Pay equity regression or cohort review?
A regression on a small team gives a precise-looking answer to a question it cannot really test. Say so and compare people instead.
Short answer: A pay equity regression needs enough people in each comparison group to mean anything. Many practitioners treat about 30 people per group, with several of each gender, as the floor. Below that, the result swings with one hire or one leaver and rarely flags a real gap. For small groups, a structured cohort review of each person against like-for-like peers is the sounder test.
A clean result nobody quite trusted
Picture a 120-person architecture and engineering firm in Salt Lake City, with a small studio of 15 people in Denver. Its board asked for a pay equity analysis before the firm expanded into two new states. The head of people exported the payroll, added job level, discipline, years of experience, license status, office and last year's rating. Then they ran a regression across all 120 people.
The output looked reassuring. Women earned about 2% less than men once the other factors were taken into account. The model said the difference was not statistically significant. The summary for the board said there was no evidence of a gender pay gap.
Then the head of people looked at the structural engineers. There were 14 of them across three levels. Four were women. At the senior level, the two women were paid 9% and 11% below the three men, with similar experience and the same license. A regression across 120 people had no way to see that. It had averaged those two women in with everyone else and reported the average.
A month later one of the senior men left. The head of people reran the model to update the board and the company-wide gap moved by almost a full percentage point. Nothing about anyone's pay had changed. The board asked which number was right.
This is a composite of a common situation at companies this size, not a client. The details are illustrative. The pattern is common because a regression is the method most people have heard of. It gives a single, precise-looking number even when the data underneath it cannot carry the weight.
Why choosing a pay equity method is hard for a small company
Two definitions first. A regression is a statistical model that estimates how much of the difference in pay is linked to each factor, such as level or experience. It then tests whether gender, race or another protected characteristic still makes a difference once those factors are accounted for. A cohort review compares each person directly with the people doing the same or equal work and asks whether any difference in pay has a legitimate, documented reason.
Two good goals pull against each other. The first is rigor. A regression looks objective. It handles many factors at once, it produces a test of statistical significance and it is the method large employers and expert witnesses use. A company that runs one feels it has done the serious version.
The second goal is fit. The law asks whether people doing equal work, or work of equal value, are paid equally. If they are not, it asks why. That is a comparison between people. At a company of 50 to 500 people, the groups doing equal work are usually small: Five product designers, eight account managers, three finance analysts. A model needs many observations to estimate each factor it includes. When a group has a dozen people and five factors, the model is mostly estimating noise.
The US Equal Employment Opportunity Commission's guidance on pay discrimination says it plainly: Statistical analyses are less reliable when they cover a small number of people [1]. California's state guidance on measuring pay gaps notes that when a group is small, a regression among people doing substantially similar work cannot be run at all to predict each person's pay [4]. Many practitioners use a floor of about 30 people in a group, with a handful of each gender, before they rely on a regression for that group. California's guidance calls this the 30/5 rule [4].
Most companies this size have very few groups that clear that floor. They still want a review they can defend. Increasingly they need one. In a survey of 3,413 organizations run in late 2025, 60% said pay equity analysis was a current or planned initiative [5]. The question is not whether to review pay. It is which test fits the size of each group. It is also what to say about the limits of whichever test is used.
Five ways a small company picks the wrong pay equity method
Each of these starts from a reasonable wish: To be rigorous, to be thorough, to be efficient or to be objective. The damage comes from letting the method decide what the review can see.
"Run one regression on the whole company"
What it looks like: Too few people in any single role, so everyone goes into one model. Engineers, sales staff, designers and finance are pooled, with job family and level added as factors so the model can "adjust" for them.
Why smart teams fall into it: Pooling gets the numbers up to a size where the regression runs and the significance test means something. It also gives one headline figure that is easy to report.
What it costs: A company-wide model answers a different question from the one the law asks. It estimates an average gap across the whole company after adjustments. An equal pay claim compares one person with the people doing equal work beside her [1][2]. A pooled model can show a small, non-significant average while one team has a large gap, as the engineers in the opening did. It also assumes that each factor, such as a year of experience, is worth the same in every job, which is rarely true. A pooled regression can be a useful first screen for where to look. It should not be the finding.
"Add every factor we have"
What it looks like: The model includes level, function, tenure, time in role, education, location, rating, license status and prior employer, because each one might explain something.
Why smart teams fall into it: Leaving out a real factor feels like an error. More factors look more thorough. Each one shrinks the unexplained gap a little more.
