The Brief
Ahold Delhaize publishes an annual report that covers operations across 19 brands and roughly 384,000 associates worldwide. The 2025 report needed a rigorous, auditable gender pay parity analysis — not just a raw pay gap number, but a decomposition that separates structural factors (job level, tenure, geography) from the unexplained residual that regulators and investors actually care about. Alongside the pay analysis, the report required brand movement and job level movement analytics to give the board and external stakeholders a complete picture of internal mobility and workforce composition.
What Was Delivered
Four workstreams fed into the final report:
- Gender pay parity analysis — OLS regression models controlling for job level, tenure, brand, country, and function, followed by a Blinder-Oaxaca decomposition that isolates the explained and unexplained components of the pay gap. The headline result: a 13.64% unadjusted average gap narrows to 1.24% adjusted once structural factors are accounted for — down from 1.36% in the prior year.
- Brand movement analysis — tracking associate flows between Ahold Delhaize's portfolio of brands over the reporting period, showing where internal mobility is happening and where it isn't.
- Job level movement analysis — mapping vertical and lateral progression across the organisation's grading structure, giving visibility into promotion velocity and retention at each level.
- Data integration pipelines — the plumbing that made the above possible: ingesting, cleaning, and harmonising HR data from multiple source systems across brands and geographies into a single analytical layer.
The results appear on page 316 of the 2025 Annual Report, which explicitly references the Blinder-Oaxaca methodology.
How It Was Built
The core constraint was auditability. Every number in a public annual report is subject to external assurance, which means the analytical pipeline had to be deterministic, reproducible, and documented at a level where an auditor could trace any published figure back to source data and methodology.
Blinder-Oaxaca decomposition was chosen over simpler regression-based adjustments because it provides a clean two-part answer: how much of the raw gap is explained by observable differences in characteristics (the "endowments" component), and how much remains unexplained (the "coefficients" component). This is the standard the academic literature uses and what sophisticated stakeholders expect.
PySpark on Azure Databricks handled the scale. With hundreds of thousands of records across multiple brands, geographies, and compensation structures, the data preparation — currency normalisation, FTE adjustment, outlier treatment, and feature engineering — needed a distributed compute layer. The statistical modelling itself ran in pandas and statsmodels once the data was aggregated to the right grain.
Reproducibility was enforced through versioned notebooks, parameterised pipelines, and a clear separation between data preparation, modelling, and reporting layers. Any analyst can re-run the full pipeline against next year's data extract with a parameter change.
Why It Matters
Gender pay parity analysis is table stakes for a Fortune-500 annual report. What matters is whether the methodology is defensible. A simple median-of-medians comparison invites the question "but have you controlled for...?" — the Blinder-Oaxaca approach answers that question before it's asked.
The 1.24% adjusted figure is not just a number for the report. It gives the organisation a precise, decomposed view of where residual gaps exist, which factors drive them, and where targeted interventions would have the highest impact. That's the difference between compliance reporting and actionable analytics.