# The Education Escalator: Methods Edition: published-evidence-v1, 12 September 2026. ## Evidence And Population This edition analyzes published aggregate estimates in OECD PISA 2025 Results, Volume I. It does not claim independent student-microdata estimation. The workbooks use the June 2026 analytical database; the September 2026 public-use release is a different vintage. Four original workbooks are hash-bound in `evidence.json`, with URLs and retrieval dates. `exact-values.csv` retains every displayed source cell, its standard error, and coverage marker. PISA covers enrolled students aged 15 years 3 months to 16 years 2 months who have completed at least six years of schooling. The German sample contains 6,659 students in 246 schools, representing about 717,900 fifteen-year-olds and roughly 93% of the age population. Those population totals come from the OECD Germany country note, not a recalculated sum of student weights. Nineteen selected national or territorial education systems provide comparison cases. Chinese Taipei, Hong Kong and Macao retain their official names; none is treated as national China. The selected set is not the OECD membership set. ## Published Estimates And Intervals Every plotted interval is a pointwise estimate plus or minus 1.96 times its published standard error, following Annex A3. These are not simultaneous bands, student-level prediction intervals, or intervals that incorporate non-response bias. A missing point is unavailable, never zero. A published zero with a suppressed standard error remains a point with unavailable uncertainty and cannot enter a newly calculated inferential contrast. The OECD reference is the published arithmetic country-average estimate. Its country availability can differ between indicators. Germany is included, so an independent-samples test of Germany minus the OECD average is not performed. For Germany minus Estonia, independent national samples give SE(difference) = sqrt(SE(Germany)^2 + SE(Estonia)^2). Two-sided normal p-values use that standard error. A Holm correction is applied to the three focal science contrasts: mean score, top-minus-bottom national ESCS-quartile gap, and academic resilience. This is a declared analysis family, not external preregistration; the other comparisons remain exploratory. ## Distributions, Background And Trends Science is the main domain, with reading and mathematics available in the distribution and opportunity comparisons. Quantiles are the five published national achievement percentiles, not a reconstructed density. Socioeconomic group means are discrete published bands, not a fitted student regression. National ESCS quarters refer to relative position inside each country. The international bands share cutoffs but contain different national shares. Germany's ESCS index is missing for 16.97943% of students. Observed-background group results and the all-student means therefore have different denominators. Trend charts use actual dates, including the four-year 2018-2022 interval. Assessment cycles are repeated cross-sections, not longitudinal pupils. Claims about change use the separately published change estimates and their standard errors, which include linking uncertainty, rather than comparing two unrelated confidence intervals. Academic resilience is the proportion of students in the bottom national ESCS quarter reaching the top national achievement quarter. It is not adult mobility or a diagnosis of individual talent. Low performance means below the official Level 2 threshold in the stated domain and has a different denominator and substantive meaning. ## Schools And AI Between-school variation is 100 minus the published inclusion index. It is not reconstructed from displayed variance components that use different scaling or rounding. School-model standard errors are model-based, as specified in Annex A3, not design-replication standard errors. The academic index uses the modal ISCED restriction; the social index shown does not. Neither is a causal school or peer effect. Principal responses about staff shortages are weighted by represented students. They are separate from student-reported teacher interest. Response categories sum to 100% within a question, but distinct questions are not additive causes. AI frequency categories retain the source questionnaire definitions. Different purposes overlap. The OECD's purpose coefficients compare weekly-or-daily use with less frequent use, before or after adjusting for student and school ESCS. Separate coefficient intervals do not test the difference between adjusted and unadjusted coefficients. Reverse causality, prior achievement, motivation and other omitted variables prevent a treatment-effect interpretation. No new AI-by-background interaction model was estimated. ## Four-Band Composition Accounting Let p denote international-ESCS-band shares and m the mean achievement in each band. The observed-background mean is sum(p*m). The Germany-minus-peer gap can be written as sum((pG-pE)*mE) + sum(pG*(mG-mE)), or using the German reference as sum((pG-pE)*mG) + sum(pE*(mG-mE)). The symmetric view averages the two component pairs. Each identity is checked to numerical tolerance. For Germany-Estonia the observed-background mean gap is -29.0988 points; it is not the -41.5956 all-student mean gap. The symmetric composition and within-band components are -13.2022 and -15.8966 points. Broad bands do not eliminate within-band composition. Joint covariance between shares and conditional means is