What it costs: Every factor uses up some of the information in the data. In a group of 40, eight or nine factors leave the model very little left to estimate the one it is meant to test. The result becomes unstable: One hire, one leaver or one corrected record moves the gender estimate, as the departure in the opening did. Some factors also carry the gap inside them. Level and ratings are set by managers. If they are set unevenly, controlling for them hides the problem. Our guide on which pay gaps you can defend covers why a factor must be tested before it is used to explain a gap.
"The group is too small to test, so leave it out"
What it looks like: The analysis runs on the groups large enough for a regression. Everyone else is reported as "not tested" or simply does not appear. In a company of 150, that can be most people.
Why smart teams fall into it: Following a statistical threshold feels disciplined. Reporting only results that meet it seems more honest than reporting weak ones.
What it costs: The people left out are often the most exposed: The only woman at a senior level, a small team of specialists, a role with one person in each office. The law does not exempt them because they are few. An equal pay claim can rest on a single comparator. EEOC guidance tells investigators to look at all similarly situated employees, not only the ones that make a tidy sample [1]. In the EU, where a real comparator cannot be found, a worker may rely on a hypothetical comparator or on statistics [3]. A review that tests 30% of the workforce and reports "no significant gap" has said nothing about the other 70%. A sound review covers everyone, using the test that fits each group.
"A cohort review is just eyeballing the numbers"
What it looks like: The company accepts that regression does not fit, so someone scans a spreadsheet of salaries by title, notes a few outliers and calls the work done. No comparison groups are written down and no reasons are recorded.
Why smart teams fall into it: Without a model, the review feels informal by nature. A quick look by someone who knows the people seems good enough for a small company.
What it costs: An undocumented look at the numbers is the weakest position there is. It shows the company looked, without showing what it compared or why it accepted each difference. A cohort review has its own discipline. The comparison groups are defined on objective criteria before anyone looks at pay. Every difference within a group gets a reason that would hold up, or is flagged. The EU directive requires pay structures that let a company assess whether workers are in a comparable situation on objective, gender-neutral criteria such as skills, effort, responsibility and working conditions [3]. Our guide on which job evaluation method to use covers how those groups are usually set. Done properly, a cohort review is harder work than a regression, not easier.
"Let the tool choose the method"
What it looks like: The company buys or uses a pay equity tool, loads the data and accepts the output. The tool runs regressions where groups are large enough and quietly sets the rest aside, or pools groups until they are.
Why smart teams fall into it: A tool feels neutral and repeatable. It removes the need to make method choices that nobody in-house feels qualified to make.
What it costs: The method choices are still being made. They are made by default settings that were built for larger workforces. The report looks complete and the people the tool could not test are easy to miss in an appendix. Tools are useful for organizing data and for running models on the groups that are big enough. Someone still has to decide how groups are formed, which factors count as legitimate and how the small groups are reviewed. Those are judgment calls the company will have to explain, whatever produced the numbers.
What a sound pay equity method looks like at 50 to 500 people
A good method is not the most sophisticated one. It is the one whose results the company could explain to a regulator, a court or an employee who asks why she is paid less than a colleague. Several qualities set it apart.
The method follows the size of each group. Large groups can be tested with a regression. Small groups get a structured cohort review. Most companies this size end up using both. The report says which test was used where.
Groups are set before pay is looked at. Who is compared with whom is decided on the work: Its skills, effort, responsibility and conditions. It is not adjusted after the results come in.
Everyone is reviewed. No one drops out because their group is small. The people in the smallest groups often get the closest look.
Every difference has a reason or a flag. Within each group, a pay difference is either explained by a documented, legitimate factor or marked for action. "Not statistically significant" is not a reason.
The limits are stated. Where a group is too small for statistics to mean much, the report says so. It does not present a small-sample result as proof of fairness.
The result is stable. One leaver or one corrected record does not change the conclusion. If it does, the method was asking more of the data than it could give.
Counsel is involved from the start. Whether to run the review under legal privilege, how to record the results and what to do about any gaps found are legal questions as well as pay questions. We would expect employment counsel to be involved before the analysis starts, not after the results are in.
The calls only you can make
A method can be chosen on technical grounds. The decisions inside it belong to the company.
Who counts as doing equal work? Narrow groups by title are easy to defend as like-for-like but leave many people with no comparator. Broader groups based on the value of the work include more people but need a sound way to compare different jobs. Our guide on what triggers a joint pay assessment shows why the EU's categories of workers make this choice harder to avoid.
