Read this topic

Classification & Assessment in Psychiatry · Paper A

Ordinal rating-scale data

The idea

Ordinal rating scale data gives you a trustworthy order, but not a trustworthy ruler. A HAM-D or PANSS score tells you honestly that a higher number means worse, the same way finishing 2nd is worse than finishing 1st. But it never proves the gap between 1st and 2nd is the same size as the gap between 9th and 10th. So the moment a mean, a standard deviation or a parametric test appears next to a rating-scale score, ask what it is silently assuming about those gaps.

How it works

The four classical levels of measurement, in order of increasing information: nominal (named categories, no order) < ordinal (ranked categories, gaps not proven equal) < interval (equal gaps, arbitrary zero) < ratio (equal gaps, true zero).
Ordinal rating scale data — HAM-D, PANSS, mRS, a single Likert item — earns its rank order from clinical judgement calls, but nothing about how these scales are built proves the psychological distance between adjacent categories is equal.
Because only the rank order is trustworthy, the valid descriptive statistics are the median and interquartile range, and the valid comparative tests use only rank information: Mann-Whitney U (two independent groups), Wilcoxon signed-rank (paired data), Kruskal-Wallis (three or more groups), Spearman's rank correlation (association).
A mean, a standard deviation, or a parametric test (t-test, ANOVA, Pearson's correlation) all assume equal intervals — an assumption ordinal data has never earned — so applying them to a single raw ordinal item is a category error, not just a stylistic choice.
The one deliberate exceptiona SUMMED total across many ordinal items (a full HAM-D or PANSS score) is, by long-standing research convention, often analysed with the parametric toolkit anyway. For example, a trial reporting a HAM-D total falling from 24 to 12 will usually be given a mean, a standard deviation (SD) and a t-test. The pragmatic argument is that summing many items approximates a continuous distribution — an exception that belongs to the summed total, never to a single raw item or an unsummed category.
In practice, this means the exam question always turns on one detail: is the reported statistic describing a single raw item, or a summed multi-item total?

How to spot it

A vignette naming a specific psychiatric rating scale (HAM-D, PANSS, YMRS, mRS) and reporting individual item scores, or a single 1-to-5 severity judgement, is almost always describing ordinal rating scale data.
A vignette that reports a mean and standard deviation for a scale's overall summed total (not a single item) is invoking the deliberate summed-total convention, not a plain ordinal analysis.
A vignette describing categories with no rank at all (diagnosis, marital status, ethnicity) is nominal, not ordinal — check whether the categories can genuinely be put in order.
A vignette describing a millimetre distance marked along a continuous line (a visual analogue scale) is continuous/interval-like, not ordinal, despite using the word 'scale'.
A vignette describing a count of events (admissions, relapses) with many zeros and a skewed distribution is ratio-scaled discrete count data, even though it is often analysed with non-parametric methods for a different reason (skew, not unproven intervals).

Tell it apart

vs nominal datanominal categories have NO inherent order at all (diagnosis, sex); ordinal categories are genuinely rankable (mild < moderate < severe).
vs interval datainterval data's equal-sized gaps are proven (e.g. degrees Celsius); ordinal data's gaps are not proven equal, even though both are described with numbers.
vs ratio dataratio data adds a true, meaningful zero on top of proven equal gaps (e.g. drug dose in mg); ordinal data has neither the proven equal gaps nor necessarily a meaningful zero.
vs a summed Likert scale total scorea single raw ordinal item stays ordinal and needs non-parametric tests; a total summed across many such items is, by convention, often treated as approximately continuous instead.
vs a visual analogue scale scorea VAS is a millimetre distance along a continuous line, not a choice from a small set of ranked labels, so it is usually treated as continuous/interval-like rather than ordinal.
vs discrete count dataa count (number of relapses) has a true zero and equal intervals (it is ratio-scaled) even though, like ordinal data, it is often analysed with non-parametric or count-specific methods because of skew, not because of its measurement level.

