard one to digest. It points to the importance of setting diversity as a priority from the start if it is actually important to the strategy of the company.

Let’s come back to the philosophical question: Do you optimize for a person’s trajectory at any given company and perpetuate the status quo? Divertia is real. Companies (and humans) have a tendency to replicate themselves. But we’re also adaptive. If the force is great enough or if it’s applied early enough, there’s hope.

If you want to be the force that shifts these patterns of inequity, here’s what you can do:

If you’re looking for your next job

  • Look for companies with certain signals (founding team, HR, etc) as they, at least statistically, speak to your chances for leadership and mobility.
  • Alternatively, if a company seems to have unfavorable signals for someone like you, look for other signals that can make up that difference: For example, openly admitting that they are aware of issues, expressing a clear desire to evolve, and having a concrete plan with resources allocated to address it.

If you’re an employee

  • If you’re a woman in an all-male founded company, be vocal about ways to rise to leadership positions, and prepare for a possibly-harder battle than expected
  • If you’re a person of color, look at HR and see if you’re represented there, it could speak to your chances of rising up. If you’re not represented, consider bringing this up and use the insights presented here to spark a conversation at your workplace.
  • Consider the age and the school of the founders you’re working for, for clues as to your chances.
  • Alternatively, pressure your own employers to break all of these trends. You can vote with your feet, but you can also persuade with data. Show them these findings

If you’re in HR (or thinking about it)

  • Think critically about how you’re looking for diversity in candidates. Are you hiring yourself, over-and-over? Be mindful about how “similarity bias” impacts you, for better or for worse.
  • Check your own numbers. How many people of color are in your slate? How many women?
  • Or, if you are any kind of minority and want to create change in the faces of tech, joining the HR team is a way to do so.

If you’re a founder (or thinking about it)

  • You have a huge impact on how diverse different parts of the company will be in the future. Your founding team is the metadata—so pick your partners with this in mind. And if you end up with a more homogeneous founding team, be extra careful and make hiring a diverse set of “first employees” a priority
  • If building a diverse team is not a priority from the start, it won't be easier later (harder, in fact).
  • But the opposite is also true, so you get to choose.

If you’re an investor or a Limited Partner

  • Are you backing the right teams? Think about how you can use your influence to nudge founders to hire beyond themselves. Data shows more diverse teams build more financially-successful companies (run a Google search and stay-tuned for more from us).
  • If you are backing a team that is not very diverse at the start, use your influence to encourage the early team to diversify sooner vs. later (since “later” might be never).
  • If you’re vocal about how much diversity matters to you, founders might listen and assemble teams accordingly.

And that, perhaps, is the most actionable finding for the majority of technology companies that are still small and emerging: Build a diverse team from the start, and it will self-reinforce over time.

About this report & methodology.

We built this because we’re working to create the most intelligent, ai-powered, career manager. Our vision is to connect every engineer with her or his next ideal job, even (or especially) if the company is outside of a given engineer’s network. We match all kinds of engineers with companies that truly match their priorities and their chances of success. Understanding the work environment and the opportunities for growth a given engineer might face given her or his background is what initially inspired this analysis.

This analysis has been a collaborative effort between Roi Chobadi, founder of Stellares.ai, Danae Sterental, MBA at Stanford, and Dr. Matthias M. Herterich PhD, Visiting Scholar at Stanford.

Methodology

Through a combination of web crawlers, and manual data entry, we have collected information on employees of more than 13,000 US-based tech companies. We trained models to identify the gender and ethnicity of all employees. We kept the threshold for these algorithms high, so all employees who had less than 99% confidence in either gender or ethnicity were not tagged. Only employees who were tagged above 99%, and companies where the amount of tagged employees were statistically representing the population of the company (interval of 10% and confidence level of 99%) were included in the data set. We also built an Natural Language Processing (NLP) parser with an entity extraction algorithm, allowing us to parse the employee’s titles, and classify them to roles (e.g. business, engineering, recruiting, etc) and to seniority level (e.g. CXOs, VPs, directors, etc). Dependent variables that express diversity levels were built by calculating relative measures. We then tested 12 dependent variables resulting in 504 possibilities for correlations between independent and dependent variables. Out of this came our analysis. All insights are based on a 95% confidence interval (CI). For more, see our Full Methodology.

