Methodology
Version 0.3 · data snapshot 14 October 2026 · covers the 2025/26 performance year
1. What this site is
Prison Rankings compares the 120 adult prisons in England and Wales using official data only. Every figure links to its source. We publish our weights, our models and our data.
It is built to answer one question: how does each prison perform compared with what you would expect, given who it holds and what it is for?
It is not a verdict on any individual, a measure of any member of staff, or a prediction of what will happen to any prisoner.
2. The short version
- Prisons are compared only with similar prisons: five groups (Reception, Category C, Category B and High Security, Open, Female). We never rank across groups.
- Each prison gets a band (top quarter to bottom quarter of its group) and a likely rank range, not a precise rank. The ranges are wide: typically close to two-fifths of the group. The band is the more reliable guide.
- Safety outcomes are adjusted for who the prison holds (sentence length, remand, recalls, age, turnover, crowding) before being compared.
- Reoffending, workforce stability and disorder incidents are shown but not scored. Section 8 explains why.
- We also show where a prison is unusually higher or lower than expected on an outcome. We only show this when three different benchmarks agree.
- The official HMPPS rating is shown alongside as context. It is not an input to our scores.
3. Who is covered
| Included | The 120 adult prisons that HMPPS rates in its annual performance framework, including the 15 contracted-out (privately run) prisons. |
| Excluded | Youth custody (under-18 young offender institutions and the secure training centre), which has a separate framework. Immigration removal centres. |
| Out of scope | Scotland and Northern Ireland. They have different systems and there is no UK-wide equivalent data. |
| Closed or very new | Prisons that closed, and HMP Millsike, which is not in the 2025/26 ratings. |
Some prisons are reported by the Ministry of Justice only as combined units: Grendon with Spring Hill, Usk with Prescoed, Peterborough (male and female) and Feltham (including its under-18 site). Where data cannot be separated we say so and do not draw conclusions from it (section 7).
4. Data
All data is from official publications made available under the Open Government Licence v3.0 (GOV.UK states that all its content is, unless marked otherwise; CQC publishes its data under the same licence). Contains public sector information licensed under the Open Government Licence v3.0. Files are stored unchanged with SHA-256 checksums in each snapshot.
| Data | Publisher | Period used | What we take from it |
|---|---|---|---|
| Annual prison performance ratings, supplementary tables | HMPPS | 1 Apr 2025 – 31 Mar 2026 | HMIP grades and inspection dates, prisoner survey grades, resettlement outcomes, purposeful activity, staff sickness and resignations, functional group, comparator groups, the official rating |
| Safety in Custody (monthly, by prison) | MoJ | Apr 2025 – Mar 2026 | Prisoner-on-prisoner assaults, assaults on staff, self-harm incidents |
| Prison population data tool, quarterly | MoJ (OMSQ) | Quarter-ends Jun 2025 – Mar 2026 | Population mix: sentence length, remand, recall, age, offence group |
| HMPPS Annual Digest 2025/26 and data tools | HMPPS | 12 months to Mar 2026 | Average population, crowding, protest incidents, finds |
| Releasing-prison data tool (proven reoffending) | MoJ | Apr 2023 – Mar 2024 cohort | One-year proven reoffending by prison of release |
| HMPPS workforce statistics | HMPPS | 12 months to Mar 2026 | Sickness history, staffing (context) |
| Care Quality Commission locations | CQC | Oct 2026 | Healthcare provider, compliance flag, coordinates |
How current is each measure?
Not everything describes the same year. We show the date next to every figure.
- HMIP grades come from inspections between April 2022 and March 2026 (median August 2024). A prison inspected in 2022 is described by a 2022 inspection.
- Prisoner survey grades come from surveys between May 2021 and February 2026 (median June 2024).
- Incident rates and workforce figures cover the 2025/26 year.
- Reoffending covers people released between April 2023 and March 2024, followed for a year plus a six-month wait for court outcomes, so it is about two years behind. The next edition is expected around April 2027.
