Methods

PRISM-R is built from a reproducible data pipeline. Every figure on this site is produced by Python scripts that read publicly available source files and emit flat JSON. The pipeline, the source list and the processed data are open. This page records how PRISM-R works and the decisions behind it.

Published figures versus PRISM-R analysis

PRISM-R draws a clear line between two kinds of number. Where an authoritative body publishes a headline statistic, PRISM-R cites the published figure with explicit provenance rather than substituting its own. The four headline numbers on the home page are cited this way, from Youth Justice Statistics 2024 to 2025. The adult Relative Rate Indices that PRISM-R holds for reference are cited from the Ministry of Justice in the same way: they are derived from data that is not public, so they cannot be reproduced and carry no confidence interval.

Where no published figure exists, PRISM-R produces its own analysis and labels it as such. The road-to-remand cascade, the child-specific Relative Rate Indices, and the remand decision mix by ethnicity are PRISM-R analysis, computed transparently from open data. Every chart on the national picture page is captioned with its source so the distinction is immediate.

A note on terminology

PRISM-R uses current best-practice terminology in its prose, drawing on the racial terminology research published by Action for Race Equality in 2024 and 2025. In the text, children of Mixed Heritage are described as such rather than as "Mixed children", and the residual group is described as children in the Other category.

Charts and data tables retain the source-data category labels exactly: White, Black, Asian, Mixed and Other. These are the labels used by the Youth Justice Board, the Ministry of Justice and the Office for National Statistics in the underlying statistics, and PRISM-R is bound to them for fidelity. The Other category is the ONS residual category: it holds ethnic groups not separately identified, and it is not a single community.

The Relative Rate Index

The Relative Rate Index compares the rate at which an ethnic group experiences an outcome with the rate for the White baseline group. A value of one is parity. The method was recommended in the Lammy Review and is used by the Ministry of Justice. PRISM-R applies it to children specifically, which published official statistics do not.

Each PRISM-R-derived index carries a 95% confidence interval. It is an exact interval, computed by conditioning on the total of the two counts being compared, rather than a normal approximation: the counts in the child series are small enough that an approximation overstates precision. Exact intervals are conservative, so they are a little wider than the approximate ones they replaced in September 2026. Child custodial sentencing counts are small and unstable year to year, so that index is presented as a three-year pooled estimate; the single-year values are retained for inspection. Confidence intervals are wider for groups with smaller populations, which is a property of disaggregation, not a flaw.

The road-to-remand cascade

The cascade follows children through four decision points: stop and search, arrest, remand, and custodial sentence. Stop and search and arrest rates use the 2021 Census child population as the denominator; remand and custodial sentencing use a count of children already at that stage. A fifth stage, charge, is out of scope for this version: no open data gives child charges by ethnicity at a usable granularity. It is a candidate for a later version.

Downloads

Everything PRISM-R publishes is downloadable. Each CSV carries the same provenance columns as the JSON, so a file can be cited without returning here: source and reference period travel with every row, alongside a disclosure status. A blank value with a status of suppressed is a cell withheld under disclosure control. It is not a zero and must not be read as one.

Chart data, as CSV

• Road to remand cascade Relative Rate Index by decision point and ethnic group, with confidence intervals.
• Remand outcomes Use of remand for children, England and Wales, by ethnicity, age band, sex and remand type.
• Remand target tracker The monthly remand population against the 25 per cent commitment, with its 12-month rolling average.
• Remand duration Nights on remand by band and ethnic group, for episodes ending in the year.
• Context indicators Exclusions, children looked after, stop and search, arrests and child income deprivation, sub-nationally.
• Child population Population aged 10 to 17 by ethnic group: the denominator behind every rate.

The geographic explorer exports its current map view and any area's summary panel as CSV from the page itself, because both depend on the filters chosen. Those exports carry the same columns as these files.

Structured data

• explorer/index.json Every geography with its parents, child population, deprivation decile, the indicator catalogue and the national comparators.
• explorer/rgn.json, utla.json, lad.json, pfa.json The indicator records for each geography level, already disclosure-controlled.
• manifest.json The provenance record: every output with its SHA-256, size, record count and sources, plus the build's step log.
• suppression_audit.json One record per suppression decision, so a reader can see the rule applied and not only its effect.

Raw source files

The unmodified publisher files the pipeline reads are published as a release asset, so the whole pipeline is reproducible from source: the raw-data releases, most recently raw-data-2026-09-05-2. The archive carries a RAW_BUNDLE.md listing every file with its source URL, retrieval date, SHA-256 and licence. Unpack it at the repository root and run python pipeline/build.py. Always take the newest release: raw-data-2026-09 is missing four ONS boundary files and will not build. It is left published, and marked superseded, rather than rewritten, because a citable asset should not change contents under a fixed name.

