Compensating Your Employees Fairly

1. 1Treat every statistically significant disparity with no obvious explanation as a question, not a conclusion: run the follow-up investigation before contemplating any remediation. Investigation always precedes remediation.
2. 2Identify legitimate nondiscriminatory factors that were omitted from the statistical model but still explain variation in pay — seniority, merit, quantity/quality of production, or a bona fide factor other than sex such as education, training, or experience. Add these to the analysis and re-test whether the disparity persists.
3. 3Examine the specific pattern behind the flagged group, not just the aggregate coefficient. Compare the actual increases granted person-by-person within the similarly situated grouping to see whether the gap is consistent or explained by a legitimate system.
4. 4Gather corroborating evidence beyond the numbers — decision-maker statements, documented pay-setting practices, application of stated systems — to distinguish a legitimate factor from discriminatory pay setting (as when a manager's expressed views about women confirm the statistical pattern).
5. 5Confirm the compensation metric matches the question: use base rate of pay when investigating discriminatory pay-setting practices, total compensation when overall equity is the concern.
6. 6Engage legal counsel before making any adjustment. Conduct remediation only under the auspices of counsel because the implications of changing compensation are significant and counsel is best positioned to advise on appropriate remedies.
7. 7When remediating, raise the underpaid employees to the equitable level — never reduce anyone's pay to close a gap. The Equal Pay Act expressly forbids reducing the wage rate of any employee to achieve compliance.
8. 8Document the legitimate factors identified, the evidence reviewed, and the rationale for any adjustment so the correction itself withstands scrutiny.

A warehouse group shows a statistically significant gender disparity: male employees received a $0.25/hour increase while female employees received only $0.15/hour, leaving an unexplained $0.10/hour difference. Rather than adjusting immediately, the investigator reviews the increases granted to the remaining people in the similarly situated grouping and finds the same pattern repeats. The investigator then gathers non-statistical evidence and documents the views expressed by the manager, Shawn Kendrick, regarding women in the warehouse. Combined, the consistent pattern and the manager's stated views indicate the likely cause is discriminatory pay-setting practices rather than any legitimate factor. Because no seniority, merit, production, or other bona fide factor explains the $0.10 gap, remediation is warranted — the underpaid female employees are raised to the equitable rate under the direction of legal counsel, and no one's pay is reduced.

• If a disparity is statistically significant but a legitimate factor omitted from the model (seniority, merit, production, education/training/experience) explains it once added, then no remediation is warranted — document the explanation.
• If the pattern is consistent across the similarly situated grouping and corroborating evidence (decision-maker statements, pay-setting practices) points to discrimination, then remediate under legal counsel.
• If you are investigating discriminatory pay-setting, then analyze base rate of pay; if you are assessing overall equity, then use total compensation.
• If remediation is needed, then raise the underpaid employees — never lower anyone's wage, because the statute forbids reducing pay to comply.
• If a bona fide factor is claimed as the explanation, then verify it is job-related and furthers a legitimate business purpose, not merely convenient, before accepting it.

• Adjusting pay before investigating. The tell: a coefficient triggers a raise with no documented review of omitted legitimate factors — leaving you unable to defend either the finding or the fix.
• Reducing someone's pay to close a gap. The tell: 'equalized' rates where a previously higher-paid employee's wage dropped — a direct statutory violation.
• Accepting the aggregate result without inspecting the person-by-person pattern. The tell: a group-level disparity treated as uniform when individual increases actually reveal (or dispel) the systematic pattern.
• Ignoring non-statistical evidence. The tell: a clean statistical explanation adopted while documented decision-maker views or pay-setting behavior contradicting it go unexamined.
• Remediating without legal counsel. The tell: HR institutes adjustments unilaterally, missing the significant legal implications counsel would flag.

• ▸Flagged grouping & metric — name the similarly situated group and whether base rate or total compensation is under review
• ▸Observed disparity — the significant gap in dollars/rate and affected class
• ▸Omitted factors tested — list seniority, merit, production, education/training/experience added to model and result
• ▸Pattern check — does the disparity repeat person-by-person across the grouping? (yes/no + notes)
• ▸Corroborating evidence — decision-maker statements, documented pay-setting practices reviewed
• ▸Legitimate explanation found? — yes (describe & close) / no (proceed to remediation)
• ▸Legal counsel sign-off — name/date confirming remediation done under counsel
• ▸Adjustment plan — target equitable rate and confirmation no employee's pay is reduced
• ▸Documentation reference — where the investigation record and rationale are stored