Data collection

Through a combination of web crawlers, and manual data entry, we have collected information on employees of more than 13,000 US-based tech companies.

  • Per each employee we’ve have collected 3 data points: their name, title, and image. We did not collect these data triplets for each and every employee in a company, but enough so that we had a statistically significant representative sample using the following parameters: confidence interval of 10% and confidence level of 99%.
  • We then tagged 500 of employees across companies with gender tags and ethnic tags, based on their names and images. Using these tagged employees, we trained models to identify the gender and ethnicity of all employees. We have kept the threshold for these algorithms high, so all employees who had less than 99% confidence in either gender or ethnicity were not tagged. Only employees who were tagged above 99%, and companies where the amount of tags employees were statistically significant (interval of 10% and confidence level of 99%) were included in the data set to represent that company’s employee population.

We have built an Natural Language Processing (NLP) parser with an entity extraction algorithm, allowing us to parse the employee’s titles, and classify them to roles (e.g. business, engineering, recruiting, etc) and to seniority level (e.g. CXOs, VPs, directors, etc).

To corroborate our results, 250 employees were selected at random, and we compared their detected gender, ethnicity, role and seniority level, to ensure accuracy. Indeed the models were correct in over 99% of cases.

We’ve augmented data about employees with data about the company and about its founding, and sometimes also leadership, team from several data sources, including Crunchbase.

Founder’s age was estimated based on graduation date of first undergrad degree. By scraping online lists of top school ranking, and matching school names of founders to the equivalent location on the list, founders were identified as having studied at a top school or not, where top school was selected as top 25 in the list.

Research

To analyze the data, we first did some data cleansing. Specifically, we removed inaccurate tuples and harmonized data from different sources. Tuples with insignificant or incomplete data as well as unreasonable data were removed. Dependent variables that express diversity levels were built by calculating relative measures.

  • An examples for such a relative measure would be the ratio between current female engineers and total current engineers within an organization.
  • Inspired by exploratory data analysis, we built a set of 42 classes of hypotheses and a multivariate correlation matrix, and thus developed an initial understanding of the data.
  • Hypotheses and simple correlations were tested by drawing on a set of 12 dependent variables resulting in 504 possibilities for correlations between independent and dependent variables.
  • Based on this rough understanding of the data, we developed a set of refined hypotheses that were tested leveraging diversity variables also focusing on historic churn data as well as multi-dimensional dependencies.

Table 1 is an example for such a class of hypotheses with 12 dependent variables that were calculated as relative measures.

>
Independent variableDependent variables
The average founders age has an impact on ... [Current Female Engineers]/[Current Total Engineers]
[Current Female Leadership]/[Current Total Leadership]
[Current Female Workers]/[Current Total Workers]
[Current POC Engineers]/[Current Overall Engineers]
[Current POC Extended Leadership]/[Current Total Extended Leadership]
[Current POC Leadership]/[Current Total Leadership]
[Current POC Workers]/[Current Total Workers]

All insights are based on a 95% confidence interval (CI). For data cleansing we used R and for the actual analysis we used the business intelligence & analytics software Tableau.

We'd like to thank all the following wonderful people who've helped with reviewing and offering feedback on our research:
Jeff Diana, previously Chief People Officer at Atlassian, at SuccessFactors, etc.
Efrat Dagan, Head of Talent Acqusition at Lyft
Paul Giron, Head of Technical Recruitment at Numerator
Luke Beseda, Head of Talent at Lightspeed Venture Partners
Alex Konrad, Associate Editor at Forbes
Alex Olshonsky, Director of Sales at VentureBeat
Menna Samaha, Director of Alternative Investments at Sponsors for Educational Opportunity
Richard Wellington, Engineering Manager at Netflix
Avesh Singh, Technical Lead at Cardiogram
Benjamin Vishny, Software Engineer at Fin
Darren Hau, Application Engineer at Tesla
Greg Brandt, Software Engineer at Airbnb
Elad Ossadon, Software Engineer at Lyft
Ilya Konstantinov, Software Engineer at Lyft
Shlomo Priymak, Engineering Manager at Facebook

Apply to STELLARES