Gender gaps in student performance,
seen in context

Of the 81 matched economies with a WEF education score, six have a published score below 0.97: Albania (0.955), Cambodia (0.953), Kenya (0.847), Morocco (0.960), Rwanda (0.960) and Zambia (0.930). The other scores cluster near 1.000. The comparison below shows how much the coefficients change when selected observations are omitted.

These are post-hoc sensitivity checks, not recommended exclusions. The full sample remains the reference. Removing the six economies narrows the range of education scores and changes the population being compared. Dependence on these observations does not establish a data error or a causal explanation.

This table always uses signed gaps and the full matched samples. It is independent of the main chart controls, filters and manual omissions. No new p-values or confidence intervals are assigned to these selected omissions. Δr is the comparison coefficient minus the full-sample coefficient.

Fixed full matched sample for the selected indicator and GDP scale; signed gaps. The difference is |r(P90)| − |r(average)|. Its 95% interval comes from resampling the same economies jointly. Positive values favour a stronger P90 association; negative values favour the average. These exploratory difference intervals are not adjusted for multiple comparisons.

Sampling uncertainty matters. In this workbook, P90 gap estimates generally have larger standard errors than mean gaps. The intervals below resample economies and do not propagate the uncertainty of the PISA scores within each economy. Under a classical independent-error model, measurement noise can weaken correlations, but that does not establish the direction or size of bias in these comparisons. A stronger observed P90 association cannot therefore be described as necessarily understated. Mean and P90 differences alone also do not identify score variance or its causes.

Fixed full matched samples for the selected indicator and GDP scale; signed gaps. Each row compares P10 with the stated reference measure on the same economies. The difference is |r(P10)| − |r(reference)|: positive values favour a stronger P10 association, negative values favour the reference. Its 95% interval jointly resamples the same economies 20,000 times. These exploratory intervals are not adjusted for multiple comparisons and do not propagate within-economy PISA sampling uncertainty. They do not replace the original average-versus-P90 comparison above.

Original average/P90 results: this retained section covers the original eight tests per indicator/scale, 40 WEF tests and 128 tests overall. Its coefficients, intervals and corrections are unchanged. The additional table below reports expanded corrections that also include P10.

These reference results use all valid matches and signed gaps, regardless of current filters or omissions. The table reports Holm correction across the eight mean/P90 tests for the selected indicator and scale, and across all original indicator/scale tests. WEF also has a correction across its available comparisons. These are exploratory analyses, not preregistered confirmatory tests. The different testing families can produce different significance labels.

Read the interval and the test together. The p-values use the conventional Pearson correlation test; the intervals use percentile resampling of economies. These are different procedures with different assumptions, so they need not give the same answer. Highlighted rows flag an eight-test Holm p-value below 0.05 alongside an interval containing zero. Treat such results as sensitive to the inference method. A small adjusted p-value does not address influential observations, sampling selection or measurement error. Pearson-test assumptions.

Holm correction does not require independence between tests, provided the individual p-values are valid. The 128-test correction counts all original specifications; closely related GDP versions make this a conservative exploratory overview. No specification has been retrospectively designated as preregistered or primary. The unadjusted bootstrap intervals and the adjusted p-values do not control the same error rate. Multiple-comparison documentation.

These retained full-sample signed-gap results cover the original indicator snapshots: 12 tests per indicator/scale, 60 WEF 2025 tests, and 192 tests overall. They exclude the new PISA x-axis benchmarks. Their coefficients, intervals and corrections remain unchanged. The next table adds the PISA benchmarks and a separate 240-test correction. A highlighted interval includes zero despite a 12-test adjusted p-value below 0.05. These are exploratory testing families; intervals remain unadjusted.

Fixed full matched samples and signed gaps; current filters and omissions do not alter this table. Holm corrections cover 12 tests per indicator/scale, 60 WEF 2025 tests, 48 PISA-benchmark tests (four x-axis benchmarks × four gap domains × three attainment measures), and 240 tests across all indicators/scales. Earlier coefficients and bootstrap intervals are retained exactly. Only the new PISA associations and the 240-test correction are additions. All are exploratory; the 95% bootstrap intervals remain unadjusted and do not propagate survey uncertainty.

GDP is nominal output per resident in current US dollars. The World Bank and IMF provide 2025 values; the archived UN series ends in 2024. All three 2024 versions are included so source differences can be examined on the same GDP year. No source is filled using another, and no 2024 value is labelled 2025.

Where source values diverge

Materiality here means a symmetric percentage difference above 10%: 100 × |A − B| / ((A + B) / 2). This is a descriptive threshold, not a significance test. Country-level differences may reflect revisions, estimation, population denominators or conversion methods; no specific cause is assigned here.

Missing values are left missing. The CSV preserves source years, status, full-precision scores, signed gaps and any transformed plotting values.

Updated 14 September 2026 following a separate validation report. This revision adds the WEF education sensitivity comparison, flags disagreement between inference methods, clarifies coverage, measurement uncertainty and source periods, and fixes the slider label markup. HCI remains the default. Country values and the original full-sample coefficients, bootstrap intervals and multiple-comparison corrections are unchanged.

As part of the follow-up, 1,036 numeric P90 score/gap/standard-error cells and 1,036 corresponding mean cells were compared with the supplied workbook; all matched exactly. This verifies transcription against that file, not independent authentication of its origin or an audit of the survey. The additional review and automated checks do not constitute peer review or remove the disclosure at the top of this page.

This version adds Japanese alongside English in the same file. The language switch changes the presentation only: data, numerical precision, statistical calculations, chart geometry and machine-readable downloads are shared. Your current controls, country selection and omissions are preserved when you switch. A fresh visit starts in English; a shared link can specify either language.

The separate bilingual support bundle contains the English v1.1 input, translation dictionaries, presentation code and parity checks. Run bash run_build.sh to rebuild this single HTML file. The original analysis and v1.1 revision evidence remain available through the existing downloads. Translation and automated checks do not replace specialist review.

Added 14 September 2026. P10 is now available in the chart selector, country details, correlation matrix, GDP source comparisons, WEF education sensitivity table and plotted CSV. The original average/P90 data and statistical metadata are retained unchanged. New P10 data, paired comparisons and expanded corrections can be downloaded separately. The P10 support bundle includes the frozen input HTML and workbook, extraction and analysis scripts, a separate verifier and rebuild instructions. Automated verification checked 1,554 P10 score/gap/SE cells against a second workbook reader, all 64 new full-sample correlations and the paired bootstrap intervals. This does not constitute independent human or specialist review, and the disclosure at the top still applies.