All PRISM-R outputs contain public sector information licensed under the Open Government Licence v3.0. Please cite PRISM-R, Oxon Advisory, and the underlying publisher named in the row you use. Citing PRISM-R gives the form of citation.

Citing PRISM-R

PRISM-R is refreshed as its sources publish, so any citation should name the version and the date the figures were read. Each release is tagged in the repository and every page carries the date of the data it shows.

Gilmour, S. (2026) PRISM-R: an open analysis of remand disproportionality in the youth justice system of England and Wales (version [tag]). Available at: https://prism-r.howpreventionworks.com (Accessed: [date]).

For a specific figure, cite the underlying source as well, which is named in the caption of every chart and in the provenance columns of every CSV. Where a figure is PRISM-R's own analysis rather than a published statistic, the chart caption says so.

The repository at github.com/SGilr/prism-r carries a CITATION.cff file, which GitHub renders as a citation block, and the raw data bundle used for each build is published as a release asset.

Disclosure control

PRISM-R applies five disclosure control rules: counts below six are suppressed; a suppressed count must leave no published field it can be recovered from, which covers both back-calculation from a group total and any rate, share or percentage derived from the count; rates are not shown where the denominator is below 100; suppression already applied in a source is carried through and flagged as inherited; and every suppression decision is logged. National figures rarely trigger suppression because the counts are large; sub-national cells do, and they pass through the same rules.

The full audit trail is published, one record per decision, so a reader can see the rule applied and not only its effect: suppression audit.

Defects found in PRISM-R's own figures or methods are recorded, dated and kept permanently, whether or not anyone outside the project noticed them. Two are recorded so far, both from 3 September 2026 and both instances of a suppression being lost between the file where it was applied and a figure derived from it: suppressed counts that could be recovered from their published rates, and 141 suppressed cells republished in the explorer payloads because a derived step ran before the suppression stage. Each is written up in full, with the affected cells, how long they were exposed and the guard added afterwards, in the corrections section of the methods record.

Data sources

The headline statistics and the youth justice analysis draw on Youth Justice Statistics 2024 to 2025 (YJB and MoJ). The cascade also uses Home Office Police powers and procedures, year ending March 2025, and the Ministry of Justice statistics on ethnicity and the criminal justice system. Child population is from the 2021 Census. The build manifest records every source with its reference period, publication date, retrieval date and a checksum for each output: build manifest.

What PRISM-R does not yet show

This version works at England and Wales level, with a static force map and a Lambeth worked example as previews of the geographic explorer. It does not yet show remand at local authority level or post-remand court outcomes by ethnicity; those require data that is not yet published openly. The monthly remand tracker cannot show the remand population by ethnicity, by age or by region, because the Youth Custody Service monthly tables do not cross legal basis with those dimensions; the ethnic composition shown alongside it is the whole custody population, clearly labelled. HMI Prisons' thematic review figure of 216 children on short-term remand in YOIs and STCs cannot be reconciled from the published sector table, which does not cross sector with legal basis; it is on a different basis from the figures here.

Reproducibility

The pipeline is one command. It runs every ingest and computation step in dependency order, applies disclosure control, and writes the manifest. The processed outputs are deterministic and reproduce byte for byte. The code and the processed data are public: github.com/SGilr/prism-r.

Baseline and tracking

The Youth Justice White Paper Cutting Youth Crime. Changing Young Lives., published on 18 May 2026 by the Ministry of Justice, commits to reducing custodial remand for children by 25% over this Parliament. PRISM-R tracks that commitment on the home page: the monthly remand stock from the Youth Custody Service report as a 12-month rolling average, against a baseline of the twelve months to March 2025 (215.75) and a 75% target line (161.8), with the annual flow of remands to youth detention accommodation (baseline 988, target 741) as the accountable measure. The stock is the whole youth secure estate, including the small number aged 18 who remain in the youth estate, because the published tables do not cross legal basis with age. The stock refreshes monthly and the latest month is provisional; the flow refreshes annually. The full definition, including the equity condition and what is not treated as progress, is in the target metric definition.

What the White Paper says about disproportionality

Paragraph 1.32 of the White Paper acknowledges the disparities PRISM-R analyses: "Black children, who make up just 6% of the 10-17 population, represent 9% of arrests, 12% of sentences and cautions, 26% of custodial remands and 22% of the youth custody population." It explicitly names adultification bias as a contributing factor at key decision-making points. PRISM-R's findings on the road-to-remand cascade are consistent with this government analysis, presented openly and reproducibly so that progress against the 25% target can be tracked independently.

Note: the White Paper quotes a 26% Black share of custodial remands; the YJS supplementary tables used by PRISM-R show 28%. The discrepancy is likely due to different cuts of the same underlying data, possibly stock against flow or different reference periods; it does not materially change the picture.