Why Suicide Risk Scores Cannot Predict an Individual

The short answer

Suicide risk scores cannot tell you what a single person will do, and the reason is arithmetic rather than a failure of clinical skill. Suicide is a rare event, so even a score that reliably separates higher-risk from lower-risk groups will label far more people high risk than will ever die, while missing a large share of the people who do. The largest longitudinal synthesis of the evidence found that roughly 95 percent of patients placed in a high-risk category did not die by suicide, and close to half of all suicides occurred among people the same tools rated as lower risk. A number can describe a population; it cannot forecast an individual.

Why rare events defeat prediction

Every screening tool lives or dies by three numbers. Sensitivity is the share of true cases it catches. Specificity is the share of non-cases it correctly clears. Base rate is how common the outcome actually is in the group being tested. The first two get most of the attention, and the third quietly controls the result.

Suicide has a very low base rate, even among psychiatric patients. When the outcome is rare, the pool of people who screen positive fills up with false positives, because there are so many more people who will not have the outcome than people who will. This is the same reason a highly accurate test for a rare disease still produces mostly false alarms. No amount of tuning sensitivity and specificity rescues a prediction when the event itself is uncommon. Put concretely, when the pooled data placed patients in a high-risk category, only about 5 or 6 of every 100 later died by suicide, so anyone acting on the label was responding largely to the roughly 94 who would not.

What the largest analysis actually found

In 2016, Large and colleagues published a meta-analysis in PLOS ONE pooling 37 studies, 53 samples, and 315,309 psychiatric patients, among whom 3,114 died by suicide. High-risk categorization carried a pooled odds ratio of 4.84, which sounds decisive: high-risk patients died at several times the rate of lower-risk patients. The detail behind that ratio matters. Sensitivity was 56 percent, meaning 44 percent of suicides occurred among people rated lower risk. Specificity was 79 percent. The positive predictive value was 5.5 percent, so about 95 percent of the high-risk group did not die by suicide.

Two further findings sharpen the point. Between-study heterogeneity was extreme, with an I-squared near 93 percent, so the tools did not even behave consistently from one setting to the next. And there was no evidence that predictive strength had improved across roughly four decades of research. The authors concluded that a statistically strong, reliable method to distinguish high-risk patients remains elusive, and that a group-level difference means little if there is no intervention that should be given to high-risk patients yet withheld from low-risk ones.

The cost of being wrong in both directions

Read as a group statistic, a fivefold difference looks like a reason to act. Read as a decision about one patient, it splits into two kinds of error, each carrying a human cost.

False positives are the roughly 95 in 100 high-risk patients who will not die. Treating a score as a verdict can mean coercive admission, lost autonomy, and scarce resources spent on people who did not need that particular response. False negatives are the near half of suicides that occur in the lower-risk group. A reassuring score can license lighter attention and thinner follow-up for exactly the people a tool was meant to protect. An instrument that generates both errors at once cannot safely substitute for judgment.

The pattern is not confined to a single instrument. A 2017 meta-analysis by Carter and colleagues in The British Journal of Psychiatry examined 70 studies of risk scales and found a pooled positive predictive value near 5.5 percent for suicide. They concluded plainly that no high-risk classification was clinically useful for allocating care, and that the low base rate itself sets a ceiling on what any such scale can achieve.

A better use of the same information

None of this means the risk factors are fictional. Prior attempts, active suicidal ideation, and acute distress genuinely raise probability at the population level, which is why they belong in research and in public health planning. The error is treating a population probability as an individual prediction, then using that prediction to ration attention.

Both meta-analyses point in the same direction: away from stratifying people and toward addressing needs. Rather than sorting patients into risk tiers and matching the tier to a level of care, a needs-based approach asks what modifiable problems a person has in front of them and offers evidence-based help accordingly. Distress, access to means, pain, and thin support are conditions to respond to, not inputs to a score that claims to know the future.

This article is educational and not medical advice. If you or someone you know is struggling, contacting a local crisis line or, in the United States, the 988 Suicide and Crisis Lifeline connects you to trained support.