1. 1Split the question 'do we compensate fairly?' into its four justice dimensions before answering: distributive (the outcome), procedural (the rules), interactional (the treatment), informational (the communication). Never treat 'fair' as one variable.
2. 2Test distributive justice statistically: examine actual compensation outcomes (bonus payments, merit increases, promotional increases) for disparate treatment — whether money was distributed equitably given your policies. This is where intentional discrimination shows up.
3. 3Test procedural justice statistically: examine your compensation policies and practices for disparate impact — whether the rules you use to distribute money affect different groups differently. This is where unintentional discrimination hides even when intent is absent.
4. 4For each pay decision type, write down the criteria used to determine it: how bonuses are calculated, how merit increases are set, what guidelines govern promotional increase size. If you cannot state the rule, employees cannot perceive it as fair.
5. 5Build informational justice into every pay communication: do not deliver only 'your bonus is $X' or 'your merit increase is Y%.' Attach the supporting reasoning for how the number was reached, because without informational justice there is no overall fairness.
6. 6Separate what you can quantify (distributive and procedural justice) from what you cannot (interactional and informational justice), and staff each accordingly — statistics for the first two, communication discipline and manager behavior for the latter two.
7. 7Run the analysis proactively rather than waiting for litigation or a regulatory investigation; a self-analysis lets you correct inequities before scrutiny begins, and past absence of complaints does not immunize you from future scrutiny.

An organization sets objective, well-defined compensation factors, distributes pay equitably against them, and has managers deliver decisions courteously. On paper, three of the four justice dimensions are satisfied — distributive, procedural, and interactional. But each employee is told only 'your merit increase is Y%' with no supporting explanation. Despite the equity and courtesy, employees perceive the decision as unfair because informational justice is absent. The fix is not changing the number or the tone; it is supplying the reasoning behind the number. As Thomas notes, even employees who disagree with the outcome or the procedure are more likely to perceive the decision as fair if they understand how it was made.

• If the complaint is 'I didn't get what I deserved,' then you're in distributive justice — test outcomes for disparate treatment.
• If the complaint is 'the process that set my pay was rigged,' then you're in procedural justice — test policies for disparate impact.
• If pay is internally equitable and courteously delivered but employees still call it unfair, then the gap is informational justice — add the supporting reasoning to the communication.
• If a decision-maker says 'I know we're doing things equitably, so we don't need to analyze,' then treat that as overconfidence, not evidence — proceed with the quantitative self-analysis anyway.
• If the reasoning behind a pay figure cannot be articulated in plain terms, then do not communicate the figure alone — a bare number invites a perception of unfairness.

• Collapsing 'fairness' into a single yes/no. Tell: leadership answers 'yes, we pay fairly' without distinguishing outcomes from procedures from communication.
• Treating equitable numbers as sufficient. Tell: pay is objectively equitable yet employees still feel wronged because communication was 'your bonus is $X' with nothing behind it.
• Assuming no intent means no problem. Tell: managers point to good intentions while procedural rules quietly produce disparate impact across groups.
• False sense of safety from a clean history. Tell: 'we've never had a problem before, so we won't have one' — the availability heuristic masking real risk.
• Fear-driven avoidance of analysis. Tell: 'what if we find problems? I'd rather not know,' which forecloses any chance to correct inequities before litigation.

• ▸Pay decision type: name it (bonus / merit increase / promotional increase).
• ▸Distributive check: state the actual outcome and whether it was distributed equitably given policy.
• ▸Procedural check: write the exact criteria/rule used to reach the outcome.
• ▸Disparate impact flag: note whether the rule affects any group differently.
• ▸Interactional check: was the outcome delivered with politeness, dignity, respect (yes/no + note).
• ▸Informational check: list the supporting information given beyond the bare number.
• ▸Perceived-fairness gap: which of the four dimensions is weakest and why.
• ▸Corrective action: the specific step to close the weakest dimension.
• ▸Proactive review date: when this self-analysis is rerun before external scrutiny.

1. 1Engage legal counsel before any data is pulled or analysis is run, so the review proceeds under attorney-client privilege and attorney work product protection; route all correspondence, data exchanges, and results through counsel, especially if outside consultants are involved, to preserve nondiscoverability.
2. 2State the review's goal explicitly as exposure assessment: separately scope 'exposure to compensation discrimination litigation' and 'exposure to regulatory investigation,' because the two implicate different agencies (EEOC, OFCCP) and different data demands.
3. 3Map your regulatory footprint: if you hold federal contracts you fall under OFCCP; recognize that the EEOC now runs an Equal Pay Act Directed Investigation Pilot Program and can initiate a compensation investigation at its discretion, not only in response to a filed charge.
4. 4Run multiple regression analysis on similarly situated employee groupings as the preferred technique, testing compensation outcomes for disparate treatment and testing policies/practices for disparate impact; where regression is inappropriate or impossible, fall back to raw-differential tests (means/medians, tipping point, threshold).
5. 5Examine general equity across all employees irrespective of protected characteristic, not gender alone; scope the analysis to every protected group and to white/male employees so the review reflects the actual exposure surface, not just the culturally-assumed gender question.
6. 6Investigate every grouping flagged by the tests, including directionally adverse patterns that are not individually statistically significant, and document the cause of each disparity (intent, disparate-impact policy, or model misspecification) before concluding.
7. 7Because Ledbetter amendments treat each paycheck issued under a discriminatory decision as a fresh violation, treat ongoing pay disparities as continuing liability rather than a one-time event, and remediate flagged disparities rather than assuming a limitations period has closed them.

