School Absence & Deprivation

The finding

The COVID attendance crisis hit every school equally, regardless of how deprived their neighbourhood was

Persistent absence, missing at least 10% of possible school sessions, has long been treated as a proxy for disadvantage. The assumption is that it clusters in deprived areas, and that area-level interventions can reach the pupils most at risk. This analysis, covering roughly 79,000 school-year observations across four academic years, tests that assumption directly.

Free School Meals eligibility is nineteen times more predictive of absence rates than the Index of Multiple Deprivation score of the area a school sits in. Area deprivation is real but it is a blunt instrument, and the model makes that gap visible.

The year a school year falls in is the second strongest predictor in the model. Not because school years inherently differ, but because COVID-19 caused a structural break in attendance norms that is still visible in 2023-24 data. When the SHAP dependence plots are split by deprivation level, the shock is flat across that spectrum. The pandemic hit every school at roughly the same magnitude.

39.7%

of model explanatory power comes from
Free School Meals eligibility alone

0.597

Test R² for Random Forest
vs 0.486 for OLS baseline

2.1%

of model explanatory power from
area-level IMD score

The model

What drives absence

Random Forest feature importances, trained on roughly 79,000 school-year observations split by school URN to prevent leakage.

Random Forest vs OLS baseline

Test set performance. RF trained with GridSearchCV (n_estimators=300, max_depth=10). OLS baseline uses same feature set.

Explainability — SHAP values

SHAP summary bar — Mean absolute SHAP values across the test set. Higher = more influence on the model's prediction.

SHAP beeswarm — Beeswarm plot showing direction and spread of each feature's effect. Red = high feature value, blue = low.

FSM dependence plot. Higher FSM rates consistently drive large positive SHAP contributions — more absence.

Year dependence coloured by IMD score. The post-COVID shift is uniform across deprivation levels — no gradient.

Phase dependence — Secondary schools (phase=2) receive consistently higher SHAP contributions than primary schools.

IMD dependence — IMD has a weak, noisy relationship with absence once FSM, year and phase are accounted for.

Key findings

What the model tells us

FSM eligibility (39.7% importance) is already an administrative flag held by every school and local authority. Using it directly to prioritise attendance outreach will reach higher-risk pupils more reliably than area-level IMD targeting.
SHAP values for 2022-23 and 2023-24 remain positive, meaning the pandemic's effect on attendance norms has not fully unwound four years later.
Phase is the third most important feature (20.9%). The absence gap between primary and secondary is structural and independent of deprivation. It cannot be attributed to higher FSM rates in secondary schools alone.
Because the COVID shock was deprivation-neutral, channelling recovery funds through deprivation criteria would have left many high-absence schools out of scope. A universal programme would have had broader reach.
The Random Forest outperforms OLS by 11 percentage points of R² (0.597 vs 0.486) because it captures year × phase and year × FSM interactions that a linear model misses.

Data & method

Feature importances in full

Feature	Importance	Description
percent_fsm	0.397	% pupils eligible for Free School Meals
year_numeric	0.298	Academic year encoded 0–5 (captures COVID break)
phase_numeric	0.209	1=Primary, 2=Secondary, 3=All-through
log_pupils	0.044	Log-transformed school roll
imd_score	0.021	IMD 2019 area deprivation score
region_London	0.013	London region dummy
All other regions + urban flag	0.018	—