Statistical analysis is most useful in pay equity work when the question goes beyond describing a gap and asks whether a pay difference remains after relevant factors are considered. Large datasets with reliable job, pay and worker-category information may support regression or other multivariable techniques. Smaller or simpler datasets may be better served by descriptive comparisons, mean and median analysis, category-level review and documented case-by-case investigation. Directive (EU) 2023/970 does not require employers to use regression as a universal method. Articles 9 and 10 establish reporting and category-level assessment requirements, so statistical models should support those duties rather than replace the required metrics or legal analysis.
Jurisdiction: European Union
Start With the Question, Not the Statistical Tool
A pay equity review should begin by defining the decision the analysis must support. If the question is simply whether women and men have different average or median pay levels, descriptive statistics may be sufficient. If the question is whether a difference remains after accounting for job, grade, location, experience or another relevant factor, a more advanced model may be useful. Choosing regression before defining the question can lead to unnecessary complexity and weak interpretation. The method should be proportionate to the issue being investigated and capable of being explained to the people who will rely on the result.
The Directive Requires Reporting Metrics, Not One Universal Model
Article 9 specifies the information that covered employers must report, including gender pay gaps, median gaps, variable-pay measures, quartile distributions and category-level gaps. Article 10 addresses joint pay assessment where defined conditions are met. Neither article requires employers to run regression analysis for every workforce. This distinction matters because a sophisticated statistical model cannot substitute for a required reported metric. Employers should first produce the mandated outputs using the applicable national methodology, then use statistical analysis as an additional investigative layer where it is useful and defensible.
Advanced Analysis Works Best When the Data Can Support It
Multivariable analysis depends on data quality. The dataset should have enough observations, meaningful variation in the variables being tested and reliable definitions for pay, job, working time and other explanatory factors. A model can become unstable when categories are extremely small, when variables are highly correlated or when important fields are missing for a large part of the workforce. Analysts should therefore review sample structure and data completeness before fitting a model. A clean, simpler analysis is usually more informative than an elaborate model built on weak or inconsistent inputs.
Descriptive Analysis Is Often the Right First Step
Mean, median, distributions, category-level gaps and cross-tabulations can reveal a large amount before regression is considered. These methods can identify where a gap is concentrated, whether one sex is overrepresented in higher-paid categories and whether variable pay or working-time patterns are contributing to the difference. Descriptive analysis also makes later modelling easier because it exposes data problems and unusual observations. Employers should not assume that advanced statistics are automatically more rigorous. The strongest method is the one that answers the question accurately and transparently with the available data.
Use Statistical Models to Test Defined Explanatory Factors
Regression can be useful when the employer wants to estimate the relationship between pay and several factors at the same time. The model can test whether a difference remains after accounting for selected characteristics such as job category, grade, location or experience. Those factors should be selected because they are relevant to the pay system, not merely because they improve the model fit. Analysts should also assess whether each factor is objectively defensible and whether its use could conceal a structural issue. Statistical adjustment is an analytical exercise and does not automatically establish that a legal pay difference is justified.
Document Why the Chosen Method Was Appropriate
A defensible methodology record should explain the question, population, measurement period, variables, exclusions, transformations, model specification and interpretation rules. It should also record why a simpler or more advanced method was chosen. This documentation matters when workers' representatives, management, advisers or authorities need to understand how the result was produced. It also supports repeatability when the analysis is rerun in a later reporting period. The objective is not to create a technical appendix for its own sake, but to make the analytical reasoning visible and reviewable.
Frequently Asked Questions
Does the EU Pay Transparency Directive require regression analysis?
No. The Directive specifies reporting and equal-pay requirements but does not mandate regression as a universal employer method. Regression can support investigation when the data and question justify it.
When is descriptive analysis enough for pay equity work?
Descriptive analysis can be sufficient when the objective is to identify or locate observed gaps, compare categories, review distributions or assess straightforward pay patterns without testing multiple explanatory factors simultaneously.
Is a more complex statistical model always better?
No. A complex model built on weak, sparse or inconsistent data can be less reliable than a simpler transparent analysis. Method complexity should match the question and the quality of the dataset.
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.