DClinPsy Equal-Opportunities Explorer

About this explorer

An interactive, statistically honest look at who applies to, and who is accepted onto, the Doctorate in Clinical Psychology (DClinPsy) in the UK, by applicant characteristic, from 2015 to 2024.

The data is the Clearing House for Postgraduate Courses in Clinical Psychology equal-opportunities monitoring, published each year. The "acceptance rate" is the share of applicants in a group who were accepted onto a course. Every number on this site is computed live in your browser from clean counts, and each computation is checked against an offline reference written in R, so there is a gold-standard value behind every figure on screen.

How the numbers are made

There are two layers. Offline, R cleans and reconciles the raw published tables, builds the aggregate groups (the broad ethnic groups and Mixed are taken from the source's own published group totals; other aggregates are transparent sums of named sub-groups), and emits two things: the clean per-group, per-year counts that the site loads, and a reference results table that the browser code is unit-tested against. In the browser, the site loads only those counts and computes the rates, intervals, and significance verdicts itself. The browser only ever sums clean counts and skips suppressed cells. It never re-derives the suppression rule, so the one sensitive judgement lives in a single vetted place.

Suppressed counts are not zero

Where an accepted count is too small to publish safely, the Clearing House withholds it. A withheld value means "too small to report", not zero. These cells are dropped before any sum, never interpolated and never filled with zero. In the trend charts a suppressed year breaks the line rather than dropping it to the floor, and the table marks suppressed cells explicitly.

Uncertainty

Every rate is an estimate from a finite number of applicants, so it carries a confidence interval.

• Single rates use the Wilson 95% interval, which behaves well for small counts and rates near 0 or 100%.
• Differences between two groups (a group versus its reference) use the Newcombe interval, built from the two Wilson intervals.
• Rate ratios use the Katz log method.
• Change between eras is tested with a Welch t-test on the per-year values, which does not assume equal variance.

Significance, and why it is honest

Testing many groups at once inflates the chance of a false positive. We correct for this with the Benjamini-Hochberg procedure, which controls the false discovery rate, and the displayed intervals are widened to match, so a bar that excludes "no difference" is exactly the one flagged as significant.

Crucially, the size of each correction family is fixed in advance by a pre-specified register of families, not by how many groups you happen to have on screen. This prevents significance-shopping: you cannot make a result appear or disappear by changing the selection. Because the family size is pinned, a group can be significant in the pooled "All groups" view but not within its own smaller family, since the correction differs. British Welsh is the known borderline case.

Reading the colour: red always means a statistically significant difference, and grey means not significant. Direction (better or worse than the reference) is shown by position, sign, or the direction of an arrow, never by colour. This keeps the significant findings unambiguous at a glance.

Reading the p-values: hover any point, or open the Table view, to see the exact p-values. A p-value is the probability of seeing a difference at least this large if the groups truly accepted at the same rate: smaller means stronger evidence against "no difference". It is not the size of the gap, nor the probability that the gap is real, and 0.05 is a convention, not a law of nature. We show two: raw p is the single uncorrected test, and p (BH) is that p-value after the Benjamini-Hochberg correction across the family. The corrected one is what decides the red/grey verdict, so always read it alongside the raw one. A result that only just clears 0.05 (for example Black ethnicity at p (BH) ≈ 0.046 over 2021 to 2024) is genuine but fragile, and should be read as "just significant", not as settled fact: the confidence interval, which shows how big and how uncertain the gap is, matters more than the p-value alone.

Eras: pre-DEI versus DEI

The "Pre-DEI vs DEI" view compares the pre-DEI years (2015 to 2019) with the DEI years (2021 to 2024). Each era value is the mean of that group's per-year rates or gaps, with a 95% interval, and the change is tested on those per-year values. 2020 is treated as the changeover year and sits in neither era. In the other views 2020 is a normal, fully-included year; the trend charts simply shade 2021 onward to mark the DEI era visually.

Representativeness

This view asks a different question: does the applicant and accepted pool look like the population it is drawn from? Shares are compared to a population benchmark and shown as the difference from it (a bar to the right is over-represented, to the left under-represented). Benchmarks come from the 2021 Census (England and Wales) unless stated, and are rebased to the "answered" population so they sit on the same base as the applicant data, which excludes prefer not to say.

