A pay equity audit is a seven-step statistical review: define scope and privilege, clean your compensation data, group employees into comparable job categories, choose legitimate explanatory variables, run a multiple regression within each group, flag any group where the unexplained pay gap for a protected class exceeds roughly 2 standard deviations (p < 0.05), and then remediate the gaps and fix the policies that caused them. Done properly, the whole cycle takes 6 to 12 weeks for a mid-sized employer and should be repeated annually, ideally just before merit-increase planning.
Why the methodology matters more than the result
Pay equity analysis is governed by overlapping rules: the Equal Pay Act of 1963 (29 U.S.C. § 206(d)), which requires equal pay for “substantially equal” work in the same establishment; Title VII of the Civil Rights Act of 1964, which reaches broader compensation discrimination claims; and state laws such as California Labor Code § 432.3 and § 1197.5, Colorado's Equal Pay for Equal Work Act, and New York Labor Law § 194, several of which compare “substantially similar” rather than identical work. California employers with 100 or more employees must also file an annual pay data report with the Civil Rights Department by the second Wednesday in May.
Because these standards differ, two audits of the same payroll can produce different answers depending on how jobs were grouped. Your defensibility rests on documenting why you made each methodological choice before you saw the results.
Step 1: Set the scope and protect the work
Decide which entities, countries, and employee populations are in scope, and whether you are testing for sex, race/ethnicity, age (40+), or all three. Then engage counsel so the analysis is conducted under attorney-client privilege and the attorney work-product doctrine. Running an unprivileged audit means a plaintiff may obtain your draft regression output in discovery, including models you rejected. Privilege is not a shield against fixing what you find — it is protection for the exploratory stage.
Step 2: Build and clean the data set
Pull a point-in-time snapshot, typically the first pay period after year-end. At minimum, gather:
- Base pay annualized, plus a separate field for hourly rate and FTE percentage
- Bonus, commission, and equity grant value (test these separately — gaps often hide in variable pay)
- Job title, job code, job family, grade or level, FLSA status
- Location, business unit, and reporting manager
- Time in role, company tenure, prior relevant experience, highest degree
- Most recent performance rating
- Sex, race/ethnicity, date of birth
Expect 2 to 4 weeks here. The most common data defect is inconsistent job coding after an acquisition, where two populations doing identical work sit under different codes.
Step 3: Group jobs into comparable units
Grouping is the single most consequential decision in the audit. Group too narrowly and every cell has four people and no statistical power; group too broadly and you compare a staff accountant to a controller and generate false alarms.
| Grouping approach | How it works | Best for | Weakness |
|---|---|---|---|
| Job title | Each distinct title is its own group | Equal Pay Act “substantially equal work” claims | Tiny cells; title inflation distorts results |
| Job family + level | Combine related roles at the same grade | Most employers; balances power and comparability | Requires a clean, current job architecture |
| Similarly situated employee group (SSEG) | Analyst-built groupings based on skill, effort, responsibility, conditions | State “substantially similar work” standards | Judgment-heavy; must be documented |
| Full-population model with controls | One regression across everyone, with grade and function as variables | Organizations under ~300 employees | Weaker legal fit; masks localized gaps |
As a practical rule, a group needs at least 30 employees with a minimum of 5 in the comparison class before regression is meaningful. Smaller groups should be reviewed through a cohort or rank-order analysis instead, not discarded.
Step 4: Choose explanatory variables carefully
Legitimate controls are factors the Equal Pay Act recognizes as defenses: seniority, merit, quantity or quality of production, and other factors other than sex. Typical model variables are tenure, time in role, job level, location pay zone, performance rating, and full/part-time status.
Variables to exclude
Do not control for prior salary — more than 20 states and localities now prohibit relying on salary history, and including it imports past discrimination into your model as if it were legitimate. Be cautious with performance ratings and job level as well; if those are themselves biased, controlling for them will hide the very gap you are looking for. A useful technique is to run the model twice, with and without the suspect variable, and compare.
Step 5: Run the regression and set your threshold
Fit a multiple linear regression with the natural log of base pay as the dependent variable (log transformation makes the coefficient readable as an approximate percentage difference). The coefficient on the protected-class indicator is your unexplained pay gap.
| Threshold | Statistical meaning | Typical use |
|---|---|---|
| 2 standard deviations (p < 0.05) | <5% chance the gap is random | The benchmark courts and the OFCCP generally treat as meaningful |
| 3 standard deviations (p < 0.01) | <1% chance the gap is random | Conservative trigger for large populations |
| 1.645 SD (p < 0.10, one-tailed) | Early-warning signal | Internal monitoring, not litigation standard |
| Practical materiality (e.g., gap > 1–2%) | Business judgment overlay | Screening out trivial but significant gaps in huge groups |
Also check your model's R-squared. If legitimate factors explain less than about 0.70 of pay variance, your model is probably missing a variable rather than revealing discrimination. Follow up with a cohort review of the individual outliers flagged, because regression tells you a group-level story, not a person-level one.
Step 6: Remediate the gaps
Fund adjustments for flagged employees and make them effective on a single date, separate from the merit cycle, so they are not mistaken for performance rewards. Budget planning rule of thumb: full remediation of a 2%–3% unexplained gap typically costs 0.1% to 0.3% of total base payroll.
- Never lower anyone's pay to close a gap — 29 U.S.C. § 206(d)(1) expressly forbids reducing wages to achieve compliance.
- Adjust to the predicted value from the model, not to the group mean.
- Document the business rationale for each adjustment in the employee's compensation record.
- Keep communications simple and factual; avoid characterizing adjustments as remedying discrimination.
Step 7: Fix the system, then re-run
Adjustments without process change produce the same gap next year. Look upstream at starting-offer approval rules, bonus discretion, promotion velocity, and manager-level merit allocation. Narrow the discretion range managers hold without justification, require a written rationale for offers above or below midpoint, and publish salary ranges where state law requires it.
Train the people who make pay decisions. A solid Diversity, Equity & Inclusion (DEI) course helps managers recognize how bias enters starting offers, performance ratings, and promotion timing — the three upstream drivers that regression consistently surfaces. Pair it with Sexual Harassment Prevention training where state mandates apply, and with Conflict Resolution & De-escalation for the managers who will field the inevitable questions about pay transparency. Be clear with your team, though: an online course is awareness training, not a substitute for a statistically valid audit conducted by qualified analysts and counsel.
Frequently asked questions
How often should we run a pay equity audit?
Annually, timed 60 to 90 days before merit planning so results can inform budget. Run an interim analysis after any acquisition, large reduction in force, or compensation structure redesign.
What size workforce do we need for regression analysis?
Roughly 30 employees per analysis group with at least 5 in the protected class. Below that, use cohort comparisons, rank-order tests, or Fisher's exact test rather than forcing a regression that cannot produce reliable coefficients.
Does a pay gap automatically mean we broke the law?
No. A raw gap often reflects job composition. A statistically significant unexplained gap after controlling for legitimate factors is a risk indicator that warrants individual review, not an automatic finding of liability.
Should we tell employees about the audit?
Share the commitment and methodology at a high level; avoid releasing raw results, which are usually privileged and easily misread. If you operate in a pay transparency jurisdiction, align messaging with your posted range disclosures.
Next steps
Pay equity is sustained by the decisions managers make every day, not by the spreadsheet. Enroll your leaders in DEI and compliance training through FastCredentials and get certificates issued the same day they finish.