Guide · Pay equity

Which pay gaps can you defend?

An explained gap is only as sound as the factors that explain it.

Short answer: An explained pay gap is one that factors such as job level, tenure or location account for statistically. That is not the same as a gap you can defend. Under US and EU equal pay law, each factor must be genuine, applied consistently and free of sex bias itself. Level, ratings and starting pay often are not, so test them before you rely on them.

A gap that almost disappears

Northwind Analytics, a 160-person software company, runs its first pay equity review the year it starts posting ranges. The raw numbers are uncomfortable. Women's average base pay is 14% below men's.

Then the analyst adds controls: Job level, function, tenure, location and last year's performance rating. The gap shrinks to 1.8% and the model says it is not statistically significant. The CEO reads the summary and relaxes. "So it's explained. We're fine."

The head of people is less sure. She knows three things the summary does not say. Most of the 14% sits in job level: Women at Northwind are concentrated at the two lower levels of most job families. Level and performance rating are both decided by managers. And until 2023 the company set starting pay by adding a fixed percentage to each new hire's last salary.

So the question is not whether the gap is explained. It is what explained it. And would those explanations hold up if an employee, a regulator or a court asked?

Northwind Analytics is a fictional company and the figures are illustrative.

Why "explained" is harder than it sounds

Every pay equity review produces at least two numbers. The raw gap compares the average or median pay of women and men across the company. The adjusted gap compares people after accounting for factors that legitimately affect pay, such as the job, the level and the location. The part of the raw gap the factors account for is called "explained". What is left is "unexplained".

The difference between the two numbers can be large. A compensation data provider's 2026 report, based on more than 130,000 US survey responses, puts women's median pay at 82 cents for every dollar men earn. With job title, experience, education and other factors held constant, the figure rises to 99 cents [6] [7]. Across the EU, the raw gap in average hourly pay was 11.1% in 2024 [8].

Both numbers are true. They answer different questions. The raw gap asks whether women and men end up with the same pay at this company. The adjusted gap asks whether people in comparable situations are paid the same. A company can have a small adjusted gap and a large raw one. That usually means women and men are doing different jobs or sitting at different levels.

The difficulty is that "explained" is a statistical word, not a legal one. A factor explains a gap when it is associated with it in the data. Whether it justifies the gap is a separate question with a much higher bar. The two goals pull against each other:

  • Precision. Adding the right controls isolates like-for-like comparisons. Without them, every difference in job mix looks like discrimination.
  • Integrity. Every control assumes the factor itself is fair. If a factor was shaped by bias, controlling for it removes the bias from the number without removing it from the company.

A sound review holds both. It uses controls to find like-for-like comparisons and then asks whether each control deserves its place.

Size adds a third difficulty. A company of 160 people has too few people in most comparisons for a statistical model to say much with confidence. Large employers can run regressions across thousands of people and trust the averages. A mid-size company usually learns more from comparing named groups of people in the same job and level, then reading the outliers one by one. The model is still useful as a summary. It is a weak basis for concluding that nobody is underpaid.

And the choice of controls is rarely neutral. Whoever builds the model decides which factors go in. Each factor added tends to shrink the unexplained gap, so there is quiet pressure to keep adding them. A model with a dozen controls can explain almost any gap away. The question to ask of every model is not how small its gap is but which factors are doing the explaining.

Five ways a review explains away a real gap

Each of these is common in a first review. Each produces a result that looks clean and may not be.

"Explained means justified"

What it looks like: The model accounts for most of the gap, so the company reports that its gap is explained and treats the matter as closed.

Why smart teams fall into it: The word does the work. "Explained" sounds like a verdict. And statistical models carry an authority that a manager's reasons do not.

What it costs: The law asks more than a model can answer. Under the US Equal Pay Act, a pay difference in substantially equal jobs is defensible if it rests on a seniority system, a merit system, a system that measures production or a factor other than sex [1]. The federal guidance is clear that the employer must show that a gender-neutral factor, applied consistently, in fact explains the difference [1]. California goes further. Each factor must be job-related, consistent with a business necessity, applied reasonably and, alone or with other factors, account for the entire difference [2]. In the EU, pay differences must rest on objective, gender-neutral criteria [4]. Each of these asks about the reason behind a specific person's pay. A regression coefficient is not a reason.

"Controlling for level settles it"

What it looks like: Job level is the first and strongest control in the model. Once it is in, most of the gap disappears.