5. Who counts as "similar"
Peer groups. HMPPS assigns each prison a functional group (nine groups). Three of them are too small to compare on their own (High Security has 5 prisons, Foreign National Category C has 3), so we combine them into five:
| Our group | Prisons | Made from |
|---|---|---|
| Reception | 31 | Reception (local) prisons |
| Category C | 50 | Category C Trainer, Trainer/Resettlement, Resettlement, Foreign National |
| Category B / High Security | 14 | Category B, High Security |
| Open | 13 | Open |
| Female | 12 | Female |
The combination of Category B with High Security, and of the Category C types, is our judgement, and different groupings would give different bands for some prisons.
Comparator sets. Inside each group, HMPPS publishes up to eight "comparator" prisons for each prison, chosen by nearest-neighbour matching on age, sentence length, turnover, sex-offender and foreign-national shares, complexity, building age, capacity and urban or rural setting. We test these and use them (section 6). Comparator sets are specific to each prison: if A names B, B does not necessarily name A. They cannot be used to build a league table, so groups are used for bands and comparators for expectations and "similar prisons" displays.
In testing, a prison's comparators predicted its outcomes much better than a random eight from its group (rank correlation 0.78 against 0.53 for prisoner assaults; 0.85 against 0.67 for reoffending).
6. What we measure
Scores are built from 14 measures in four pillars. Each pillar has equal weight.
| Pillar | Measures | Adjusted for context? |
|---|---|---|
| Safety | Prisoner-on-prisoner assaults; assaults on staff; self-harm incidents (all per 1,000 prisoners); HMIP safety grade; prisoner survey, safety | Incident rates: yes. Grades: no |
| Decency | HMIP respect grade; prisoner survey, respect | No |
| Regime | HMIP purposeful activity grade; prisoner survey, purposeful activity; share of prisoners in at least half-time purposeful activity | No |
| Resettlement | Housed on first night after release; employed six weeks after release; employed six months after release; HMIP preparation for release grade | No |
Some measures are missing for some prisons. A pillar uses the measures that are available, and the overall score averages the pillars that have data, each with equal weight. Every prison has HMIP grades, so all four pillars are present for all 120; individual measures (for example the survey, or employment after release) are missing for some, so each prison is scored on between 11 and 14 of the 14 measures.
What is missing and why.
- Security. There is no usable drug-testing data: random testing is not published by prison, and the Ministry of Justice states it cannot robustly estimate prevalence. Escapes and absconds are too rare, outside open prisons, to compare. We therefore do not score a Security pillar rather than build one on weak data.
- Education. HMPPS dropped its education measures for 2025/26 when prisons moved to the Prisoner Education Service.
- Recall after release. Recalls within months of release, by releasing prison, would be an excellent resettlement outcome. It is not published at prison level; obtaining it would need a request to the Ministry of Justice.
7. Adjusting for who a prison holds
A prison that holds many short-sentence and remand prisoners has more assaults and more reoffending than an open prison for reasons that are not about how well it is run. We therefore compare each prison's incident rates with an expected rate given its population.
The expectation
For each of three outcomes (prisoner assaults, staff assaults, self-harm) we calculate two expectations and average them (on a log scale):
- A regression expectation. A statistical model, fitted across all prisons, of the outcome against peer group, share of the population on remand, on sentences under 12 months, on sentences of four years or more, on indeterminate sentences and recalled, age (under 25, over 50), violent offences, turnover (releases per head) and crowding. Crowding is included for the safety models because overcrowding is outside a governor's control.
- A comparator expectation. The average outcome of the prison's HMPPS comparators.
Both are leave-self-out: a prison's own result never contributes to its own benchmark. Where fewer than three comparators exist, the regression is used alone.
The model's coefficients are published (data download, v0_2_coefficients.csv).
What we deliberately do not adjust for
Staffing levels, staff sickness, drug finds and anything else that management can influence. Those are performance, not context. Crowding is the exception, because it is a feature of the estate and the system, not of management.