Consider a company running a compensation review across several departments. The regression models return no statistically significant gender disparity in any single grouping, so a manager wants to close the file. But the reviewer notices that the majority of similarly situated employee groupings show a gender effect that is directionally adverse to women, even though none crosses the significance threshold. Per the passages, you do not take comfort in this: a collection of universally directionally adverse results, left uninvestigated, creates substantial exposure to a 'pattern or practice' discrimination claim. The reviewer keeps the file open and pursues the three candidate explanations — discriminatory intent by the decision makers in those departments, a policy or practice creating disparate impact on women, or misspecification of the models — until the underlying cause of each estimated disparity is identified. Because the work was routed through legal counsel from the outset, the interim results remain protected as attorney work product while the investigation continues.

• If you are conducting any exposure assessment, then bring legal counsel on board as early as possible and channel all data, correspondence, and results through them to secure privilege and nondiscoverability.
• If the majority of similarly situated groupings are directionally adverse to a protected group, then investigate further even when no single result is statistically significant, because the pattern itself creates 'pattern or practice' exposure.
• If multiple regression is inappropriate or impossible, then use raw-differential techniques (means/medians, tipping point, threshold), but treat their higher false-positive rate as expected and investigate every flagged grouping.
• If you hold federal contracts, then assume OFCCP may apply ad hoc analysis methods 'consistent with broad Title VII principles' and demand expanded compensation data — do not rely on the old Interpretive Standards methodology.
• If a pay disparity persists, then treat each subsequent paycheck as a new violation under the Ledbetter amendments rather than assuming the 180-day clock has expired.

• Treating pay equity as a gender-only issue: the tell is a scope that examines women and racial minorities but excludes white and male employees, which understates true exposure and misses strategic value.
• Closing the file on 'no statistically significant disparity' while ignoring a body of directionally adverse results — the tell is a summary that reports significance tests but never counts the direction of effects across groupings.
• Running the analysis outside privilege: the tell is consultants emailing draft results directly to HR rather than through counsel, making the work discoverable in later litigation.
• Relying on raw-differential tests without follow-up: the tell is treating a two-dimensional compensation-vs-protected-status comparison as conclusive, when unexamined false positives (disparities that vanish once compensation determinants are added) inflate apparent exposure.
• Assuming a discriminatory pay decision is time-barred: the tell is dismissing an old decision without recognizing that each paycheck paid under it is a fresh, actionable violation post-Ledbetter.

• ▸Review objective — state whether this assesses litigation exposure, regulatory investigation exposure, or both
• ▸Regulatory footprint — note federal contractor status (OFCCP) and EEOC directed-investigation applicability
• ▸Privilege setup — record counsel engaged and the routing channel for data/results
• ▸Protected groups in scope — list every group examined, including white/male employees for general equity
• ▸Similarly situated groupings — define each grouping and the technique applied (regression or raw-differential)
• ▸Significant disparities — list each grouping with a statistically significant adverse result
• ▸Directionally adverse patterns — count groupings adverse to each group even if not significant
• ▸Cause determination — for each flagged grouping, record intent / disparate-impact policy / model misspecification
• ▸Remediation and follow-up — note action taken and continuing-violation (paycheck) risk status

1. 1Extend your pay-equity analysis beyond gender and race to include white males, since the retention, engagement, and productivity relationships hold for every employee group irrespective of protected characteristic — general equity, not just protected-group equity, is what produces the strategic benefit.
2. 2Build the compensation package around objective and measurable criteria (base pay, commissions, bonuses with documented seniority/merit/productivity bases) so candidates and current employees can have the pay decision explained in clear terms — the mechanism that converts equity into retention.
3. 3Document the decision-making process in writing and communicate it transparently, adding supporting rationale to every pay statement rather than a bare 'your merit increase is Y%,' because informational justice is required for a decision to be perceived as fair even when the outcome is disputed.
4. 4To measure outcomes, collect company-specific data before and after introducing an internal pay equity policy and track it across multiple years — a single snapshot cannot show the retention or turnover-cost effect.
5. 5Track proxy outcome measures you can actually capture — turnover rate, turnover cost, absenteeism rate, performance/efficiency scores, engagement survey results — and pair them with the equity metric, while accepting you cannot claim a precise 'X% equity → Y% engagement' coefficient.
6. 6Confirm the qualitative drivers that co-determine engagement are in place — work-role fit, clear goals, effective leadership, training and development opportunities — so equity gains are not swamped by these confounds.
7. 7Because equity's link to productivity has no direct line of sight, present the business case qualitatively and manage expectations: the benefit is real (attraction/retention of top talent, lower absenteeism, higher efficiency, engagement, innovation) but not quantifiable to a single ratio.