Why smart teams fall into it: Level is the most legitimate pay factor there is. People at higher levels do bigger jobs and should earn more. Leaving it out would compare a junior analyst with a director.

What it costs: Level is a decision, not a fact of nature. It is set when someone is hired and changed when they are promoted. If women are placed lower at hire or promoted more slowly, a model that controls for level hides exactly that. The gap moves from "pay within a level" to "who gets which level", where the model no longer looks. At Northwind, most of the 14% disappeared into the level control. That is the part a reviewer should read most closely, not least. Promotion speed is often where it shows. If men at Northwind reach the next level in two and a half years on average and women in three and a half, the level control turns that year into an explained gap. The EU directive points the same way. A joint pay assessment has to analyze the share of women and men in each category, not only their pay [4].

"Performance ratings are objective"

What it looks like: The model includes last year's performance rating. Women's ratings are slightly lower on average, so part of the gap is attributed to performance.

Why smart teams fall into it: Pay for performance is a stated policy and ratings are recorded. Using them seems fair and consistent with how the company says it pays.

What it costs: Ratings are judgments made by managers, often in a calibration meeting, often on criteria that are hard to write down. If ratings themselves differ by gender for reasons that are not about the work, a model that controls for them launders the difference. US guidance treats a merit system as a defense only to the extent it really accounts for the difference [1]. It also has to be a genuine system, not a label on discretion. A sound review tests the factor before it uses it. Do ratings differ by gender within the same level and team? If so, the rating is a finding, not a control.

"Starting pay was market rate"

What it looks like: A gap between two people in the same job traces back to their starting offers. The company explains that each was hired at the market rate of the time, often based on what they earned before.

Why smart teams fall into it: Matching or beating a candidate's current pay is a common way to close a hire. It feels neutral because the company did not set the earlier salary.

What it costs: Prior pay carries earlier gaps forward. Federal guidance says prior salary cannot by itself justify a pay difference, because it can reflect sex-based discrimination elsewhere [1]. California does not allow prior salary to justify a sex-based difference at all [2]. In the Ninth Circuit, which covers California and eight other western states, prior pay is not a "factor other than sex" under the Equal Pay Act [3]. The EU directive goes further still. Employers may not ask applicants about their pay history [4]. A gap that began with a starting offer does not shrink on its own. Merit increases are usually a percentage of pay, so the gap tends to grow in dollars every year it is left alone.

"It isn't statistically significant, so there's no problem"

What it looks like: The adjusted gap is small and the model says it could be due to chance. The review concludes there is nothing to fix.

Why smart teams fall into it: Significance is the standard test in statistics. Acting on a result that could be noise feels unscientific.

What it costs: At 160 people, many comparison groups hold a handful of people. Significance tests in groups that small rarely flag anything, even when a real problem exists. And the law does not work on averages. Equal pay claims compare one person with specific comparators doing equal work [1] [2]. A company-wide adjusted gap of 1.8% can sit alongside one woman paid 12% less than the man doing the same job beside her. If she asks, the company needs a reason for her pay, not a p-value. In the EU, an employer that has not met its transparency duties carries the burden of proving there was no discrimination [4]. Our guide on what triggers a joint pay assessment looks at how category gaps are tested there.

What a defensible review looks like

A review cannot make a company fair. It can show clearly where it is and is not. In our view a defensible review has these qualities.

  • Both numbers are reported and explained. The raw gap and the adjusted gap appear side by side, with a plain account of what moves one to the other. Neither is used to hide the other.
  • Every control earns its place. Each factor in the model has a stated reason to be there. Factors set by people, such as level, rating and starting pay, are tested for differences by gender before they are used as controls.
  • Placement is read as closely as pay. The review shows who sits at each level and how fast people move between levels, by gender. That is where explained gaps usually come from.
  • Individuals are visible, not only averages. The review lists the people whose pay sits furthest from their comparators, whatever the model's overall result. Each has a documented reason or a fix.
  • The reasons are on record. Pay decisions carry a short neutral reason written when they were made, so the review can rely on them rather than reconstruct them.
  • Prior pay is treated as a warning, not a reason. Where starting pay was anchored on earlier salaries, the review checks those people first.
  • The review is repeated before each pay cycle. Run before raises are decided, it can shape them. Run after, it can only report what they did.
  • Fixes are costed and dated. Unexplained gaps above a level the company set in advance get a remedy with a budget and a date, not a note for next year.

