Diversity and profit: what McKinsey's 21% does and does not show

Your HR team sends the draft board paper, and one line reads: "companies with more gender-diverse executive teams are 21 per cent more likely to beat their industry on profit". You are the chief executive, or an owner raising money. If an investor asked where that comes from, would the line hold? The figure comes from a McKinsey & Company report (Hunt et al., 2018); see also the case on this report. I do not assess whether companies should pursue diversity, only whether the number supports the sentence attached to it. For a chart in a different report, see five questions to ask of a chart.

What do the 21% and 33% actually measure?

A gap between two rates, not a 21 per cent higher profit. McKinsey sorted companies into four groups by diversity of the executive team, then counted, in each group, how many had an EBIT margin above the median for their industry in their own country. For gender (991 companies) the top quartile scored 55% and the bottom 45%; for ethnic and cultural diversity (589 companies) it was 59% against 44% (Hunt et al., 2018, p. 8). Fifty-five over 45 is about 1.22; the report prints 21% because it computes from unrounded figures (Hunt et al., 2018, p. 8). Per 100 companies, the top group has about 10 more above the median on gender, or 15 more on ethnicity; the post on relative and absolute risk covers this conversion.

The definitions sit in the methodology. The executive team is set by each company, typically the CEO and two levels below (Hunt et al., 2018, pp. 8, 36). Profit is EBIT margin averaged over 2011 to 2015, with return on equity for financial companies (Hunt et al., 2018, p. 35). "Outperform" means above the median of the same industry and country (Hunt et al., 2018, p. 38). Some results are positive but not statistically significant: ethnic diversity against economic profit margin (Hunt et al., 2018, p. 13), and gender diversity on US and UK boards (Hunt et al., 2018, p. 10).

What does the report itself say it proves?

A correlation, and it says so several times. Page 2: "correlation does not demonstrate causation" (Hunt et al., 2018, p. 2). Page 5: "correlation is not causation. While not causal, we observe a real relationship" (Hunt et al., 2018, p. 5). Page 38: "we are not asserting a causal link", and showing one would be hard, "likely requiring detailed longitudinal studies" (Hunt et al., 2018, p. 38).

The report adds hypotheses about why (Hunt et al., 2018, p. 5) and a call to act: the relationship "does prompt action" (Hunt et al., 2018, p. 2). Both are the authors' views, not outputs of the calculation. A slide that quotes 21% and drops the sentence on page 2 or 38 says more than the report does.

Why can the same table be read in either direction?

Because it shows two things moving together, and the order of measurement puts profit before diversity.

Taris and colleagues (2021) write that two concepts may well correlate significantly without one causing the other, and that at the very least one must show the "cause" precedes its "outcome" in time (Taris et al., 2021, p. 1). A design in which each variable is measured once, even across two waves, is no better than a cross-sectional one at establishing causality (Taris et al., 2021, p. 2). Their editorial is about surveys of workers; applying it to company data is my inference.

In my reading, profit was measured first, so the data go further than "does not show the order"; Green and Hand (2024) read it the same way (Green and Hand, 2024, pp. 25–26). McKinsey concedes the possibility in theory but writes that in practice it "seems unlikely", because most companies undertake a major transformation only when they face a "burning platform" (Hunt et al., 2018, p. 39). In the pages the post uses, that is an observation with no figures.

A third explanation is a shared cause. The gender quartiles are cut across all 12 countries, while "outperform" is judged within each country's industry (Hunt et al., 2018, pp. 37–38); Australian, UK and US companies are 47 per cent of the sample but over 70 per cent of the top gender group (Hunt et al., 2018, p. 20). The report mentions multivariate regressions (Hunt et al., 2018, p. 38); the pages used here contain no list of control variables.

The two examples below are invented by me. They are not McKinsey's data or any company's.

Invented example 1: reverse causation. Eight companies, A to H. At year 0, A, B, C and D already have profit margins above the industry median. Assumption: A, B and C can afford new management posts and add women managers, D does not, and among the four less profitable companies only E does. At year 3, women are at least 40% of management in A, B, C and E.

The year-3 snapshot says "three times as likely", but A, B, C and D were profitable at year 0. The same snapshot would appear if diversity raised profit; only year-0 data tell the readings apart. The 2018 report gives one collection period for diversity data (Hunt et al., 2018, p. 35).

Invented example 2: a shared cause. Eighteen companies, split by size. Assumption: large companies have more management posts and higher margins because of scale, and within each size group diversity has nothing to do with profit.

Pooled, it is 56% against 44%, "25 per cent more likely", the same shape as McKinsey's figure. Split by size, the two groups are equal at both levels. That is the pooling-of-groups step in Where numbers go wrong. I do not know whether McKinsey's regressions controlled for size; the example shows how to ask, and is not an accusation.

Has anyone independent repeated the test?

One team tried part of it and could not reproduce it. Green and Hand (2024) applied McKinsey's approach to 497 S&P 500 companies, measuring profit over 2015 to 2019 and racial and ethnic diversity in mid-2020. In their top quartile 54.0% of firms had a positive industry-adjusted EBIT margin, against 51.2% in the bottom; the p-value is 0.65 (Green and Hand, 2024, pp. 8, 10). Across 40 test statistics on six financial measures, 37 are not significant, one is reliably positive and two reliably negative (Green and Hand, 2024, p. 9). McKinsey declined to share its data, even the names of the companies (Green and Hand, 2024, p. 6).

They state their limits: US only; the S&P 500 is not a random sample; one co-author judged race from photographs and names, which probably undercounts non-white executives (Green and Hand, 2024, pp. 13, 27). Their test also measures profit first and diversity later, so it is no test of cause, and it does not test the 21% figure on gender.

What should you ask before putting the figure in a paper?

Six questions for whoever put the slide together; grouping them into six is mine, drawn from the data-selection and interpretation steps of Where numbers go wrong.

A sentence for the paper that I think stands up: "McKinsey & Company, in a sample of 1,007 large companies in 12 countries (991 with gender data), found that 55% of those in the top quartile for women on the executive team beat their industry median EBIT margin, against 45% in the bottom quartile; the authors state this is a correlation and do not assert causation (Hunt et al., 2018, pp. 8, 35, 38)." It suits you if you want to cite a well-known study for exactly what it says; it does not if the paper must claim that raising diversity will raise profit.

The case for diversity can stand on other grounds. The report itself notes that social justice and legal compliance are typically the initial impetus for these efforts (Hunt et al., 2018, p. 1), and Green and Hand write that their results say nothing about the social or moral contributions of diversity (Green and Hand, 2024, pp. 28–29). In my reading, state those grounds on their own rather than asking a correlation to carry them.

Going deeper: which parts of the sources does the post use?

Hunt and colleagues (2018): the full-text PDF from mckinsey.org (page numbers are the printed ones). The post uses the executive summary, introduction, results (pages 8 to 19) and the methodology and limitations (pages 35 to 39); it does not use pages 21 to 32, and from page 20 it takes only the paragraph on Australia, the UK and the US. The ratio 55 over 45 is the post's own division.

Green and Hand (2024): the journal's PDF in full text, pages 5 to 29; not the appendix, the references, the 2021 working paper, or the Bermiss, Green and Hand (2024) study they mention.

Taris and colleagues (2021): the full editorial, publisher's version from the Utrecht University open repository. Dixon-Fyle and colleagues (2020): pages 10, 12, 14, 20, 28, 48, 49 and 51. Beraki and colleagues (2022): the abstract only. The McKinsey reports of 2015 and 2023 are not used.