Where does the data come from?
| Coverage | Source | Used for |
|---|---|---|
| KS4 (GCSE), 2023/24 | DfE Key stage 4 performance, GCSE subject transition matrices | GCSE outcomes by earlier KS2 attainment in England’s state-funded schools; 35 subject categories |
| KS5, 2024/25 and 2025/26 | DfE 16 to 19 performance: 2024/25 and 2025/26, transition matrices | Outcomes by earlier GCSE attainment, with separate qualification types, sizes and grade scales |
| GCSE and A level, 2019, 2023 and 2024 | Brylka & He’s Ofqual inter-subject comparability study · PDF | Published General Ability difficulty ranks |
| IB Diploma Programme, May 2019 | Badham, Meadows & Baird (2025), construct comparability study · PDF | Published General Ability thresholds, Table 4, Dataset 1 |
The DfE matrices are aggregate subject-entry counts, not linked individual records. A student can appear in several subjects. The Ofqual and IB studies use separate datasets and models; we reproduce their published results rather than fit those models to the DfE matrices. Subject coverage differs between sources.
Data, notation and grade scales
Counts and denominators
For subject
For GCSE transitions, X outcomes remain in percentage denominators but are excluded from numeric summaries. Unknown prior attainment is shown separately and excluded from the attainment-adjusted analyses. Each page specifies its usable bands and exclusions.
Grade coding
Grades use consecutive steps from U = 0 up the course’s own ordered grading scale. Combined Science (GCSE) uses the average of its two grades. Equal grade steps are a calculation convention. Means may fall between grade labels; these scales do not establish equivalence between qualifications.
Transition matrices and summaries
Exact-grade percentages show the share at each grade; cumulative percentages add that grade and all higher grades. The mean weights numeric grade points by entry counts. The median is the first grade reaching 50% when counting from the highest grade down, returning the higher grade at an exact tie.
Any display pooling sums counts before recalculating percentages and summaries. It does not change the source bands used for analysis. Explore GCSE Subject transitions or KS5 Subject transitions.
Difficulty from prior attainment
We adapt the prior-attainment weighted mean grade approach described by He and Cadwallader (2023, §2.2). Every eligible subject receives the same mix of earlier attainment: KS2 for GCSEs, and GCSE attainment for KS5. Our implementation uses DfE’s published bands and pooled entry weights, rather than the study’s ten equal-population bands.
Calculating the weighted mean grade
Multiply each subject’s mean grade within a band,
Here
This adjustment accounts for differences between attainment bands, while differences within bands and other entrant characteristics remain. Changing the comparison group can change weights and ranks. The GCSE comparison retains all 11 published KS2 bands; KS5 uses six bands: <4, 4–<5, 5–<6, 6–<7, 7–<8 and 8+. Subjects need usable results in every band. Size filters change eligibility and weights; search keeps them fixed. Each page shows its reference weights and excluded subjects.
Correlation, association and spread
These measures use observed grade counts and each subject’s actual intake, rather than the common WMG weights.
- Pearson correlation (r): the direction and strength of a straight-line relationship between coded prior-attainment bands and grade points, from −1 to 1. Band coding cannot recover variation in exact prior scores.
- Association (η²): the proportion of grade variance associated with differences between band means, allowing patterns beyond a straight line.
- Within-band SD: how much grades vary among entrants in the same attainment band, calculated from the count-weighted within-band variances. Normalised SD expresses this relative to the full grading-scale range.
A required variance of zero leaves the corresponding statistic undefined. Correlation does not establish causation or individual prediction accuracy; within-band spread is not a tested prediction error.
For count-weighted prior-attainment scores
Total variance = between-band + within-band variance. Within-band variance averages each band’s grade variance using its share of usable entries; normalised SD divides by the full grade-point range. These calculations use population denominators.
Closed prior-attainment bands use their midpoints. KS5 codes Below 1 as 0.5 and 9-or-above as 9.0. GCSE correlation excludes the open-ended “Less than 80” band, No prior score and X outcomes. Neither analysis includes unknown attainment or aggregate totals.
General Ability
Rasch models estimate a shared academic ability from the subjects a student took, alongside each subject’s grade-specific difficulty. Higher estimated ability increases the likelihood of higher grades; a harder subject requires greater ability for comparable outcomes. This assumes that different subjects can be linked through a common ability, with limits caused by selection, subject-specific strengths and model fit.
For Ofqual, difficulty at grade
The IB page uses published Rasch–Thurstone thresholds in logits; its overall measure averages grades 5, 6 and 7. Each study has its own population and scale. Ranks cannot be compared across qualifications, and small rank differences alone do not establish meaningful differences or grading unfairness.
References
-
Brylka, A. & He, Q. (2025). Inter-subject comparability in GCSEs and A levels in summer 2024. Ofqual. GCSE ranks: Figure 4 and Table A2; A level ranks: Figure 11 and Table A4.
-
He, Q. & Cadwallader, S. (2023). Inter-subject comparability in GCSEs and A levels in summer 2022: methods. Prior-attainment weighting and Rasch comparisons.
-
Badham, L., Meadows, M. & Baird, J.-A. (2025). Construct comparability and the limits of post hoc modeling: insights from International Baccalaureate multi-language assessments. Table 4, Dataset 1.
Ofqual source material © Crown copyright 2025, licensed under the Open Government Licence v3.0.
The IB study is © 2025 its authors and licensed under CC BY. Data extracted; layout and rankings adapted.