A data story · World Bank Group results, FY25

Who benefits? Health, jobs & gender

Can health investment create jobs, especially for women, and does labour mobility play a part? Before asking whether it works, ask what gets measured. Scroll through four chapters built on the World Bank Group's own results data.

0projects
0countries
0project-indicator records
0measure health workers
↓Begin ↧Download the full report (PDF)
Chapter 01

What gets measured?

Every project's objective and results indicators were read by a transparent keyword classifier. Each tag keeps the phrase that triggered it, and a hand-checked sample measures how often the tags are right.

1,131projects, one dot each
Explore the tagged projects

Chapter 02

Who is being reached?

People reached so far by Bank Group projects, and how many are women. Choose a result type, then tap a region to see it next to its development context.

Everyone reachedWomen

People reached to date, FY25 (30 June 2025). Sourced · WBG Scorecard

Chapter 03

When do results arrive?

For each result, compare the share of its target delivered with the share of its implementation period that has passed. A result is flagged behind if delivery trails time by more than 10 points. Then ask: can a model spot these early?

Share flagged behind, by project age

Years since Board approval · Derived from Scorecard project data

An early-warning model

AUC: the chance the model ranks a truly-behind result above one on track. 0.5 is a coin flip. Model estimate

What the model relies on

Drop in AUC when each feature is shuffled

Does it work equally well everywhere?

Out-of-fold AUC by subgroup; the faint line marks 0.5

Chapter 04

Are women counted, and reached as planned?

Where projects report both a total and a figure for women, compare women's planned share with their achieved share. First, though, check whether a gap could even appear.

1,734results reported for women and in total

Planned vs. achieved share of women

Each dot is one informative result. Dots start on the line (as planned), then move to what was achieved.

Not fragileFragile / conflict-affected (FCS)

Does fragility widen the gap?

Fragile-setting results compared with similar results elsewhere, using propensity-score matching

Making the comparison fair

Fragile-setting results differ from others before any comparison, for example more are in Africa and in low-income countries. Matching pairs each with a similar non-fragile result. Watch the imbalances shrink toward zero. The shaded band marks |SMD| < 0.1.

Chapter 05

How this was built, and its limits

Every number on this page is computed from public World Bank Group Scorecard exports by one Python script. Nothing is illustrative or simulated. The full methodology, robustness checks and annexes are in the written report (PDF, 22 pp.).

Data & provenance
  • Results and project records: World Bank Group Scorecard exports (scorecard.worldbank.org), FY25 cycle, results as of 30 June 2025. Covers health, nutrition & population services; WASH; gender-equality and economic-opportunity actions; financial services; and health-emergency capacity.
  • Context: Scorecard vision indicators for poverty at $3.00/day, food and nutrition insecurity, and basic hygiene. The latest year is shown for each region.
  • Cleaning: the export's “ACW” row is All countries, and the seven regions sum to it (checked in code). Only all-project rows are kept, because the disability-flag “Yes” rows are a subset. Results the export marks as double-counted are excluded, and exact duplicates across files are dropped.
Sourced vs. estimated
  • Sourced numbers come straight from the exports, or from simple arithmetic on them.
  • Model estimate numbers come from models fitted here. Each is shown with its validation or uncertainty.
  • Check reporting marks figures that look like reporting practice, not reality.
Limitations
  • Matching is not randomisation. Only observed differences are balanced, and unobserved ones (capacity, security) could still bias the comparison. Balance on a few covariates stays slightly above 0.1. The result is a matched association, not a causal effect.
  • Small informative sample. Only results where women are counted directly can show a gap, so the confidence interval is wide.
  • “Behind” is a straight-line benchmark. Results are back-loaded, so the flag is a screening signal, not a verdict.
  • The classifier is rule-based, and its recall is unmeasured. Precision was checked on small samples (5–12 per theme). Projects described in unexpected words will be missed.
  • Coverage. The Scorecard reports corporate results indicators, not every indicator a project tracks.
Privacy, fairness & reproducibility
  • Privacy: the analysis uses only public, aggregated, project-level data. No individual or household records.
  • Fairness: model performance is reported by region, income group and FCS status, and its weakest subgroup is named.
  • Validation: the classifier's review samples, labels and failed first version are saved in the repository. The review is a draft pending the author's confirmation.
  • Reproducibility: analysis/who-benefits/pipeline.py rebuilds every number with fixed seeds. Code & data on GitHub →
Related research
This connects to two strands of the author's work. The first is ongoing research on Global Skill Partnerships, skills and international labour mobility, which evaluates a skills-mobility pilot with propensity-score matching. The second is Bleeding at the Margins (Journal of Development Studies, 2023) on menstrual health among SC/ST women in India. Chapter 01 shows how rarely either topic appears among measured results.