The 1951 U.S. Department of Labor case studies are the concrete precedent to cite. Management in those studies reported that after internal pay equity was improved, women's attitudes toward their work improved, their efficiency increased, and turnover decreased. This is the pattern to expect: you cannot isolate a percentage, but the direction is consistent — fairer, explainable pay reduces turnover and raises efficiency. Use it to frame the modern business case Secretary Solis and others make: retention rises when employees are paid on objective, explainable criteria, and higher retention lowers turnover cost. The lesson for your own program is to instrument the same before/after measures management captured in 1951 — attitude, efficiency, turnover — rather than waiting for a single quantified equity-to-productivity ratio that the empirical literature still does not provide.

• If overall equity is your concern, then examine total compensation (base + commissions + bonuses); if you are investigating discriminatory pay-setting, then examine base rate of pay only.
• If a benefit claim depends on a precise numeric link between equity and productivity/engagement/turnover, then do not make it — no study has quantified that 1% equity yields X% of any outcome.
• If you scope the review only to women and minorities outside litigation/regulatory need, then you forfeit the strategic advantage, because white and male employees are equally sensitive to distributive, procedural, interactional, and informational justice.
• If pay communication is limited to the bare amount with no supporting rationale, then treat the decision as likely to be perceived unfair regardless of how equitable the underlying outcome is.
• If you run a formula-based system, then apply it consistently to all eligible employees without exception and document the criteria and business reasons in writing.

• Treating pay equity as a compliance, diversity, or social-responsibility issue only — the tell is a program justified purely on legal risk with no retention/turnover-cost metrics attached.
• Automatically equating internal pay equity with gender and stopping there — the tell is an analysis that never checks white-male equity and then puzzles over persistent turnover in that group.
• Promising a quantified ROI ('1% equity boosts productivity Z%') — the tell is a projection citing a coefficient the passages explicitly say does not exist.
• Base-pay data captured in inconsistent units (hourly for some, annual for others, no full-time-equivalency adjustment) — the tell is an apparent disparity within a similarly situated group that vanishes once units are reconciled.
• Attributing an engagement or retention change entirely to pay equity while ignoring confounds like work-role fit, leadership, and training — the tell is one employee satisfied with a conference while another finds it inadequate, i.e., the outcome varies person to person.

• ▸Employee grouping: name the similarly situated group under review
• ▸Compensation metric selected: base / commissions / bonuses / total, and why
• ▸Data units check: confirm hourly vs. annual and full-time-equivalency reconciled
• ▸Baseline outcome measures (pre-policy): turnover rate, turnover cost, absenteeism, efficiency/performance, engagement score
• ▸Policy change and date implemented: what equity action was taken
• ▸Post-policy outcome measures + years tracked: same metrics, dated
• ▸Qualitative confounds present: work-role fit, clear goals, leadership, training/development
• ▸Transparency check: is decision rationale documented and communicated with each pay statement?
• ▸General-equity scope confirmation: are white/male and all protected groups included?

1. 1Define the objective criteria that govern each pay decision type before making decisions: bonus payment criteria, merit increase determination rules, and guidelines governing the size of promotional increases. Document each one so a decision can be traced back to a written rule.
2. 2Anchor job-comparison criteria to job content, not job title, using the four EPA factors: skill (ability, education, training, experience the job requires), effort (physical or mental exertion required), responsibility (degree of accountability), and working conditions (physical surroundings). Record what the job requires, not what the incumbent happens to possess.
3. 3Separate the two fairness questions your policy must satisfy: distributive justice ('Did I receive what I should have received?') tests outcomes for disparate treatment; procedural justice ('Was what I received determined fairly?') tests the rules themselves for disparate impact. Assign a distinct check to each.
4. 4Run a proactive statistical self-analysis — multiple regression is the standard tool — on your compensation outcomes to surface pay differences by protected group, then test whether legitimate nondiscriminatory factors explain them before those factors are omitted.
5. 5Add informational justice deliberately: for every pay decision, communicate not just the number ('your bonus is $X') but the supporting rationale showing which documented criteria produced it. Without this, even objective, respectfully-delivered decisions read as unfair.
6. 6Screen any criterion that operates to exclude a group for business necessity and manifest job-relatedness, following Griggs — a requirement fair in form but discriminatory in operation (like a diploma or test with no demonstrable relationship to job performance) is prohibited unless you can document its relationship to the work.

Consider two bookkeeping jobs where one incumbent holds a master's degree in physics. Under the EPA analysis, you evaluate the jobs as substantially equal because the criterion is what skill the job requires, not what the individual has — the physics degree is not required for bookkeeping, so it cannot justify a pay differential. Now compare an assembly line where one position must also lift each assembled product off the line and place it on a board. Here the added lifting is a substantial, regular part of the job, so it increases the required effort; paying that position more is not a violation regardless of who holds it. The tell of a clean policy is that each pay gap maps to a documented job-content factor (skill, effort, responsibility, working conditions), while a minor difference like turning out the lights at day's end does not justify differential pay because it fails the substantiality standard.

