Classification & Assessment in Psychiatry · Paper A
Principles of risk assessment: actuarial, structured and dynamic approaches
Because suicide and serious violence are statistically rare events even in high-risk psychiatric populations, no risk-assessment tool, however carefully designed, can reliably pick out in advance which specific individual will go on to have that outcome; that mathematical fact, not a failure of any particular scale, is why current guidance moves away from using risk categories to decide who gets care.
A mental health team reviews a year of deaths by suicide among the patients it cares for, and finds something that feels deeply wrong: most of those patients had been rated low risk at their last assessment. Does that mean every one of those assessments was a mistake? Almost certainly not, and working out why is one of the most important pieces of statistical reasoning in psychiatric risk assessment. This page builds it from zero.
A base rate (also called the , or the prior probability) is simply how common an outcome genuinely is across a whole group of people, before any test, scale, or clinician's judgement is applied to a single individual. If completed suicide happens to roughly 1 in every 1,000 psychiatric patients over the course of a year, that 1-in-1,000 figure IS the base rate, a fact about the population as a whole, not a prediction about any one person sitting in front of you.
Here is the trap almost everyone falls into: assuming a screening tool with 80% and 80% specificity (sensitivity is how good a test is at correctly catching people who truly will have the outcome; specificity is how good it is at correctly clearing people who truly will not) must therefore be roughly 80% trustworthy whenever it flags someone. It is not. The gap between those two ideas is exactly what a very low base rate does to any screening tool, however well it is built.
Suppose you screen 10,000 psychiatric inpatients for suicide risk over the coming year, and the true base rate is 1 in 1,000. That means, before any test is applied, exactly 10 of those 10,000 people will genuinely go on to die by suicide, and 9,990 will not. Now apply your tool, with 80% and 80% . Sensitivity of 80% correctly catches 8 of the 10 true cases, 2 are missed. Specificity of 80% correctly clears 80% of the 9,990 people who were never going to die by suicide, which sounds excellent, but it also means the remaining 20%, a full 1,998 people, are wrongly flagged as high risk. Add up everyone the tool flags: 8 true cases plus 1,998 false alarms is 2,006 people in total. Only 8 of those 2,006 flagged people are the real cases. That is a positive predictive value, PPV, the chance a positive result is actually a true case, of roughly 8 divided by 2,006, under half of one per cent.
Run the same numbers the other way. Of the 10,000 people screened, 2,006 were flagged and 7,994 were cleared as low risk. Among those 7,994, 7,992 truly were never going to die by suicide, and only 2 of the true cases slipped through as false negatives. That gives a , NPV, the chance a negative result is genuinely correct, of about 7,992 divided by 7,994, which rounds to over 99.9%. So the very same tool, on the very same population, is almost perfectly reassuring when it says low risk and almost useless when it says high risk, and both effects come from the same cause, the outcome's rarity, not from the tool being built badly. Because the low-risk group is by far the biggest group, 7,994 of 10,000 people, it still contains most of the handful of true cases that were missed, which is exactly why an audit can honestly find that most deaths occurred in patients previously called low risk, without that proving anyone made a clinical error.
| Term | What it actually measures |
|---|---|
| Base rate | How common the outcome is in the whole population, before any test |
| How good the test is at catching true cases, a fixed property of the test | |
| The chance a positive result is a true case, swings wildly with the base rate | |
| The chance a negative result is genuinely correct, stays high when the outcome is rare |
There is no diagnostic test being applied to a disease here. What gets measured, in large prospective cohort studies that follow thousands of assessed patients forward in time to see who actually goes on to self-harm or die by suicide, is exactly the arithmetic above: how many of the people a scale flagged as high risk really did have the outcome. Repeated of these follow-up studies, for example work by Carter and colleagues, and by Large and colleagues, find the same pattern again and again, reasonable-looking and , but a positive predictive value too poor to usefully single out which specific individual will act.
NICE's guideline on self-harm, NG225, Self-harm: assessment, management and preventing recurrence, responds directly to this arithmetic: it recommends against using risk-assessment tools and scales, or a global low, medium or high risk category, to predict future suicide or repeat self-harm, or to decide who receives treatment and follow-up. Instead, every patient who has self-harmed should receive a comprehensive psychosocial assessment, covering their needs, their strengths, and the specific factors relevant to them, rather than being sorted into a numerical bucket and treated, or not treated, accordingly. The guidance is not saying risk assessment is pointless; it is saying a single score cannot do the job of picking out the one person, among thousands who look statistically similar, who will actually act.
The classic version of this question gives you a and that both sound reassuring, pairs them with a very low base rate, and asks you to spot that the collapses anyway, or it gives you an audit finding that most deaths occurred in low-risk patients and asks you to explain it without assuming clinical negligence. Before you move on: could you explain to a friend, using only the words rare and biggest group, why most people who die by suicide under a service's care were very likely rated low risk beforehand?
When an outcome is genuinely rare, even a good test flags mostly false alarms and clears mostly true negatives, so a low-risk score is not the same as no risk, and current guidance says not to let a risk category decide who gets care.