Plan, apply and roll back changes to a production ontology, with a blast radius report and a proof an auditor can re-check.
Technical note: intergenerational educational and skills mobility in England (PIAAC)
This note documents the data, variables, statistical methods, validation and limitations behind the case study. It is written for analysts and official-statistics methodologists who want to check or reuse the work.
1. Data
| Item | Detail |
|---|---|
| Source | OECD Programme for the International Assessment of Adult Competencies (PIAAC), Survey of Adult Skills, public-use files |
| Files | prggbrp1.csv (Cycle 1, 2012, 8,892 respondents) and prggbrp2.csv (Cycle 2, 2023, 4,941 respondents) |
| Host | OECD public file store, https://webfs.oecd.org/piaac/ (open directory; no registration or licence) |
| Value labels | Taken from the official SPSS public-use files (prggbrp1.sav, PRGGBRP2.sav) on the same store |
| Access date | 17 June 2026; URLs, byte sizes and MD5 checksums recorded in data/raw/MANIFEST.csv |
| Population | Adults aged 16 to 65 |
The UK participated in PIAAC Cycle 1 as England and Northern Ireland and in Cycle 2 as England. For every cross-cycle comparison the 2012 sample is restricted to England (OECD TL2 regions UKC to UKK, excluding Northern Ireland, UKN) so the geography matches. Northern Ireland was heavily oversampled in 2012; the final weights correct for this, and restricting to England removes it from the comparison entirely.
2. Variables and harmonisation
The two cycles use different variable names and, in places, different codings. All analysis variables are mapped to a common scheme, recorded in outputs/tables/harmonisation_map.csv and published as a coded SKOS scheme in ontology/.
| Concept | 2012 variable | 2023 variable | Coding used |
|---|---|---|---|
| Origin (parental education) | PARED |
PAREDC2 |
3 bands: Low (neither parent above lower secondary), Medium (at least one upper secondary or post-secondary non-tertiary), High (at least one tertiary). Codes 6/7/8/9 = missing |
| Own highest qualification | EDCAT6 |
EDCAT6_TC1 |
Collapsed to the same 3 bands: Low = code 1; Medium = 2,3; High = 4,5,6,7. (2012 uses code 7 for grouped tertiary; 2023 uses 5 and 6) |
| Numeracy | PVNUM1-PVNUM10 |
same | 10 plausible values, 0-500 |
| Literacy | PVLIT1-PVLIT10 |
same | 10 plausible values, 0-500 |
| Occupation (1-digit) | ISCO1C |
ISCO1C |
ISCO-08 major group; high-status = groups 1-3 (managers, professionals, technicians) |
| Occupational skill | ISCOSKIL4 |
ISCOSKIL4 |
4 bands, recoded so higher = more skilled |
| Hourly earnings | EARNHRDCL |
EARNHRDCLC2 |
National decile, 1 = lowest to 10 = highest |
| Sex | GENDER_R |
GENDER_R |
Male, Female |
| Age | AGEG10LFS |
AGEG10LFS |
10-year bands |
| Region | REG_TL2 |
REG_TL2 |
OECD TL2 |
| Final weight | SPFWT0 |
SPFWT0 |
|
| Replicate weights | SPFWT1-SPFWT80 |
same | 80 replicates |
The _TC1 suffix on the 2023 education variable denotes the OECD's own trend-coding to Cycle 1, so it is the version comparable with the 2012 EDCAT6.
An empirical correction worth recording. The SPSS value labels for the earnings-decile variable are internally inconsistent (they label code 2 as the "9th decile", code 3 as the "8th", and so on, which would make the scale non-monotonic). We did not trust the labels. We validated the variable against the continuous earnings measure available in 2012 (EARNHRBONUSPPP): mean continuous earnings rise monotonically from 7.6 PPP-USD per hour at decile 1 to 54.0 at decile 10, with a Spearman correlation of 0.99. The numeric values are therefore standard decile ranks (1 = lowest, 10 = highest) and the mid-range labels are a labelling error in the public file. We use the numeric values.
3. Statistical methods
Plausible values. Proficiency in PIAAC is reported as ten plausible values, not a single score. Each analysis is run separately on each of the ten values and the results are combined by Rubin's rules: the point estimate is the mean of the ten, and the total variance is the mean sampling variance plus (1 + 1/10) times the between-value variance. Averaging the ten plausible values into one score (a common error) understates uncertainty and is not done anywhere here.