• If a pay difference cannot be traced to a documented skill, effort, responsibility, or working-condition factor tied to job content, then treat it as unexplained and correct it before charges are filed.
• If a criterion excludes members of a protected group in operation, then it is prohibited unless you can demonstrate its manifest relationship to job performance and business necessity (Griggs standard).
• If you distinguish jobs, compare job content rather than job title, because identical work is not required — only substantially equal work.
• If the only communication of a decision is the outcome number with no supporting rationale, then expect it to be perceived as unfair even when it is objectively equitable — add the rationale.
• If a factor reflects what the individual possesses rather than what the job requires, then exclude it from the equal-work analysis.

• Comparing job titles instead of job content — the tell is two roles labeled differently but requiring the same skill, effort, responsibility, and conditions being paid unequally with no documented basis.
• Justifying differentials with the incumbent's credentials the job does not require (e.g., paying more because a bookkeeper holds a physics degree) — the tell is a factor that describes the person, not the position.
• Confusing distributive and procedural justice: fixing outcome numbers while leaving discriminatory rules intact — the tell is a rule that disproportionately affects a group even after individual pay is 'corrected.'
• Neglecting informational justice — communicating 'your merit increase is Y%' with no rationale — the tell is employees perceiving fair decisions as arbitrary because the how was never explained.
• Waiting for a formal charge before running any analysis — the tell is no before-and-after data and no regression on file, leaving no opportunity to identify and correct problem areas proactively.

• ▸Decision type: bonus / merit increase / promotional increase — name which rule applies
• ▸Governing documented criterion: cite the written rule used
• ▸Skill required by the job: abilities, education, training, experience the position demands (not the incumbent's)
• ▸Effort required: physical or mental exertion, noting any substantial, regular added tasks
• ▸Responsibility level: degree of accountability the job carries
• ▸Working conditions: physical surroundings, hazards, ventilation, temperature
• ▸Business-necessity check: if a criterion excludes a group, state its manifest job-relationship
• ▸Rationale communicated to employee: the supporting information given beyond the number
• ▸Statistical review status: date of last regression/self-analysis and unexplained gaps found

1. 1Define the measurement target before touching data: state explicitly what compensation outcome you are testing (base salary, bonus, merit increase, promotional increase) and which relationships you will explore, because the analysis is only as good as this clarity.
2. 2Build cradle-to-grave employment histories for each employee rather than a single-point snapshot; capture hire date, seniority clock start, halts or losses of seniority, and events affecting accrual so time-based determinants are reconstructable.
3. 3Assemble determinant variables that map to the Equal Pay Act's factors-other-than-sex: seniority, merit-system ratings, quantity/quality-of-production measures, and bona fide job-related factors such as education, training, and experience.
4. 4Run a completeness, accuracy, and consistency review across every field: check for missing values, internally inconsistent codes (e.g., a job title that conflicts with cost center), and out-of-range figures before any regression is estimated.
5. 5Establish similarly situated employee groupings so each employee's pay is compared against the correct peer group; do not default to raw job-title or cost-center bins without confirming they actually represent comparable work.
6. 6Flag tainted variables (those subject to employer discretion, like performance ratings that may themselves reflect discrimination) and, rather than automatically dropping them, document their presence and assess and minimize their effect on the estimates.
7. 7Verify machine-readability: confirm dates, codes, and numeric fields parse consistently so the data can be analyzed without manual reinterpretation that reintroduces error.
8. 8Re-run the completeness and consistency checks after any merge or transformation, since incorrect inferences from bad data, if fed into business decisions, can create the very inequity you set out to remedy.

You want to examine annual salaries of the 'headquarters' cost center for gender pay discrimination. The instinct is to pool everyone assigned to that cost center into one grouping and compare men to women. But the passage shows this is rarely correct: employees in the same cost center hold different job titles and perform non-equivalent work, so a single pooled comparison would compare people who are not similarly situated. Before running the regression, you regroup employees so that each person's salary is assessed against genuine peers doing equal or equivalent work. You then confirm each record carries a reliable seniority clock and merit-rating field, because those are the lawful factors-other-than-sex that legitimately explain pay differences. Only after the groupings and determinant fields are verified for completeness and consistency do you estimate the model, because otherwise the 'garbage in, garbage out' problem produces a false finding of discrimination or masks a real one.

• If the grouping is built purely on job title or cost center without confirming comparable work, then reconstruct similarly situated groupings before analyzing.
• If a determinant variable is subject to employer discretion in its measurement or use, then treat it as tainted: assess and minimize its effect rather than routinely excluding it, since it may be an important productivity proxy.
• If key predictor variables (seniority, merit, production, education/training/experience) are missing, then the model is incomplete and inferences are unreliable — do not report results.
• If a seniority or merit system is invoked as an explanation, then confirm it is applied uniformly and considered regularly, not case-by-case, before crediting it in the data.
• If data show inconsistencies or missing values, then correct them before estimation, because incorrect inferences applied to business decisions can create the inequity you intended to fix.