Watch for

Assuming any number produced by a rating scale can be averaged just because it looks numeric — the numbers are ranks wearing a numeric costume, not proven equal-interval measurements.
Reporting a mean and SD for a single raw ordinal item, then running a t-test or ANOVA on it, when the median, IQR and a non-parametric test are what the data actually supports.
Missing the summed-total exception — wrongly insisting a full HAM-D or PANSS total score must be treated exactly like a single raw item, when research convention treats the summed total as approximately continuous.
Confusing a count (ratio-scaled, but often skewed) with ordinal data just because both are sometimes analysed with non-parametric methods — the reason for using a non-parametric test is different in each case (unproven intervals for ordinal; skew for counts).
Calling a visual analogue scale 'ordinal' purely because the word 'scale' appears in its name, missing that its millimetre measurement is genuinely continuous.

Use it again

First askcan these values be genuinely ranked? If no, it is nominal, and the answer is the mode, frequencies and a chi-squared test.
If yes, ask a second questionhas anyone proven the gaps between the ranks are equal? If no — a rating scale, a severity grade, a single Likert item — it is ordinal, and the answer is the median, IQR and a non-parametric test.
If yes, check for a true, meaningful zerono zero-as-absence (like temperature) means interval; a genuine zero-as-absence (like age, dose, or time) means ratio — both get means, SDs and parametric tests.
If the stem is specifically about a SUMMED total built from many ordinal items, remember the exception: research convention often analyses that total with the parametric toolkit anyway, even though a single raw item would not qualify.
If the stem describes counting events or measuring time-to-an-event, check for the separate reasons non-standard methods are needed — skew for counts, censoring for survival data — rather than assuming the measurement level itself has changed.

In detail

Hook — the score that lies about its own arithmetic

Picture two patients on a psychiatric ward, both rated on a 0-4 severity scale for how disturbed their sleep is: one moves from 0 to 2, the other from 2 to 4. Both changes look identical on paper, "up by 2." But nothing about the scale actually proves the jump from "no problem" to "moderate problem" feels the same size, to the patient living it, as the jump from "moderate" to "severe." If those two jumps are not provably equal, then adding, averaging, or making "twice as bad" claims about this number are all resting on an assumption nobody has checked.

This page is about exactly that assumption, and about the specific rung of the measurement ladder — ordinal data (values that can be ranked in order, like 1st, 2nd, 3rd, but where the size of the gap between neighbouring ranks is not proven equal) — where a huge amount of real psychiatric data actually sits.

Plain definition

Every number a researcher collects belongs to one of four classical levels of measurement, and each level earns a different set of statistics. Nominal data is just named categories with no order (diagnosis type, sex, marital status). Ordinal data adds a genuine order — you can say one value is higher than another — but not proven-equal gaps between the ranks. Interval data has equal, known gaps between values, but its zero point is arbitrary (0 degrees Celsius is not "no temperature"). Ratio data has equal gaps and a true zero (a zero value that genuinely means "none of the thing being measured", so 40mg really is twice 20mg).

A psychiatric rating scale — the (HAM-D), the Positive and Negative Syndrome Scale (PANSS), or a single Likert item (one question answered on a fixed ranked scale, such as 1 to 5) — sits squarely on the ordinal rung: it ranks severity honestly, but it does not prove its intervals are equal.

The critical caveat — a rating scale is not automatically 'just a number'

Here is the trap almost every exam candidate and every junior researcher falls into: a rating scale is scored with numbers, so it feels safe to treat it exactly like any other number, take the mean, run a t-test, report a standard deviation. That instinct is wrong for a single ordinal item, because the mean of 0, 2 and 4 assumes those values are evenly spaced points on a real number line, and ordinal data has never earned that assumption.

The honest statistics for ordinal data are the median (the middle value when all the scores are lined up in order) and the interquartile range (the range covering the middle half of all the scores), and the honest comparative tests are non-parametric tests (statistical tests, such as the Mann-Whitney U test, that only use the rank order of the data, never the exact size of the gaps). A parametric test (a statistical test, such as a t-test or ANOVA, that assumes the data's gaps are all genuinely equal) silently smuggles in an assumption ordinal data cannot support.

