Regression analysis can help a pay equity audit estimate whether a pay difference associated with sex remains after selected characteristics such as job, grade, location or working time are considered. It is most useful when the dataset is large enough, the variables are reliable and the model specification reflects the compensation system being analysed. Directive (EU) 2023/970 does not prescribe regression as a mandatory universal method, so employers should not use a model to replace Article 9 reporting or the Article 4 assessment of same work and work of equal value. A defensible regression analysis requires documented variables, diagnostics, sensitivity testing and careful interpretation of both statistical and practical significance.
Jurisdiction: European Union
Regression Estimates Pay Differences While Holding Selected Factors Constant
A regression model treats pay as an outcome and estimates how that outcome is associated with several measured characteristics at the same time. In a pay equity audit, the model may include a sex indicator together with job, grade, location, working-time status, tenure or other variables that are relevant to the organisation's pay system. The coefficient associated with sex can then be interpreted as an estimated remaining difference under that specification. The model does not discover causation automatically. Its result depends on the data, variables, functional form and assumptions chosen by the analyst.
Use Regression Only When the Data Can Support It
Regression becomes unstable or misleading when the dataset is too small, key fields are missing or categories contain very few observations. A model with many controls can consume the available information quickly, particularly if job or location fields have many levels. Before modelling, analysts should examine record counts, missingness, pay distributions, worker-category sizes and whether the chosen variables have enough variation. Small groups may be better reviewed descriptively or through individual compensation records rather than forced into a complex statistical model that produces precise-looking but unreliable estimates.
Control Variables Need Legal and Business Justification
The strongest statistical model is not necessarily the one with the most controls. Every variable should have a reason for inclusion and a clear definition. Job and grade may reflect legitimate pay architecture, but analysts should consider whether those structures themselves were established through objective gender-neutral criteria. Performance ratings may be relevant in some systems but may also contain managerial discretion. Prior salary can be particularly problematic because historic inequality can be carried into current pay. The model specification should therefore be reviewed by compensation, analytics and legal stakeholders rather than chosen solely to maximise statistical fit.
Check Diagnostics Before Interpreting the Sex Coefficient
A coefficient should not be interpreted in isolation. Analysts should review residual patterns, influential observations, outliers, sparse categories and whether highly correlated variables are making estimates unstable. They should also test whether pay relationships are plausibly linear on the chosen scale and whether a transformed pay measure is more appropriate for the analytical question. Diagnostics do not create one universally correct model, but they reveal where the chosen specification may be fragile. A sensitivity analysis using reasonable alternative specifications can show whether the substantive conclusion is stable.
Statistical Significance Is Not the Same as Equal-Pay Compliance
A conventional significance test estimates how compatible the observed result is with a statistical null hypothesis under the model assumptions. It does not determine whether a pay difference is legally justified, whether two jobs are work of equal value or whether a compensation practice is objectively gender neutral. A small workforce may contain an important practical difference that does not meet a conventional significance threshold. Conversely, a very large dataset can make a small difference statistically detectable. Employers should therefore consider effect size, confidence intervals, worker-category evidence and compensation records alongside formal tests.
Regression Should Sit Beside the Required Reporting Metrics
Article 9 requires specific observed metrics, including gender pay gaps, median gaps, variable-pay gaps, quartile distributions and category-level information. Regression can help explain or investigate those results, but the adjusted estimate is not a substitute for the statutory measures. The employer should preserve both layers: the required reporting outputs and the separate analytical model. This makes it possible to explain why the organisation reported one observed figure while an internal adjusted analysis produced another result. It also avoids giving stakeholders the impression that statistical adjustment changed the legally required headline metric.
Make the Model Reproducible and Reviewable
A pay equity regression should have a written specification showing the dependent variable, group indicator, controls, coding rules, reference categories, exclusions, transformations and treatment of missing values. The organisation should retain the dataset version, code or software settings, diagnostics and outputs used for decision-making. Important alternative models should also be preserved where they influenced interpretation. This documentation allows another qualified reviewer to reproduce the result and understand why the model was chosen. It is especially valuable if the analysis later informs remediation, worker consultation, an investigation or litigation.
Frequently Asked Questions
Is regression analysis required by the EU Pay Transparency Directive?
No. The Directive specifies reporting and equal-pay obligations but does not prescribe regression as a mandatory universal employer method.
What does the sex coefficient mean in a pay equity regression?
Under the chosen model, it estimates the remaining association between sex and the pay outcome after the included variables are taken into account. Its meaning depends on the model specification and data quality.
Should a regression include every available HR variable?
No. Controls should be relevant, reliably measured and substantively defensible. Adding variables merely because they reduce the gap can produce misleading conclusions.
Related Guides
Official Sources
Requirements and practices differ by jurisdiction and organisation. Check current local law, official guidance and professional advice for a specific situation.