• Treating a raw job-title or cost-center bin as a peer group; the tell is a 'headquarters' grouping mixing dissimilar roles that produces a pay gap driven by job mix, not sex.
• Routinely dropping tainted variables to be safe; the tell is an underspecified model missing important productivity proxies, weakening the reliability of every inference.
• Analyzing a point-in-time snapshot instead of cradle-to-grave histories; the tell is an inability to reconstruct when the seniority clock started or was halted.
• Skipping the completeness/accuracy/consistency review and trusting exported data; the tell is 'garbage in, garbage out' — results that look precise but rest on inconsistent codes or missing fields.
• Feeding uncorrected inference into business decisions; the tell is an employer unintentionally creating the very disparity it set out to remedy.

• ▸Measurement target: name the exact pay outcome being tested
• ▸Relationships to explore: state which determinants you expect to explain pay
• ▸Similarly situated grouping basis: describe how peer groups were defined and validated
• ▸Seniority history: hire date, clock start, halts/losses, accrual rules captured (Y/N)
• ▸Merit/production fields: rating or production measures present and machine-readable (Y/N)
• ▸Bona fide factors: education, training, experience fields captured (Y/N)
• ▸Tainted variables flagged: list variables subject to employer discretion and effect-minimization plan
• ▸Completeness/accuracy/consistency review done: note missing values, code conflicts, out-of-range figures corrected
• ▸Post-merge re-check: confirm checks re-run after any transformation (Y/N)

1. 1Start with the OFCCP definition as your anchor: group only employees who perform similar work, occupy positions with similar responsibility levels, and involve similar skills and qualifications. Do not default to grouping by job title or cost center alone.
2. 2List every candidate grouping variable before running any comparison: job function, responsibility level, required skills, required qualifications. These four are the core criteria.
3. 3Layer in the OFCCP 'pertinent factors' that legitimately separate otherwise-similar employees: different departments/functional divisions with different budgets or importance, different pay plans (team-based vs. incentive-based), different pay basis (hourly, salary, commission), collective bargaining coverage vs. not, and employment status (full-time vs. part-time).
4. 4Add other pertinent factors specific to your organization, such as geography or location, when locality adjustments apply to pay.
5. 5Test each proposed grouping by asking whether a raw average comparison inside it would be misleading — if wildly different roles sit together (CEO next to receptionist), the grouping is invalid regardless of shared cost center.
6. 6Confirm each grouping has enough employees of each protected-group status to support statistical inference. A grouping with only one woman cannot yield a t-statistic or p-value.
7. 7Re-examine groupings if diagnostics flag heteroscedasticity, since the presence of heteroscedasticity may reflect how similarly situated groupings were defined (alongside pay structure type and compensation type).
8. 8Document the rationale for each grouping so a reviewer can trace which factors and metrics justify who was compared with whom.

Table 4-1 groups all employees in the 'headquarters' cost center: a male CEO at $250,000, a female Executive Assistant to the CEO at $80,000, a female Admin Assistant to the CEO at $65,000, a female Admin Assistant II at $57,000, and a female Receptionist at $30,000. A simple comparison of average annual salaries by gender shows a $192,000 gap in favor of males, which tempts an inference of gender discrimination. But the gap is not attributable to sex — it is due to grouping employees inappropriately for comparison. The CEO, executive assistants, admin assistants, and a receptionist do not perform similar work, occupy similar responsibility levels, or require similar skills and qualifications. Because the grouping mixes fundamentally different jobs, the inference is inaccurate and misleading. The correct move is to compare each employee only against genuine peers who satisfy the OFCCP similarly-situated criteria.

• If two employees differ in work performed, responsibility level, required skills, or qualifications, then do not place them in the same grouping — even if they share a cost center, department, or title.
• If employees are otherwise similar but fall under different pay plans, pay basis, collective bargaining coverage, or employment status, then treat those as pertinent factors and separate them.
• If locality adjustments apply, then include geography/location as a grouping factor.
• If a grouping contains only one employee of a protected group, then flag it — statistical significance cannot be calculated (as with the single-woman forklift-operator grouping).
• If diagnostics show heteroscedasticity, then revisit how the similarly situated groupings were defined before trusting the estimated parameters.

• Grouping by cost center alone — the tell is a raw average gap that looks enormous (e.g., $192,000) because a CEO and a receptionist are being compared as peers.
• Grouping by job title alone and assuming that ends the analysis — the tell is that duties within the 'same' title vary by experience and specialty, so ostensibly identical roles are not actually comparable.
• Treating a non-significant estimated gender effect as proof of no discrimination — the tell is a grouping with tiny protected-group counts (e.g., three women among 36 repair personnel, p = 0.4237) where power is too low to detect an effect.
• Ignoring OFCCP pertinent factors and lumping incentive-based, commission, and salaried employees together — the tell is unexplained variance or heteroscedasticity traceable to mixed pay plans or pay basis.
• Building groupings with no documented, articulable rationale — the tell is that a reviewer cannot identify what factors and metrics justified who was compared, meaning the grouping (like a compensation policy) should be revisited.