First principles — why order without equal gaps breaks arithmetic

Start from what arithmetic actually needs to be valid. Addition and averaging only make sense when a "1 unit" step means the same thing wherever it happens on the scale, and that is what "equal intervals" means. A ruler earns this: the centimetre between 1cm and 2cm is physically identical to the centimetre between 41cm and 42cm. A rating scale has never earned this, because "moderately depressed" and "severely depressed" are clinical judgements bundled into a category label, not physical distances; nobody has measured whether the psychological gap between them is the same size as the gap between "mild" and "moderate". So you can trust the order (severe is genuinely worse than mild), but the moment you add, subtract, average, or divide the numbers as if they were centimetres, you are making a claim the scale itself never proved.

Structured walk-through — the same fact, worked through a real scale

Suppose a trial uses a 0-3 anxiety item (0 = none, 1 = mild, 2 = moderate, 3 = severe) on 100 patients, and 25 patients score each of the four values. The median sits between the 1 and 2 categories, so you can honestly report that half of patients scored 2 or higher.

Now suppose you instead average the raw numbers to get a mean severity of 1.5, and use that mean to conclude a drug halved the group's severity from a starting mean of 3.0. That conclusion assumes the jump from 2 to 3 represents exactly the same real-world change as the jump from 0 to 1, an assumption this scale has never tested, so the "halved" claim is arithmetic performed on a number that was never built to support it. Swap in the median and interquartile range instead, and you can only say the middle of the group moved from "severe" toward "mild-to-moderate", a weaker but honestly earned claim, and that gap between what the mean implies and what the median actually supports is the whole lesson of this page in miniature.

Look-alikes & differentials

Looks like ordinal rating scale data	How to tell it apart
Nominal data (diagnosis, sex, marital status)	No inherent order at all — categories differ in kind, not in "more or less"
Interval or ratio data (temperature, age, dose in mg)	Gaps are equal and provable, so a mean genuinely represents the group
A summed multi-item scale total (a full HAM-D or score)	By research convention, the summed total across many items is often treated as approximately continuous, even though a single raw item is not
A visual analogue scale mark	Measured as a millimetre distance along a continuous line, not chosen from a small set of ranked labels

Diagnosis — how you actually tell a variable is ordinal, not just call it one

There is no formal test that stamps a variable "ordinal"; you work it out from two questions asked in order. First: can the values be ranked, so one is genuinely more or less than another? If no, it is nominal. If yes, second: is there proof, from how the scale was built and validated, that the gaps between ranks are equal? A rating scale built from clinician judgement calls (mild/moderate/severe, or a 0-4 anchor-point item) almost never has that proof, so it stays ordinal by default. A physical measurement with a real unit (millimetres, milligrams, days) usually does have it, which pushes it up to interval or ratio. When in doubt, the safer, more defensible classification is the lower rung; treating genuinely ordinal data as if it had earned interval status is the exam's favourite trap.

Management — which statistics you are actually allowed to run

Once a variable is confirmed ordinal, the handling of the data is a fixed toolkit, not a judgement call. Describe it with the median and interquartile range, never a mean and standard deviation. Compare two independent groups with the Mann-Whitney U test, paired measurements (for example, before-and-after scores on the same patients) with the Wilcoxon signed-rank test, and three or more groups with the Kruskal-Wallis test; test association between two ordinal variables with Spearman's rank correlation, not Pearson's correlation, which assumes interval-level gaps.

The one deliberate, examinable exception is the summed total built from many individual ordinal items (a full HAM-D or score, not a single item): by long-standing convention in psychiatric and psychological research, it is often analysed with the parametric toolkit anyway, means, standard deviations, t-tests, on the pragmatic argument that summing enough ordinal items smooths the result toward something closer to continuous. That exception belongs to the summed total only; it never rescues a single raw ordinal item or an unsummed category.

Exam pearl

Whenever a vignette hands you a rating scale, a severity grade, or a single Likert item and then reports a mean and standard deviation, ask yourself one question before reading the options: has anyone actually proven the gaps between this scale's categories are equal, or is this just a rank order wearing numbers as a costume? If it is the costume, the median and a non-parametric test are the answer the exam wants. Before you move on, try explaining to a friend why you can trust that a score of 90 is worse than a score of 60, but not that it is exactly "50% worse."

In one line

Ordinal rating scale data gives you a trustworthy order but not proven-equal gaps, so the median, the interquartile range and rank-based non-parametric tests are the honest choice, except for a summed multi-item total, which convention treats as approximately continuous.