Working paper no. 1 · Preregistered 2026-06-11 · N = 150 local authorities

Refused, Not Appealed

The SEND appeals that never happen.

Tribunals find for families in roughly 99% of decided SEND appeals. Yet when English local authorities refuse to assess a child for an Education, Health and Care plan, 88.6% of those refusals are never contested at tribunal — and the families least likely to fight live in the most deprived places.

291,796Assessment requests, 2023–24

71,836Refused: decision not to assess

8,164Refusals appealed to tribunal

99%Of decided appeals find for families

Figure 1 · The national funnel, pooled 2023–24

Most refusals are never formally contested

Assessment requests291,796

Refused — decision not to assess71,836

Taken to mediation18,093

Appealed to tribunal8,164

Only 11.4% of refusals reach tribunal, in a system where the tribunal sides with families ~99% of the time it decides. The binding constraint on justice here is not winning the appeal — it is filing it. Across local authorities the dispute rate runs from 2.7% at the 10th percentile to 20.6% at the 90th: a seven-fold gap between the places where refusals are routinely tested and the places where they are quietly absorbed.

Figure 2 · Dispute-after-refusal rate by deprivation quintile

Refusals are contested half as often where deprivation is highest

Least deprived

13.9%

2nd

14.2%

3rd

8.4%

4th

10.9%

Most deprived

5.8%

The least-deprived fifth of local authorities contests 13.9% of refusals at tribunal; the most-deprived fifth contests 5.8%. In the preregistered Poisson model, one standard deviation of deprivation predicts a 24% lower dispute-after-refusal rate (β = −0.278, one-sided p < .0001) — a gradient that survives controls for need mix, the authority's own refusal rate, scale, and region fixed effects. Had every quintile contested at the least-deprived rate, there would have been roughly 1,956 additional appeals in 2023–24 — appeals that, on national success rates, would overwhelmingly have been won.

The sharpest finding · Channel substitution

Deprived areas don't challenge less. They're confined to the channel without teeth.

Mediation (free, informal)

+0.16

per SD of deprivation · rises with deprivation

Tribunal (binding, costly)

−0.28

per SD of deprivation · falls with deprivation

All formal challenge

+0.04

per SD of deprivation · flat — no gradient

Case merit affects mediation and tribunal symmetrically — a weak case is weak in both rooms. Costs do not. Families in deprived areas use free mediation more than anyone else, and the heavyweight tribunal route far less; total willingness to challenge is the same everywhere (interaction β = −0.43, p < .0001). That is the signature of a resource barrier, not weaker cases. Refusal rates themselves show no deprivation gradient at all.

Table 1 · Preregistered confirmatory models (coefficient on deprivation, z-scored)
Model	β	Robust SE	95% CI	p	N
M1Poisson, offset log(refusals)	−0.278	0.053	[−0.382, −0.175]	<.0001	150
M2+ need, refusal mix, scale	−0.277	0.071	[−0.417, −0.137]	.0001	150
M2b+ region fixed effects	−0.187	0.083	[−0.348, −0.025]	.024	150
M3Refusal-rate gradient (H3)	+0.104	0.094	[−0.080, +0.288]	.266	150
H4IMD × tribunal (vs mediation)	−0.434	0.072	[−0.576, −0.292]	<.0001	300
H5Refusal rate → EHE rate (OLS)	+0.193	0.284	[−0.364, +0.751]	.497	142

Table 2 · Preregistered robustness battery
Model	β	Robust SE	95% CI	p	N
R1WLS winsorized rate	−0.028	0.005	[−0.039, −0.018]	<.0001	150
R2Unweighted OLS rate	−0.014	0.010	[−0.033, +0.005]	.156	150
R3IDACI replacing IMD	−0.257	0.056	[−0.367, −0.146]	<.0001	150
R4a2023 only	−0.344	0.065	[−0.472, −0.216]	<.0001	144
R4b2024 only	−0.243	0.065	[−0.370, −0.117]	.0002	148
R6Exclude inherited IMD / rate > 1	−0.291	0.054	[−0.397, −0.185]	<.0001	140
R7Negative binomial	−0.166	0.068	[−0.300, −0.032]	.015	150
R8Mediation + tribunal combined	+0.036	0.043	[−0.048, +0.119]	.401	150

What didn't hold

Reported because we preregistered it, not because it flatters the story

H5 (home-education spillover) found nothing. We predicted higher refusal rates would push families toward elective home education; the estimate is directionally positive but null (p = .50). It was flagged as underpowered in the preregistration and it stays null in the paper.

One robustness variant is fragile. The unweighted OLS rate form (R2) reaches only one-sided p = .078 — small authorities inject noise that the count and weighted forms are designed to handle, but we report it plainly.

Our priors underestimated the data. We predicted a channel interaction of −0.10; the estimate is −0.43 — far outside our locked 90% interval. The honest reading: we expected a modest resource gradient and found a large one.

What follows

For families

A refusal to assess is, statistically, an invitation to appeal. The tribunal is free, and when it decides refusal cases it overwhelmingly decides them for families. The deadline is generally two months from the refusal letter.

For policy

A ~99% success rate against an 11% take-up rate is not an appeals system — it is a persistence tax. Automatic independent review of refusals would remove the gate; funded advocacy in deprived areas targets the measured constraint.

For the DfE

Publish tribunal outcomes by local authority, and link appeal flags to refusal cohorts. Those two releases would close the gaps this paper has to bound rather than measure.

Methods & materials

Local-authority × year panel built from the DfE SEN2 person-level collection (2023–24), the January school census, the English Indices of Deprivation 2019 and MoJ tribunal tables. Hypotheses, exact specifications and priors locked before estimation; raw snapshots pinned by SHA-256; every number script-generated. Analysis executed by an LLM research agent under human direction — and designed so you don't have to take anyone's word for it.

Working paper ↓Plain-English brief ↓Preregistration ↓Results (JSON) ↓Pipeline code ↓

See how your local authority compares in the LA explorer.