Remand and recall
Remand is decided by courts and recall by probation and the Parole Board, so both are context for a prison. We adjust for them in the safety models. For reoffending the result is sensitive to this choice: with them in the model, almost no between-prison variation remains; without them, some does. A recalled prisoner is by definition someone who has already failed on licence, so the share of recalls partly measures risk that also predicts reoffending. We therefore make both versions available (the version without remand and recall is in the data download).
Statistical treatment
- Overdispersion. Incident counts vary more than a simple count model assumes. We estimate this from the data and widen the limits accordingly.
- Shrinkage. Small prisons swing more by chance. Each prison's ratio of observed to expected is shrunk toward 1 in proportion to how little information it carries.
- Funnel limits. A result is flagged only if it lies outside 95% control limits after the above (z beyond 1.96).
Shared reporting units
Where the Ministry of Justice reports two prisons as one unit, there is one set of counts and no coherent expectation, so we do not flag those prisons on incident outcomes. In an earlier version some of these appeared as strong outliers. That was an artefact.
8. Scores, bands and ranges
Percentiles, not z-scores. Each measure is converted to a position within the prison's group (0 = worst, 1 = best). This avoids treating ordinal grades (1 to 4) as interval data. For adjusted outcomes the percentile is of the shrunk ratio of observed to expected.
Pillar and overall score. A pillar is the mean of its available measure percentiles. The overall score is the mean of the pillar scores. No numbers are shown to more precision than the data supports.
Bands. Within each group, prisons are placed in quarters: top quarter, upper middle, lower middle, bottom quarter. A band is a statement about position within a group of similar prisons. It is not a judgement that a prison is good or poor in absolute terms.
Likely rank range. We repeat the scoring 1,000 times, each time varying pillar weights at random and perturbing the adjusted outcomes within their statistical uncertainty. The range shown is the 5th to 95th percentile of the rank a prison obtains. The median range spans about 38% of the group. A prison in the middle of its group could plausibly be in either neighbouring band. Treat the band, not the rank, as the finding.
What we do not publish. There is no single national league table and no cross-group ranking.
9. Outcomes that are shown but not scored
| Shown as a lens or context | Why it is not in the score |
|---|---|
| Reoffending (adult prisoners, by prison of release) | The Ministry of Justice itself says rates by releasing prison should not be used to assess prisons: they do not control for sentence or offender characteristics and the prison of release may not be where the sentence was mainly served. It is two years old, and the mix of people being released (which drives it) is not published by prison; we use the mix of those held as a proxy. We show prisons that are unusually higher or lower than expected, with the caveat. |
| Staff sickness and resignations | Related to quality (correlated with HMIP grades and our composite at about 0.3 to 0.45) but the direction of cause is unclear: assaults on staff cause sickness and a poor regime drives people out. The official rating already scores both. Contracted prisons report sickness differently and publish no resignation rate, so they are shown without a band. |
| Incidents at height and barricades | Persistent over years but we cannot separate real disorder from differences in recording practice. |
| Finds of drugs, weapons and phones | More finds can mean better searching, not worse security. Without drug-testing data they cannot be interpreted. |
| Food hygiene ratings, CQC | Food hygiene covers kitchens, not food quality, was found for about half of prisons and is almost entirely the top grade. CQC registers prison healthcare but gives no graded ratings, and its compliance judgements exist for only 29 prisons. Both are shown as context where they exist. |
| Staffing against target | Shows no relationship with any outcome we examined. |
10. "Higher or lower than expected" flags
For each of four outcomes we show a flag only when the result is outside the funnel limits and the direction is the same under all three expectations: regression, comparator mean and the average of the two. This is deliberately conservative.
In this edition, nine prison-outcome results meet that standard. All are higher than expected; none is lower. Counts cannot fall below zero, so unusually good results are harder to distinguish from chance. This is a limitation of the method, not evidence that no prison does better than expected.