• ▸Proposed Grouping Label: name the tentative peer group (e.g., 'customer service representatives').
• ▸Work Performed: describe the core duties; confirm they are similar across members.
• ▸Responsibility Level: state the responsibility tier; confirm it is consistent.
• ▸Required Skills & Qualifications: list them; flag any member who deviates.
• ▸Pay Plan / Pay Basis: record plan type (team, incentive) and basis (hourly, salary, commission); separate mismatches.
• ▸CBA & Employment Status: note collective bargaining coverage and full/part-time status; split where different.
• ▸Location Factor: record geography if locality adjustments apply.
• ▸Protected-Group Counts: enter headcount by protected-group status; flag if any count is too small for inference (e.g., 1).
• ▸Grouping Rationale & Sign-off: written justification a reviewer can trace, plus reviewer initials.

1. 1Specify the compensation outcome you are testing (base salary, bonus, merit increase) as the dependent variable, and treat protected class status (sex, race, national origin) as the variable of interest whose coefficient you interpret.
2. 2Build the model around the statutory factors that constitute legitimate pay determinants: the Equal Pay Act recognizes seniority systems, merit systems, systems measuring earnings by quantity or quality of production, and bona fide factors other than sex such as education, training, or experience — include these as explanatory variables so the protected-class coefficient reflects pay differences after accounting for them.
3. 3Restrict the analysis to comparators doing 'equal work' requiring equal skill, effort, and responsibility under similar working conditions; define the analysis group before running the model rather than after seeing results.
4. 4Collect data spanning enough time and enough employees to support inference — the passages note meaningful analysis requires tracking data over a number of years and before/after a policy change; verify you have the company-specific data needed rather than accepting an incomplete extract.
5. 5Run the multiple regression, then read the coefficient and its statistical significance on the protected-class variable: a significant negative coefficient for the protected group, controlling for legitimate factors, is the pattern introduced as indirect (circumstantial) evidence of discrimination.
6. 6Cross-check the regression finding against the 'other tests of equal pay' rather than relying on one model; corroborating results across methods reduces the risk that a single specification is driving the conclusion.
7. 7Interpret results in the correct legal frame: a statistical disparity establishes only a prima facie inference — the employer can rebut with an affirmative defense (seniority, merit, production, or another bona fide factor), so anticipate which defense variables must be tested.

Suppose you test base salary for a job group where women are paid less than men. You regress base salary on sex plus the statutory factors — seniority, merit ratings, production/quality measures, and education, training, and experience. If the raw gap is, say, eleven cents (echoing the reported eleven-cent-per-hour differential among federal employees the task force is studying), the disciplined question is how much of that gap survives after the legitimate factors enter the model. If the sex coefficient remains significantly negative, you have circumstantial evidence supporting a prima facie case; if it disappears once education and experience are controlled, the gap is explained by a factor other than sex — a recognized affirmative defense. The passages stress the government's own inability to understand the 'full scope' of the differential because of insufficient data, which is exactly why the model's explanatory variables and data adequacy determine whether the interpretation holds. Report the residual gap, its significance, and which controls absorbed it rather than reporting only the raw difference.

• If a pay factor is one of the four statutory defenses (seniority, merit, quantity/quality of production, or a bona fide factor like education/training/experience), then include it as a control before interpreting the protected-class coefficient.
• If the protected-class coefficient loses significance once a legitimate factor is added, then treat the disparity as explained by that factor other than sex rather than as evidence of discrimination.
• If you cannot obtain company data over multiple years and before/after the policy, then flag the analysis as underpowered rather than presenting a definitive finding, since meaningful analysis requires that longitudinal data.
• If a single regression shows a disparity, then confirm with the other tests of equal pay before concluding, because corroboration across methods is stronger than one specification.
• If the comparators are not doing equal work under similar working conditions, then narrow the analysis group before modeling, since the statute frames equity around equal skill, effort, and responsibility.

• Reporting the raw wage gap without controlling for legitimate factors — the tell is a headline differential (like eleven cents) presented with no explanatory variables, inviting a defense that a factor other than sex explains it.
• Omitting a statutory defense variable (education, training, experience, seniority, merit) so the protected-class coefficient absorbs its effect — the tell is a large, significant coefficient that shrinks dramatically when an obvious control is added.
• Defining the comparison group after seeing results — the tell is a job group assembled to maximize the disparity rather than one built on equal-work criteria set beforehand.
• Drawing conclusions from data that are too thin — the tell is a single cross-section with no longitudinal or before/after coverage, which the passages identify as insufficient for meaningful analysis.
• Treating a statistical disparity as proof rather than as indirect evidence establishing a prima facie case — the tell is skipping the affirmative-defense stage where the employer can articulate a legitimate nondiscriminatory reason.

• ▸Dependent variable: name the compensation outcome tested (base salary, bonus, merit increase).
• ▸Protected-class variable: state the basis being tested (sex, race, or national origin).
• ▸Analysis group: define the equal-work group by equal skill, effort, responsibility, working conditions.
• ▸Statutory control variables: list included factors (seniority, merit, quantity/quality of production, education, training, experience).
• ▸Data span: record years covered and whether before/after policy data exist.
• ▸Sample size: state number of employees and comparators in the group.
• ▸Result: record protected-class coefficient, sign, and statistical significance.
• ▸Corroboration: name the other test(s) of equal pay run to confirm.
• ▸Affirmative-defense check: note which legitimate factor could rebut the finding and whether it was tested.