Disability is age-adjusted

Disability prevalence rises steeply with age, and applicants are young, so the whole-population rate (about 18%) is the wrong yardstick. Instead we take the Census age-specific disability rates (Equality-Act limited a lot or a little) and re-weight them by the applicant, and separately the accepted, age profile. This is indirect standardisation, and it gives an age-matched expected rate of about 12%. It adjusts for age only, not for the definition gap (the Census measures day-to-day limitation while applicants self-report condition types) or for sex.

Socio-economic background uses POLAR

The socio-economic measure is POLAR, which groups areas by the share of young people who enter higher education, assigned from an applicant's home area. The five quintiles are equal fifths of the young population by construction, so a representative pool would sit at 20% per quintile, the benchmark used here. Quintile 1 is the lowest-participation, most disadvantaged areas. It is an area-based proxy, not a measure of individual circumstances.

Other caveats

• For detailed ethnicity, British English, Scottish and Welsh are combined as White British, which the Census does not split by nation.
• Benchmarks other than disability are all-ages, so for a young graduate pool they slightly understate representation.
• The Census covers England and Wales, while applicants are UK-wide.

What this cannot tell you

• The published data is one characteristic at a time. It cannot show intersections (for example ethnicity within a socio-economic quintile).
• Acceptance gaps are associations, not causes. They do not by themselves identify where in the process a disparity arises, nor why.
• Group labels and categories are the Clearing House's own, and some are self-reported.

Glossary

Acceptance rate

Share of applicants in a group accepted onto a course.

pp (percentage points)

The arithmetic difference between two percentages. From 40% to 43% is 3 pp.

Reference group

The group each other group is compared against for a characteristic (for example British English for ethnicity).

Wilson / Newcombe interval

95% confidence intervals for a single rate, and for a difference between two rates.

p-value

The probability of a difference at least this large if the groups really accepted at the same rate. Smaller means stronger evidence against "no difference"; it is not the size of the gap. raw p is uncorrected; p (BH) is after the Benjamini-Hochberg correction and is what sets the verdict.

Benjamini-Hochberg

A correction that controls the false discovery rate when many groups are tested together.

POLAR

Participation Of Local Areas: an area-based measure of how many young people enter higher education, in five equal quintiles.

DEI era

Here, the 2021 to 2024 application cycles. 2020 is the changeover year.

How this was built

This explorer was built by a single author working with an AI coding assistant (Claude). In the interest of transparency, here is what that involved.

Done by the author

• Conceiving the project, sourcing the Clearing House data and the Census benchmarks, and deciding which questions to ask.
• The statistical and editorial judgements: which groups to compare and against which reference, the pre-specified significance families and the decision to fix their size in advance, the choice of effect measures, the era definitions, the suppression handling, the decision to age-adjust disability and to use the POLAR 20% baseline, and which comparisons to leave out (for example marital status).
• Running the R pipeline and checking its output.
• Reviewing the results and the wording, and deciding what to publish.

Done with AI assistance

• Writing the code: the R cleaning and statistics pipeline, the browser application, and porting the statistical methods (Wilson, Newcombe, Benjamini-Hochberg, rate ratios, the Welch t-test, the disability age-standardisation) into JavaScript so they run live.
• Building the interface, the charts and this page, and drafting explanatory text, which the author then edited.
• Proposing options and trade-offs for the author to weigh and decide on.

AI assistants can and do make mistakes, so none of the figures here rely on trusting the generated code. Every statistic the browser computes is checked against an independent reference implemented in R, matched to eight decimal places, and the full data chain is verified end to end: the source tables reconcile to the clean counts the site loads, and the live browser calculations reproduce the R reference exactly. The R layer is the gold standard; the AI-written browser code only ever has to reproduce it, and where the two disagree the R reference wins and the build does not ship.

The AI accelerated the building, not the judgement. The analytical decisions are the author's, and the arithmetic is held to an offline standard rather than taken on trust.