unavailable, so no component confidence interval is fabricated. This is not an individual-level Oaxaca model and the remainder is not school-system quality. ## Benchmark Counts The difference between German and comparator below-Level-2 shares is multiplied by the rounded German covered population, 717,900, held fixed. The resulting count describes the size of a reference gap, not an intervention forecast. The Germany-Estonia science difference is about 121,326 students, with a pointwise interval of about 101,839 to 140,813 under the stated assumptions. Negative comparisons remain negative. No earnings, GDP, or years-of-schooling conversion is made. ## Comparability And Unestimated Work Canada and the Netherlands carry OECD sampling warnings. Canada's results may be upward biased by non-response. More than 40% of responding US students did not receive the student questionnaire; US context estimates are flagged and excluded from the default comparison field unless explicitly selected. The supplied microdata programme remains unexecuted: within/between school models, migration adjustment, AI interactions, quantile regressions, school- blocked cross-fitting and a full Oaxaca analysis require registered student and school data plus codebook validation. Synthetic test fixtures are not empirical evidence. No raw student or school records are published in this edition. The source package quoted an 84-point OECD science gap from Education GPS. The frozen workbook reports 84.72297, which this edition displays as 84.7. Neither the source package nor the official workbook is silently overwritten. ## Reproduction From the repository root, with its locked Node dependencies and Python with openpyxl installed: node scripts/pisa-education/acquire.mjs --fetch python scripts/pisa-education/extract.py python -m unittest discover -s scripts/pisa-education -p test_extract.py python scripts/pisa-education/extract.py --check node scripts/pisa-education/design.mjs node scripts/pisa-education/assets.mjs Acquisition refuses to overwrite a changed frozen workbook. Source workbooks remain outside the deployable tree. The source-cell CSV and generated JSON are the exact-value fallbacks; source URLs allow original-table inspection. ## Sources - https://stat.link/mrq53f (performance) - https://stat.link/k68msa (socioeconomic differences) - https://stat.link/xmra2f (school life and AI) - https://stat.link/5eaxdr (attitudes; acquired for reference, not an additional causal model) - https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en/full-report/reader-s-guide_0364240f.html - https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en/full-report/technical-notes-on-analyses-in-this-volume_46a7001c.html Programme for International Student Assessment (PISA), Organisation for Economic Co-operation and Development (OECD), Paris. This is an adaptation of an original work by the OECD. The opinions expressed and arguments employed in this adaptation should not be reported as representing the official views of the OECD or of its Member countries. ## Opportunity, skill formation and AI: interpretation added 13 September 2026 The author's personal opening is recollection supplied by Michael Schymura. It motivates the question and is not part of the PISA sample or an estimate of mobility. Rawls's fair equality of opportunity and difference principle, and Roemer's circumstance/responsibility framework, are normative interpretations. The reported ESCS comparisons are not a completed Roemer decomposition. For children, responsibility requires particular care: Roemer and Trannoy (2016, p. 1308) treat nature and nurture as circumstances before an appropriate age of consent. Neither achievement ranks nor unexplained score variation are treated here as measures of effort or merit. Circumstance-group distributional analysis and mechanism models remain proposed research, not new results. The discussion acknowledges Agarwala and Spiewak's DIE ZEIT 39/2026 essay, using the print title supplied by the author and linking to its online version. Its prose and illustration are not reproduced. PISA achievement, IQ tests and EEG measures are not interchangeable. No smartphone or generative-AI cause is assigned to the cross-country trend from chronology alone. The Bastani et al. (2025) school mathematics experiment and Kestin et al. (2025) university physics experiment are external mechanism evidence, not observations from this PISA sample. Bastani's 48% assisted-practice gain and 17% unaided-exam penalty are relative experimental grade changes. The guarded tutor's unaided exam result was statistically indistinguishable from control; it is not reported as a proven gain or an exact zero. Kestin concerns immediate learning across two lessons in one university course. No long-term or Germany-wide causal effect is inferred. Socioeconomic differences in the consequences of AI use are competing hypotheses, not effects established by these experiments or the published OECD associations. Skill formation and talent allocation supply economic mechanisms. No foreign growth estimate, IQ change, or experimental grade effect is converted into German PISA points, earnings, years of schooling or GDP. The added references are kept under `interpretationReferences` in sources.json, separate from the four frozen statistical workbooks. The publication receipt binds the revised article to the canonical article.json, and the deterministic archive includes the same interpretation and references as the page.