1. 1Define the similarly situated employee grouping first (similar work, skills, qualifications, responsibility levels), then confirm the compensation metric you are analyzing is the right one — base pay, annual salary, or overtime — because raises alone can mask disparities in underlying salary levels.
2. 2Audit the units of measure before running anything: verify every pay figure is expressed consistently (all annual or all hourly), since a mix of $10/hour and $20,800/year produces a spurious 'statistically significant' gender disparity.
3. 3Screen for full-time versus part-time mixing and normalize by hours worked per week; annualize or convert to an hourly rate so a $15,600 part-timer at 30 hours is not mistaken for pay discrimination against a $20,800 full-timer at 40 hours.
4. 4Specify legitimate nondiscriminatory determinants (human capital: knowledge, skills, education, training; plus time in job, seniority, performance rating) and enter them into a multiple regression so the estimated protected-group coefficient reflects pay net of those factors.
5. 5Read the estimated protected-group slope coefficient (e.g., the 'female' coefficient) as the disparity after controls, then test it for statistical significance against the two-to-three standard-deviation threshold — do not rely on ratio rules like four-fifths that ignore sample size and magnitude.
6. 6For overtime earnings, add a measure of overtime hours worked to the model, because differences in overtime pay by protected group are typically driven by differences in overtime hours (from choice or seniority-based allocation), not by discrimination.
7. 7Flag every grouping showing a disparity for follow-up; if the coefficient is directionally adverse to the same group across many groupings, run systemic follow-up (start with organizational structure), otherwise pursue individual follow-up to find explanations outside the model.

Take Table 4-5: six employees identical on all characteristics except gender and current performance rating. A statistical review of raises shows no gender disparity — everyone rated 10 got 3.5% and everyone rated 8 got 3.0%, regardless of sex. But examining annual salary reveals the disparity: every male sits at a $56,000 old salary while every female sits at $45,000, an $11,000 gap that persists in the post-raise figures ($57,960 versus $46,575). Because the employees differ only by gender, this points to potential gender discrimination embedded in the base salary level. Had you evaluated only raises, the detected disparity would have gone completely undetected. The lesson: percentage raises carry no information about the salary levels they are applied to, so the wrong metric hides a real disparity.

• If a regression flags a significant protected-group disparity, then first check whether pay units are mixed (hourly vs. annual) or full-time/part-time employees are combined before concluding discrimination.
• If you can only examine one metric, then use salary levels rather than raises, because equitable raises can sit on top of discriminatory base salaries.
• If the protected-group coefficient exceeds two to three units of standard deviation, then treat it as statistically significant; if it falls below (as in the 1.1-SD example), then do not infer disparity regardless of what a four-fifths-style ratio suggests.
• If overtime earnings differ by protected group, then control for overtime hours worked before attributing the difference to discrimination.
• If the estimated effect is adverse to the same group across all or most groupings (as in Table 7-1's ten groupings), then escalate to systemic follow-up starting with organizational structure rather than case-by-case review.

• Analyzing raises instead of salary levels — the tell: raises look perfectly equitable (3.5% and 3.0% by rating) while base salaries diverge by thousands ($56,000 vs. $45,000).
• Feeding mixed pay units into the regression — the tell: an implausibly clean 'significant' disparity where one group's figures look like $10 (an hourly rate) against another's $20,800 (annual).
• Overlooking part-time/full-time mixing — the tell: annual pay differences that are 'plausible' and not large enough to trigger visual inspection, but that vanish once hours per week are added.
• Using the four-fifths rule as a significance test — the tell: it declares disparate impact at a 67% ratio on six employees (only 1.1 SD) yet clears an 89% ratio on 10,000 (14 SD), because it ignores sample size and magnitude.
• Skipping follow-up on detected disparities to save cost — the tell: a documented disparity was 'detected but not remedied,' the exact exposure that leaves the organization vulnerable to litigation, as the AstraZeneca $250,000 settlement illustrates.

• ▸Similarly situated grouping: state the shared work, skills, and responsibility level.
• ▸Compensation metric: name it (base pay / annual salary / overtime) and confirm it is not raises-only.
• ▸Unit consistency check: confirm all pay is in the same unit (all annual or all hourly).
• ▸Employment status check: note full-time vs. part-time mix and hours per week used to normalize.
• ▸Legitimate determinants entered: list human capital, time in job, seniority, performance rating.
• ▸Estimated protected-group coefficient: record the sign and dollar/percentage size after controls.
• ▸Significance in SD units: record value and whether it clears the two-to-three-SD threshold.
• ▸Overtime control: note whether overtime hours worked were included when analyzing overtime pay.
• ▸Follow-up type: mark systemic (pattern across groupings) or individual, and the explanation sought.