Analysing pay gaps responsibly

1. Separate three questions

The current visual answers the first question and helps explore groups for the second. It does not estimate a regression, conduct an Oaxaca–Blinder decomposition or produce a statutory reporting return.

Blau and Kahn's review identifies the continuing importance of occupation, industry, work interruptions and working hours, while experimental evidence means discrimination cannot be dismissed. Its primary empirical context is the US. The research does not supply a ready-made German employer model or an explanation for an individual employee's pay. Blau & Kahn (2017), JEL 55(3), 789–865.

2. Write the estimand and denominator

For the displayed female/male comparison, let M be the mean male pay and F the mean female pay among valid observations under the same measure and filters:

Mean gap (%) = 100 × (M − F) / M.

The median gap substitutes the two medians. Positive means female pay is lower relative to the male denominator. Negative means it is higher. With M = €60,000 and F = €54,000 the gap is 10%. Raising €54,000 to €60,000 requires 11.11% of female pay: the percentages differ because the denominator differs. With a zero male denominator the ratio is undefined, not zero.

Eurostat's unadjusted gender pay gap uses average gross hourly earnings, with a defined statistical population. ONS's principal ASHE comparison uses median hourly earnings excluding overtime. These figures cannot be compared directly with an employer's annual FTE mean gap merely because each is called a gender pay gap. Eurostat metadata, ONS interpretation guide.

Report at least the measure, dates, population, group counts, mean/median, missingness and exclusions. Analyse base, target opportunity and actual outcomes separately. Bonus analysis should distinguish eligibility, receipt, zero awards, target percentage and payout relative to target. A recipients-only median can look equal despite unequal access to bonus plans.

The visual reads Female, Male, Other and Unknown categories without inferring them from names. Its binary gap uses valid Female and Male observations. Other/Unknown are retained as categories and do not become fabricated male/female records. Verify the displayed counts and the data dictionary. The appropriate legally reportable sex/gender variable and treatment of categories require a separate decision for the applicable reporting regime.

3. A constructed example: composition and within-job differences

Men's mean is €72,000. Women's mean €48,000. The overall gap is 33.33%. Within each job group the gap is zero. This does not prove that the organisation is equitable: access to higher-paid work may still require investigation. Nor does the overall 33.33% prove unequal pay within each job. Both observations are true and answer different questions.

Reverse patterns can also occur: aggregate similarity can hide substantial differences within comparable groups. Always inspect both composition and within-group pay. Do not average job-level percentage gaps as if that necessarily reproduced the overall gap. Calculate each target statistic from its defined population.

An overall gap can change when people are hired, leave, are promoted or move between groups, even without anyone's pay changing. Present like-for-like employee changes and workforce-composition changes separately when explaining trends. A lower headline gap is not, by itself, evidence that a remediation policy worked.

4. Build comparable-work groups before explaining differences

A shared job title or grade is a starting point, not proof of equal work or equal value. Review skills, effort, responsibility and working conditions, including interpersonal and care-related demands that simplistic job classifications can undervalue. Apply criteria to work rather than to the demographic profile of its incumbent. ILO provides a practical gender-neutral job-evaluation method. ILO guide.

Our recommended review separates the factual difference, the proposed explanation, the evidence supporting it and the decision. Relevant experience, objective performance, location policy or a progression step may be pertinent. They are not automatically legitimate just because a database column exists. Check consistency, relevance, measurement quality and potential bias. Prior salary or negotiation history can perpetuate earlier differences. Do not use them as unquestioned fairness benchmarks.

Goldin's research emphasises the role of job structures that disproportionately reward long or particular hours. Consequently, merely dividing annual pay by an FTE factor need not remove all working-time-related differences. A practical review should consider access to flexible work, desirable assignments and promotion pathways alongside pay rates. Goldin (2014), AER 104(4), 1091–1119.

5. Adjusted analysis: a model specification, not a verdict

A common illustrative specification for positive hourly pay is:

ln(pay_i) = α + β·Female_i + X_i′γ + ε_i.

X might contain carefully selected job-value, location, experience and employment characteristics. With no interactions and the same X, 100 × (exp(β) − 1) expresses the modelled relative female/male difference on the log-pay scale. To express a positive female shortfall use 100 × (1 − exp(β)). For β = −0.05 that shortfall is about 4.88%, not exactly 5%. This concerns conditional geometric/log-scale comparison. Arithmetic-mean predictions need appropriate retransformation. Interactions make the comparison vary with X.

Logib provides an official Swiss operational framework, including a regression-based module with prescribed data preparation and a separate approach for smaller organisations. It is a useful methodological reference, not evidence that our visual implements it or that Swiss thresholds constitute German compliance. FOGE Logib FAQ, Module 1 guideline, version 2024.1.

