Colloquium

Department of Statistics

Statistics Colloquium
April 6 (Fri) 11 a.m.
Alavi Commons Room, 6625 Everett Tower

Evaluating Prediction Models for the Short Term and Long Term Risk of Suicide Attempt

Hanga Galfalvy, PhD
Assistant Professor of Clinical Neurobiology
Columbia University/New York State Psychiatric Institute
Division of Molecular Imaging and Neuropathology
Department of Psychiatry

Accurate prognostic models for evaluating the risk of future suicide attempt in psychiatric patients would allow clinicians to identify patients in need of more intensive treatment without prescribing unnecessary treatment for patients less likely to attempt suicide. Prediction of the immediate or very short term risk is especially important, but it is also very hard to estimate accurately from data collected by research studies, mostly due to the low event number in the relevant time frame. Using short-term risk prediction from long-term models is the preferred solution, however, it is not clear how accurate these risk estimates would be. In this talk I will compare the predictive performance of Cox proportional hazard regression models and bagged survival trees built from data collected in a 2-year prospective study of 384 depressed patients, for the prediction of future suicide attempts 1. during the first six months after baseline assessment, and 2. during the six months to two year timeframe. Statistical validity of the models was evaluated based on cross-validated predicted probabilities but yielded unexpected results: better predictive performance was expected for early attempts, since patients’ state measures change over time, however, the results showed comparable accuracy. I will present evidence that the explanation lies in the existence of previously undiscovered clinical subgroups in the sample that have different shapes of the hazard function (so different times of highest risk). The conclusion is that building separate risk models for proximal and late suicide attempts can somewhat improve performance over predicting a common risk model but such models may still have poor predictive properties because they are essentially based on a mixture of hazard functions. A better estimate of the person-level proximal risk of suicide attempt could be obtained by predicting based on subtype alone, however, there is a trade-off with a need for an increased sample size or a more costly, targeted, recruitment protocol for our research studies.

All statistics students are expected to attend.

Past colloquiums