Law firm guidance often recommends running the review under legal privilege, with counsel, before anything is disclosed [5]. That is a decision for your lawyers. It does not change what a good review contains.

The calls only you can make

A review surfaces facts. Several of the decisions it forces are about values.

Which factors you accept as legitimate. Location, tenure and market premiums for scarce skills all raise pay for reasons the company may defend. Each also tends to favor some groups over others. Which ones you are willing to stand behind, in front of employees, is a choice.

Whether to fix gaps the model calls explained. If women are placed at lower levels, the adjusted gap may be small and the problem large. Fixing placement means re-leveling people or changing promotion practice, which costs more than a pay adjustment. Leaving it means the raw gap stays.

How far beyond gender to look. US companies often review pay by race and ethnicity where their data allows. EU law focuses on sex. The data you hold, the laws that apply and what you want to know all shape the scope.

How much to share. Some companies publish their adjusted gap and their actions. Others keep results internal. EU employers will publish some figures in any case [4]. The choice is about trust as much as law.

How fast to close what you find. Closing every gap in one cycle is clean and expensive. Spreading fixes over two or three cycles is cheaper, but each year a known gap stays open is a year the company chose to leave it.

The law and the numbers

  • US federal. The Equal Pay Act allows pay differences in substantially equal jobs based on seniority, merit, production or a factor other than sex. The employer must show the factor in fact explains the difference [1].
  • California. Each factor must be job-related, consistent with a business necessity, applied reasonably and account for the entire difference. Prior salary cannot justify a sex-based difference [2].
  • Ninth Circuit. Since the court's en banc decision of February 2020, prior pay is not a factor other than sex under the federal Equal Pay Act in its states.
  • EU. Pay structures must allow comparison on objective, gender-neutral criteria: Skills, effort, responsibility and working conditions. A neutral-looking criterion that disadvantages one sex is indirect discrimination unless it serves a legitimate aim by appropriate and necessary means. Employers may not ask about pay history and carry the burden of proof if they have not met their transparency duties [4].
  • The numbers. A compensation data provider reports a US raw gap of 82 cents on the dollar and a controlled gap of 99 cents in 2026 [6] [7]. The EU's raw gap in average hourly pay was 11.1% in 2024 [8].

Our view

We think "explained" should be the start of a review, not its verdict. Most of a company's gap is usually explained by something: Level, function, tenure, starting pay. The useful work is to ask whether each of those explanations would survive a direct question from the person it affects.

We would argue hardest against two habits. The first is controlling for factors that managers set, such as level and rating, without first testing them. The second is reading a non-significant average as proof that no one is underpaid. Both produce clean summaries that do not match what individual employees see.

A small adjusted gap is good news. A small adjusted gap built on controls that were never examined is a result no one should rely on.

Sources

  1. US Equal Employment Opportunity Commission, Compliance Manual Section 10: Compensation discrimination, sections 10-IV.F.1, 10-IV.F.2 and 10-IV.F.2.g, checked 2026-10-25
  2. California Labor Commissioner, Equal Pay Act FAQ (questions 1, 15 and 71), checked 2026-10-25
  3. Rizo v. Yovino, US Court of Appeals for the Ninth Circuit, en banc opinion (27 February 2020), checked 2026-10-25
  4. Directive (EU) 2023/970 on pay transparency, Articles 3(1)(j), 4(4), 5(2), 9, 10 and 18 (EUR-Lex), checked 2026-10-25
  5. Mayer Brown, EU Pay Transparency Directive: Practical briefing for international employers (30 June 2026), checked 2026-10-25
  6. Payscale, 2026 Gender Pay Gap Report press release (24 March 2026; more than 130,000 US respondents, January 2024 to January 2026), checked 2026-10-25
  7. Payscale, The 2026 gender pay gap report (26 March 2026; controlled gap method), checked 2026-10-25
  8. Eurostat, Gender pay gap statistics (data extracted February 2026; 2024 figures provisional), checked 2026-10-25

Raw or adjusted pay gap: Which is true?

Both are true. What the raw and the like-for-like gap each answer, which one the law asks for and the different fix each points to.

When an employee asks about pay

Employees rarely ask only for a number. Five ways the first pay conversation goes wrong and what a good first answer does.

EU pay gap reporting: When and what?

The first EU pay gap reports cover 2026 pay. Who files when, what the seven figures count and why the data year is already almost over.

See Resources for every guide and the self-check.