Which factors may explain pay? Experience, location and performance are common. Prior pay and negotiation are increasingly hard to defend. Each factor the company accepts is a statement about what it believes pay should reward.
How much evidence before acting? A large gap for one person in a group of four will never be statistically significant. The company has to decide whether it acts on a clear individual gap without a statistical test behind it. We think it should.
Which characteristics to review? Most reviews start with gender. Race and ethnicity are often reviewed in the United States where the data allows. Groups split by more than one characteristic get smaller, which pushes even more of the review toward cohort comparison.
What to report and to whom? A board summary, an employee report and an EU pay report each need a different level of detail. In small groups, publishing results can reveal individual pay, which is a privacy question as well as a legal one.
The numbers and the law
- United States, small numbers. EEOC guidance on compensation discrimination states that statistical analyses are less reliable when they cover a small number of people. It tells investigators to compare the pay of all similarly situated employees, defined as those who would be expected to receive the same pay. It notes that a significance test cannot by itself say what caused a pattern [1].
- United States, equal work. The federal Equal Pay Act compares the pay of men and women in the same establishment doing equal work requiring equal skill, effort and responsibility under similar working conditions, with exceptions for seniority, merit, quantity or quality of production and a factor other than sex [2].
- European Union, categories of workers. Pay structures must allow an assessment of whether workers are in a comparable situation, using objective, gender-neutral criteria that include skills, effort, responsibility and working conditions. Pay reports must show the gender pay gap by category of workers. A joint pay assessment is required where a category shows a gap of at least 5% that is neither justified on objective, gender-neutral criteria nor remedied within six months. Where a real comparator does not exist, a hypothetical comparator or statistics may be used. Member states may restrict access to reported information that would reveal an identifiable worker's pay [3].
- California, small samples. The state's guidance for employers on measuring pay gaps treats a regression among people doing substantially similar work as possible only when groups meet minimum sizes, which it calls the 30/5 rule. For smaller groups it describes a more cautious approach and states that its methods are not authoritative or definitive [4]. The pages carry no publication date.
- How common reviews are. A late-2025 survey of 3,413 employers found 60% running or planning pay equity analysis, 3% more than a year before [5]. These are the publisher's own figures.
- Which methods employers use. A pay equity software provider's survey of 267 organizations in early 2025 reports that many organizations rely on basic statistics, that fewer use multiple regression and that higher-revenue organizations are more likely to use it [6]. The public summary gives no percentages. This is the provider's own finding.
This is general information, not legal advice. Equal pay rules differ by state and country and change over time. Have counsel advise on how a pay equity review is set up, run and documented.
Our view
We think most companies of 50 to 500 people should not lead with a regression. Where a comparison group has fewer than about 30 people, a regression gives a precise-looking number that one hire can move and that rarely detects a real gap. Presenting that number as evidence of fair pay is the worst outcome: It reassures the board and protects no one.
For small groups, a structured cohort review is the sounder test. It asks the question the law asks, covers every person and leaves a record of why each difference exists. Where a company has groups large enough for a regression, we would use one there and say clearly where it was not used.
And we would state the limits in writing. A review that says "this group is too small for statistics, so each person was compared directly" is stronger than one that hides the small groups behind an average. Our note on raw and adjusted pay gaps covers why the headline number is never the whole answer.
Where we help
Our pay equity review uses the test that fits each group in your own pay data: A regression where groups are large enough and a structured cohort review where they are not. Every person is reviewed, every difference gets a reason or a flag and the limits of each test are written down. Your counsel decides how it is run and recorded. See pay equity for how a review is scoped and priced. If a conversation would help first, Book a free call.
Sources
- US Equal Employment Opportunity Commission, Compliance Manual Section 10: Compensation discrimination, sections 10-III.A.1 to 10-III.A.3, checked 2026-10-28
- 29 U.S.C. section 206(d)(1), Equal Pay Act (Legal Information Institute), checked 2026-10-28
- Directive (EU) 2023/970 on pay transparency, recital 28 and Articles 4(4), 9(1)(g), 10(1) and 12(3) (EUR-Lex), checked 2026-10-28
- California Commission on the Status of Women and Girls, Pay Equity Task Force, Measuring Pay Equity: Small Sample Size Individual Analysis (with the companion large sample page; undated), checked 2026-10-28
- Payscale, 2026 Compensation Best Practices Report (3,413 responses, October to December 2025), checked 2026-10-28
- Trusaic, 2025 Pay Equity Policies and Practices Report (survey of 267 organizations, first quarter 2025; public summary page), checked 2026-10-28