Replicate weights and the cycle-specific variance method. Point estimates use the final weight SPFWT0. Sampling variances use the 80 replicate weights. The replication method differs by cycle, and using the wrong one inflates standard errors by a factor of roughly sqrt(20):
- Cycle 1 (2012): jackknife. Variance is the simple sum of squared replicate deviations,
Var = sum_r (theta_r - theta_0)^2. Implemented withsurvey::svrepdesign(type = "other", scale = 1, rscales = 1, mse = TRUE). - Cycle 2 (2023): Fay's balanced repeated replication, Fay factor 0.5. Variance is
1 / (R(1 - 0.5)^2) * sum_r (theta_r - theta_0)^2 = (1/20) * sum_r (...). Implemented withsurvey::svrepdesign(type = "Fay", rho = 0.5, mse = TRUE).
We confirmed the method empirically. Computing the 2023 England mean numeracy with the jackknife formula gives a standard error of 7.25, which is implausible for a sample of about 4,900; the Fay method gives 1.65. The ratio is sqrt(20), exactly the Fay scaling. The Fay estimate matches the published OECD standard error.
Models. All regression models are survey-weighted generalised linear models (survey::svyglm) fitted on the replicate-weight design, run once per plausible value and combined across the ten with mitools::MIcombine, so that both sampling variance and between-value variance enter the standard errors.
- Origin gradient and gap: mean numeracy by origin band; and
numeracy ~ origin + sex + age(total gap) versusnumeracy ~ origin + own_qualification + sex + age(gap net of qualification). - Within-qualification gap:
numeracy ~ origin + sex + agefitted separately within each own-qualification band. - Skills beyond qualifications:
outcome ~ qualification + age + sexversus the same plus standardised numeracy, for two outcomes (earnings decile, linear; professional/managerial occupation, quasi-binomial). Numeracy is standardised to mean 0, SD 1 using the weighted pooled plausible-value distribution, so its coefficient reads per standard deviation. Variance explained is the weighted R-squared (linear) or McFadden pseudo-R-squared (binomial), averaged across the ten plausible values.
4. Validation against published figures
| Quantity | This pipeline | Published OECD |
|---|---|---|
| England mean numeracy, 2023 | 268.8 (SE 1.65) | 268 |
| England mean numeracy, 2012 | 261.8 (SE 1.10) | ~262 |
| Direction of change 2012 to 2023 | significant rise | significant rise |
The pipeline reproduces the published England toplines, which gives confidence that the weighting, plausible-value and variance handling are correct.
5. Headline estimates with confidence intervals
| Estimate (England) | 2012 | 2023 |
|---|---|---|
| Numeracy, Low / Medium / High origin | 239 / 271 / 289 | 242 / 279 / 293 |
| Tertiary attainment, Low / High origin | 23% / 62% | 32% / 60% |
| Origin gap (High-Low), total | 51.9 [46.7, 57.1] | 47.3 [36.2, 58.3] |
| Origin gap (High-Low), net of own qualification | 34.5 [28.8, 40.1] | 32.4 [21.5, 43.2] |
| Within upper-secondary group, High-Low gap | 38.3 [29.2, 47.4] | 31.3 [9.9, 52.7] |
| Within tertiary group, High-Low gap | 37.0 [28.1, 45.9] | 27.3 [14.6, 40.1] |
| Numeracy coefficient (per SD) on earnings decile | 0.86 [0.73, 0.99] | not estimated (earnings suppressed) |
| Variance explained in earnings, quals only vs + numeracy | 30.5% vs 36.7% | not estimated |
Square brackets are 95% confidence intervals.
6. Limitations
- Descriptive, not causal. These are associations. Parental education is a proxy for a bundle of family advantages; the design cannot separate them or establish causation.
- Cross-sectional. PIAAC is not a panel. "Origin to destination" is reconstructed within a single survey from retrospective parental-education reporting, not followed over time.
- Reverse and contemporaneous influences. Skills and qualifications are measured in adulthood and partly co-determined; the nested-model results describe shared and unique predictive content, not a causal decomposition.
- Public-file masking in 2023. Continuous earnings and four-digit occupation are suppressed in the 2023 public-use file. Cross-year analyses therefore use earnings deciles and one-digit occupation. The continuous-earnings nested model is 2012 only.
- Missing origin data. Parental education is unknown (don't know, refused, not stated) for about 17% of respondents, who are excluded from origin analyses. If non-response is related to origin, the gradient could be mildly under- or over-stated.
- Three-band collapse. Banding parental and own education to three levels aids comparability but discards within-band detail, in particular differences within tertiary education.
7. Reproduce
# from the case-study directory
Rscript R/01_download.R # fetch the open PIAAC files (idempotent)
Rscript R/02_harmonise.R # build the harmonised analysis table
Rscript R/03_analyse.R # estimates, with PVs + replicate weights
Rscript R/04_visualise.R # figures + per-figure CSVs
Rscript ontology/build_scheme.R # generate the SKOS variable scheme
.venv/bin/python ontology/validate_scheme.py # validate via the Oxigraph engine
A clean run regenerates every number in the case study and this note, and every figure, from the raw open files.