EDI is not a diversity KPI: three types, three levels of response

Company E, with 40 staff, looks at its year-end headcount and finds that 9 of the 10 people promoted to management in three years were men. The board sets a target of "30% women in management" for next year. A target measures an outcome; it says nothing about the mechanism that produced the gap, so it does not say whether to fix hiring, promotion criteria or leave policy. This post separates three mechanisms, shows the evidence for each, then pairs them with three levels of intervention. The Company E example is invented.

How many kinds of discrimination are there, and how do they differ?

The site's framework splits them by mechanism into three Three types of discrimination:

This is a classification for choosing measures, not for looking up the law. UK law uses two other concepts: direct discrimination, where a person is treated less favourably than others because of a protected characteristic, and indirect discrimination, where a provision, criterion or practice applies to everyone but puts people sharing a characteristic at a particular disadvantage, unless the person applying it can show it to be a proportionate means of achieving a legitimate aim (UK Parliament, 2010). The framework's "structural" type is close to indirect discrimination, but I do not treat the two as the same.

What does the evidence for the implicit type measure, and what does it not?

Bertrand and Mullainathan responded to newspaper job ads in Boston and Chicago with fictitious CVs. Each CV was randomly given either a White-sounding name (Emily Walsh, Greg Baker) or an African-American-sounding name (Lakisha Washington, Jamal Jones). They answered more than 1,300 ads, sent nearly 4,900 CVs and counted callbacks. In the version published in the American Economic Review, CVs with White names received callbacks 9.65% of the time, those with African-American names 6.45%: a gap of about 50%, meaning a White-named applicant needed about 10 CVs for one callback and the other about 15 (Bertrand and Mullainathan, 2004). In the working paper, a better CV brought about 30% more callbacks for White-named applicants and about 9% for African-American-named ones. Employers who stated "Equal Opportunity Employer" in the ad, and federal contractors, did not discriminate less (Bertrand and Mullainathan, 2004).

Two things need separating. First, the experiment measures a gap in response at the interview-invitation step, not the employer's intent, and the authors themselves point out that they measure callbacks, not hiring or pay. They also conclude that existing economic models, whether taste-based or statistical, do not explain the full set of findings. Second, the site's framework files this example under "implicit", but as I read it, that label is an interpretation; the experiment cannot tell whether an employer acted deliberately or by reflex. The practical point: a gap can be real even when nobody sees themselves as biased, and knowing your own numbers is where to start.

An example of the structural type: neutral leave, unneutral outcomes

EIGE, the European Union's agency for gender equality, devotes a section of its Gender Equality Index 2019 report to parental leave. EIGE notes that there are no reliable comparable statistics on fathers' take-up across the EU, so the table below measures different things (takers, beneficiaries, share of leave, share of days); years and original sources also differ by country (European Institute for Gender Equality, 2019).

Sweden is interesting because part of the leave is reserved for fathers: 90 days that cannot be transferred; if the father does not use them, the family loses both the leave and the benefit. EIGE records that 44% of men whose children were born in 2013 had taken their reserved days by 2015. Portugal chose another route: a bonus month if both parents share leave; the number of fathers using a month of leave rose from 596 in 2008 to more than 24,000 in 2017 after the bonus began in 2009. EIGE adds a caution that information on how well such incentives actually work is scarce, and that there is no evidence that non-transferable leave alone is enough to raise fathers' use unless it is well paid (European Institute for Gender Equality, 2019).

Pay matters a great deal. The report (EU-28 in 2019, the UK included) records seven member states that offer unpaid parental leave, and families often depend on the higher salary of the father (European Institute for Gender Equality, 2019). This is a policy that does not discriminate in its wording and still shifts the burden to one side: the structural type.

Three levels of intervention: where does each stop?

The three levels of intervention in the site's framework run from light to deep:

On the middle level, the experiment by Chang and colleagues (2019) found that a short online course produced attitude change and some limited behaviour change, but the authors conclude that one-off trainings are not panaceas for bias at work (Chang et al., 2019). As for the reason used to sell EDI to a board, do not lean on the "diversity pays" argument: Green and Hand (2024) find no statistically significant relation between executives' racial diversity and profit margin, and note the reverse-causality problem in the original tests (Green and Hand, 2024).

A note on law for readers in the UK: section 159 of the Equality Act 2010 (Great Britain, not Northern Ireland) allows an employer to favour a person with a protected characteristic in recruitment or promotion if it reasonably thinks that group is disadvantaged or under-represented, provided that person is as qualified as the other, the employer has no general policy of favouring that group, and the action is a proportionate way to overcome disadvantage or raise participation (UK Parliament, 2010). As of November 2026 I have not checked whether section 159 has been amended, and I have not checked Vietnamese law, so this post offers no legal view for Vietnam.

Trying it on Company E

This table is an invented example, not data from any real business.

The 30% target can stay, but as an outcome measure, not as a measure to take. This works if you have data for each step (application, interview, hire, promotion) so you can see where the gap appears. It does not work if you only have totals by grade: then you know there is a gap but not at which step.

Going deeper: what did I read, and what did I not?

I read the NBER working paper 9873 (July 2003) by Bertrand and Mullainathan. The library source is the version in the American Economic Review (2004); the figures 9.65% and 6.45% come from public excerpts of that version, not from a full reading, and the details on CV quality come from the working paper and may differ from the published version. For the EIGE report (2019) I read the parental-leave section; EIGE draws on national studies for each figure, and I have not checked those original studies. EIGE gives years for only some figures (2016 for Denmark and Sweden, 2008 to 2017 for Portugal).

For Chang and colleagues (2019) and Green and Hand (2024) I have only the abstracts. The classification framework gives Young (2008) as the root of the structural type; I have only the abstract of that chapter. I read sections 13, 19 and 159 of the Equality Act 2010 in full. I have not read the sources by Kahneman (2011), Rawls (2001) and Bertrand, Chugh and Mullainathan (2005) that the classification framework gives as its roots, so I do not judge the theoretical basis of the "implicit" column. The Company E table is mine.