Can we predict suicide? What the evidence shows

The finding

In 2017, Franklin and colleagues published a meta-analysis in Psychological Bulletin pooling 365 studies and 3,428 risk factor effect sizes, covering fifty years of research on predicting suicidal thoughts and behaviors.

The results were sobering. Prediction was only slightly better than chance across the analyses. No broad category of risk factor, and no subcategory, predicted far above chance. And predictive ability had not improved across five decades of work.

That last clause is the one that should stop you. Fifty years of accumulating research, and we did not get better at the thing we were trying to do.

What it does and doesn't mean

It does not mean risk factors are meaningless. A history of attempts, active ideation with a plan and access to means, recent discharge, acute loss: these genuinely raise risk at the population level, and ignoring them would be indefensible.

What it means is that they don't let you identify, in advance, which individual in front of you will die. Population-level association and individual-level prediction are different tasks, and the field spent decades treating the first as though it delivered the second.

It also means the tools built on this foundation inherit its limits. A risk score that stratifies a population is not a device for telling you that this patient is safe to discharge, no matter how confidently it's presented.

The base rate problem

There's an arithmetic reason this is so hard, and it isn't going away with better software.

Suicide is, in statistical terms, rare. When an outcome is rare, even a test with good sensitivity and specificity produces overwhelming numbers of false positives. Flag everyone the model calls high risk and you will be flagging mostly people who will not die, while still missing people who will, because most people who die by suicide were never in the flagged group. This is a property of the mathematics, not a deficiency of the instrument, and it constrains machine learning exactly as it constrained the checklists.

What to do instead

If prediction is not achievable, the clinical task has to be reframed as intervention. That's not a retreat. It's where the evidence actually is.

Safety planning has support: the Stanley-Brown Safety Planning Intervention has been associated with fewer subsequent suicidal behaviors when paired with follow-up contact. Means restriction is one of the better-supported interventions in the whole field, because reducing access to lethal means during a period of acute risk changes outcomes even when it changes nothing about the underlying despair. Treating the treatable, meaning the depression, the substance use, the insomnia, the pain, is a real lever. And follow-up contact after discharge, a period of substantially elevated risk, is both effective and routinely omitted.

We publish an educational safety planning overview and risk-assessment documentation structure in the resource library. Both are educational aids for trained clinicians and neither is a screening or prediction tool.

The burden this lifts

There's a thing clinicians carry that they rarely say out loud: the belief that if they were good enough, they would have seen it coming.

The evidence says they wouldn't have. Fifty years, hundreds of studies, and prediction barely better than chance. A psychiatrist who did a careful assessment, documented the reasoning, safety planned, restricted means, and arranged follow-up did the job, and a death after that is not proof of a missed clue. It's proof that the task is not the one we told ourselves it was.

That reframing matters clinically as well as personally, because clinicians practicing in fear of an unmeetable standard practice defensively, and defensive practice is not better care. See how involuntary commitment works for where this collides with the law.

If you're in crisis, call or text 988 in the US to reach the Suicide and Crisis Lifeline, available 24 hours a day. If someone is in immediate danger, call 911.