Our recommended model-review procedure is:

1. Specify the pay concept, population and intended interpretation before viewing the gender coefficient. Keep the original unadjusted results alongside the model.
2. Audit predictors. A current grade or performance score may itself reflect unequal opportunity or biased assessment. Adding it changes the question from total differences to differences conditional on that potentially affected variable.
3. Check common support. If a job has only men or only women, within-job comparisons rely on other groups or modelling assumptions. Extremely fine categories can remove meaningful comparators.
4. Document missingness, zero pay, outliers and exclusions. Logs cannot include zero. Do not add €1 merely to avoid a software error. For bonus data, a participation/receipt model and a conditional amount model may be more informative.
5. Review functional form, interactions, collinearity and residual patterns. Match uncertainty estimation to the sampling/dependence structure. Repeated employee observations are not independent new people.
6. Publish a prespecified set of reasonable sensitivity specifications and explain differences. Do not select the model that produces the smallest or least significant gap.
7. Have a qualified analyst review the data and model, and reward/legal specialists review proposed explanations and consequences. Keep code and versioned inputs reproducible under appropriate access controls.

These are analytical recommendations, not claims that the present dashboard has executed these steps. “Adjusted” means conditional on the chosen model. Omitted variables, measurement error, selection and modelling choices can change the residual difference. A small coefficient does not establish absence of unequal treatment. A large coefficient does not identify which individual decisions caused it.

6. Decomposition answers another question

For group-specific linear models of mean log pay, a twofold decomposition with chosen reference coefficients β* can be written:

Δ = (X̄_m − X̄_f)′β* + X̄_m′(β_m − β*) + X̄_f′(β* − β_f).

The first term is the part associated with differences in measured characteristics under the reference returns. The remaining terms reflect coefficient differences. Include the intercept in X. Choice of reference coefficients affects the split. Categorical predictors and detailed contributions need care. Jann provides the method and implementation issues. “Explained” is a statistical label, not a finding that the factors are fair. “unexplained” is not a direct discrimination estimate. Jann (2008), Stata Journal 8(4), 453–479.

The practical purpose is to identify questions for investigation, such as occupational distribution, progression or differences within a role. Do not turn a decomposition into an automatic individual salary adjustment. Individual decisions require case-level facts and the organisation's lawful pay rules.

7. Uncertainty, small groups and confidentiality

For a complete current employee census, the observed descriptive gap is a fact about that dataset, subject to data error and definitions. Sampling confidence intervals address an explicitly assumed broader population or process. They are not required to make the observed arithmetic real. Conversely, a large employee count does not remove bias from inconsistent pay definitions or missing workers.

If inferential analysis is used, state the target population, dependence assumptions, estimator and confidence interval. Statistical significance is different from practical importance. A non-significant result in a small group is not proof of equality. Many subgroup searches also increase the chance of striking results arising by chance. Prespecify priority comparisons and disclose exploratory work.

Privacy thresholds and statistical sufficiency are separate. Our recommended publication policy is to set a minimum disclosed cell size with the organisation's data-protection and employee-representative review, then test complementary suppression and drill-down paths. Hiding a tooltip or collapsing detail is not an access-control mechanism. The current visual supports employee dots and cards and must only receive information its audience is entitled to see. Use Power BI model/access controls and a suitable aggregate report for wider audiences.

8. Germany/EU legal context — limited to the cited EU instrument

Directive (EU) 2023/970 requires objective gender-neutral equal-value criteria (Art. 4). Its reporting timetable is: 250+ workers, annually from 7 June 2027. 150–249, every three years from that date. 100–149, every three years from 7 June 2031 (Art. 9). National rules may go further. The transposition deadline was 7 June 2026 (Art. 34). This guide has not established Germany's current implementation or your employer's filing obligations.

Art. 10's joint-assessment trigger combines a category-level average difference of at least 5%, absence of objective gender-neutral justification, and failure to remedy it within six months of reporting. 5% is not a general lawful tolerance or automatic discrimination verdict. A dashboard filter or regression result does not substitute for the required legal assessment. Directive text, Arts. 4, 9, 10 and 34.

9. From diagnosis to a defensible action plan

Our recommended action record identifies the issue, affected population, evidence, explanation, remedy, cost, owner and review date. Correct extraction/classification errors first. Investigate inconsistent offers, progression, incentive eligibility and discretionary awards. Address unjustified differences under the applicable legal and employment framework. Do not reduce someone else's pay to cosmetically close a gap. Track recurrence through hiring, promotion, performance and reward decisions.

Communicate the method, uncertainty and planned action in language employees can understand. Cullen's review distinguishes horizontal, vertical and cross-firm transparency, with differing effects and trade-offs. Publishing a number alone is therefore not a complete transparency policy. Cullen (2024), JEP 38(1), 153–180.

10. Required interpretation beside this visual

Unadjusted annual FTE pay comparison. The chart compares the selected pay measure among valid female and male observations in the current filter context. It does not adjust for job value, experience, hours patterns or other characteristics. It is not an hourly-pay statutory return, an equal-pay compliance certificate, or a causal estimate of discrimination. Use it to identify and communicate differences, then investigate them with the methods appropriate to the question.