A flag is a prompt to look, not a finding of fault. Possible explanations include recording differences, a recent change of population, or a small number of individuals. Where a result rests on few incidents (typically small open prisons) we say so.
11. How this differs from the official rating
HMPPS rates each prison 1 to 4 (outstanding to serious concern) on 33 measures against targets. We use the same public data, so agreement is expected.
- Our composite and the official score agree closely but not exactly (rank correlation about 0.80 within groups; 59% land in the same quarter).
- Differences by design. We compare prisons with similar ones; adjust incident outcomes for context; exclude internal audits and process measures (HMPPS includes around a dozen); include measures the rating does not (reoffending as a lens, crowding as context); and publish uncertainty.
- Overlap by construction. HMIP and the prisoner survey are inputs to both. We show the official rating as context, not an input.
- The official rating can be amended by a moderation panel. Five 2025/26 ratings were. Ours is formula-only.
12. Limitations
- Different dates. HMIP and survey grades can be up to four years old (section 4). A recent change at a prison will not be visible.
- Context adjustment is partial. We cannot adjust for everything that differs between prisons (building age, regional labour markets, local policing). The variables we have describe who is held, not who is released.
- Adjustment choices matter. Whether to adjust for remand and recall changes the reoffending result (section 7).
- Peer groups are our judgement. Different groupings give different bands for some prisons.
- Small numbers. Open and female prisons rest on few incidents. Suppressed counts (one or two) are estimated as 1.5.
- Recording. Self-harm, assaults and incidents at height depend on recording practice, which differs between prisons.
- Combined units. Seven prisons cannot be assessed on incident outcomes.
- Ratings are not independent of data we also use. HMIP and the survey feed the official rating; agreement with it is not independent validation.
- Single year. Most outcomes use one year, so short-term shocks (for example the 2024/25 capacity crisis) are not smoothed.
- Source continuity. The 2025/26 HMPPS Annual Digest is the last one planned. Some tables will move to other publications and some may stop.
13. Decisions and their reasons
| Decision | Why |
|---|---|
| Compare within groups, never across | Prisons do different jobs on different populations |
| Bands and ranges, not ranks | Mid-table ranks are sensitive to weights and to chance |
| Adjust for context, not management | A governor cannot choose who is sent to them, but can influence staffing and drugs |
| Remand and recall treated as context | Decided by courts and probation, but result is sensitive and shown both ways |
| Reoffending shown, not scored | MoJ warns against use, no release-cohort mix by prison, results depend on adjustment |
| Workforce shown as its own lens | Related to quality, direction unclear, already in the official rating |
| Disorder incidents and finds not scored | Cannot separate real events from recording; direction ambiguous |
| A flag needs three benchmarks to agree | Avoids relying on one benchmark; fewer, more reliable flags |
| Five groups for tables; HMPPS comparators for expectations | Comparators predict better but cannot form a table |
| Shared reporting units not assessed | No coherent expectation for a combined unit |
14. Review and corrections
- Open process. This page, the data, the coefficients and the code are public. We expect to be told where we are wrong, and we log what changes (see the changelog).
- Corrections. If you believe a figure is wrong, write to contact@prisonrankings.com (see the corrections page for what we do with it). We log every correction on the corrections page and in the changelog.
- Updates. Safety in Custody is quarterly, the HMPPS ratings annual, and reoffending annual. Each update is versioned with its data snapshot and checksums.
Downloads: full data (CSV), model coefficients, the list of prisons with identifiers, and every script used to produce this page are available on the data page.
Version history
| Version | Date | Change |
|---|---|---|
| 0.3 | 14 Oct 2026 | Hybrid expectation (regression plus HMPPS comparators); robust flags; shared units not assessed; workforce lens |
| 0.2 | 14 Oct 2026 | Context-adjusted outcomes, shrinkage, funnel limits |
| 0.1 | 14 Oct 2026 | Unadjusted peer-group composite |