Can a university's student intake demographics predict whether it will fall into the bottom 25% for degree completion?
The Office for Students (OfS) regulates higher education in England. One of its core accountability metrics is completion rate — the share of students who finish their programme within the expected timeframe. Providers with persistently low completion face regulatory scrutiny and potential sanctions.
This project asks a pointed question: is low completion structurally predictable from who attends the institution? If the answer is yes, then completion gaps are largely baked into intake composition, and the OfS could use demographic profiles as early-warning signals when reviewing Access and Participation Plans.
The dataset covers 312 English HE providers — research-intensive universities, post-1992 institutions, FE colleges, and specialist providers — with outcome rates rolled up across four reporting years.
Binary classification. The continuous completion rate is split at the 25th percentile: providers below are labelled as at-risk (y=1, n=77), the rest as y=0 (n=229). That 25:75 imbalance is real and shapes every design decision.
Left: class balance (3:1 split). Right: providers from deprived areas (high IMD Q1) cluster in the low-completion zone, but the overlap is substantial — no single feature is enough.
Nineteen continuous percentages describing student composition:
Age pct_mature
Disability pct_disabled
Sex pct_female
Ethnicity pct_eth_asian pct_eth_black pct_eth_mixed pct_eth_other
Deprivation pct_imd_q1 pct_imd_q2 pct_imd_q3 pct_imd_q4
Participation pct_polar_q1 pct_polar_q2 pct_polar_q3 pct_polar_q4
pct_tundra_q1 pct_tundra_q2 pct_tundra_q3 pct_tundra_q4
Reference categories (white ethnicity, IMD/POLAR/TUNDRA Q5) are excluded to avoid multicollinearity.
Every model follows the same disciplined pipeline to prevent data leakage:
Median Imputation (training stats only)
|
SMOTE 1:1 (training folds only)
|
[StandardScaler] (LR, SVM only)
|
[SelectKBest] (LR only)
|
Classifier
Five classifiers, each testing a different inductive bias:
| Model | Why it's here |
|---|---|
| Logistic Regression | Interpretable linear baseline; coefficients show per-feature risk |
| Random Forest | Captures non-linear interactions; bagging stabilises small-sample variance |
| Gradient Boosting | Sequential error correction; typically lowest bias on tabular data |
| XGBoost | L1/L2 regularisation on leaf weights; built-in feature selection for collinear inputs |
| SVM (RBF) | Kernel-based boundary; sanity check against tree-based models |
Tuning: 5-fold stratified CV on 244 training samples, scored by macro F1 (treats both classes equally — accuracy would let a majority-class predictor score 74.8% while catching zero at-risk providers).
| Model | Test Accuracy | Test F1-macro | Test ROC-AUC |
|---|---|---|---|
| Random Forest | 0.710 | 0.659 | 0.776 |
| XGBoost | 0.726 | 0.634 | 0.794 |
| Gradient Boosting | 0.694 | 0.622 | 0.777 |
| SVM (RBF) | 0.645 | 0.594 | 0.734 |
| Logistic Regression | 0.645 | 0.583 | 0.686 |
Random Forest wins on F1 (the metric that matters for balanced class performance). XGBoost leads on AUC but is more conservative — it produces better probability rankings but misses more at-risk providers at the default threshold.
Tree-based models hit 94%+ training F1 but drop to ~0.65 on test — the gap is overfitting from only 244 training rows, not a methodology flaw. The CV-to-test gap is under 0.07 for all models, confirming that grid search regularisation is working.
In the OfS context, missing a struggling provider is worse than over-flagging one that turns out fine. A missed provider gets no intervention; a falsely flagged one just gets reviewed.
Lowering the classification threshold from 0.50 to 0.30 pushes minority-class recall from 0.62 to 0.94: the model catches 15 of 16 bottom-quartile providers in the test set, at the cost of 18 false positives (up from 12).
pct_mature dominates across all three models. Providers where over half the intake is aged 21+ (median 56.2% for bottom-quartile vs 27.2% for the rest) tend to be FE colleges and specialist providers offering part-time or distance learning. Mature students juggle work, family, and study simultaneously, driving higher structural dropout rates.
IMD deprivation ranks second. Providers with high pct_imd_q1 (most deprived 20%) face a compounding effect: mature students from deprived backgrounds experience a double disadvantage.
| Feature | Bottom Q median | Rest median | Gap |
|---|---|---|---|
pct_mature |
56.2% | 27.2% | +29.0 pp |
pct_imd_q1 |
30.2% | 20.7% | +9.5 pp |
pct_polar_q1 |
14.5% | 11.8% | +2.7 pp |
pct_disabled |
18.2% | 16.6% | +1.6 pp |
All figures are pre-rendered in figures/ and reproduced by the notebook. Here is a selection:
 Pearson correlation matrix — POLAR and TUNDRA overlap heavily (r=0.92) |
 Confusion matrices for all five models on the test set |
 Boxplots of key features split by target class |
 Missing-value profile with 10% threshold line |
UK-HE-Completion-Risk-Classifier/
|
|-- README.md
|-- LICENSE
|-- requirements.txt
|-- .gitignore
|
|-- data/
| +-- ofs_provider_outcomes.csv 312 providers, 33 columns
|
|-- notebooks/
| +-- completion_risk_analysis.ipynb Full pipeline: EDA -> training -> evaluation
|
+-- figures/
|-- fig_eda_overview.png
|-- fig_missing_values.png
|-- fig_correlation_heatmap.png
|-- fig_feature_distributions.png
|-- fig_lr_tuning.png
|-- fig_rf_tuning.png
|-- fig_gb_tuning.png
|-- fig_xgb_tuning.png
|-- fig_svm_tuning.png
|-- fig_model_comparison.png
|-- fig_generalisation_gap.png
|-- fig_threshold_tuning.png
|-- fig_confusion_matrices.png
|-- fig_roc_curves.png
|-- fig_feature_importances.png
+-- fig_per_class_metrics.png
git clone https://github.com/DataScienceVishal/UK-HE-Completion-Risk-Classifier.git
cd UK-HE-Completion-Risk-Classifier
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
jupyter lab notebooks/completion_risk_analysis.ipynbThe notebook reads the CSV via ../data/ofs_provider_outcomes.csv and writes figures to ../figures/. It runs top-to-bottom with no manual steps. All random seeds are pinned to 42.
The feature vector includes proxies for protected characteristics: ethnicity shares, deprivation indices, and mature-student rates. A model that flags institutions with high proportions of students from disadvantaged backgrounds as "at risk" could easily be misread as blaming those institutions. The real story is about systemic inequalities affecting their students. Results here are descriptive of structural patterns, not a basis for punishment or withdrawal of funding.
- Sample size is the bottleneck. 244 training rows with 19 features; tree models have capacity to spare but generalisation suffers.
- No provider-type variable. Russell Group, post-1992, and FE colleges have fundamentally different missions and resources, but the dataset treats them identically.
- The model cannot distinguish student-level from institutional causes. High
pct_maturemay reflect financial pressure on students, or it may reflect weaker part-time support structures. Disentangling these requires student-level data. - Gross enrolment ratios and aggregation across four years smooth out year-to-year dynamics, including COVID-era disruption.
Office for Students (OfS) Access and Participation Data Dashboard. https://www.officeforstudents.org.uk/data-and-analysis/access-and-participation-data-dashboard/
Released under the MIT License.