Predicting Suicide Attempt Risk: Logistic Regression Requires Large Sample Sizes

Steven C. Bagley, MS, MD

J Clin Psychiatry 2011;72(12):1698

Article Abstract

Letter to the Editor

Because this piece does not have an abstract, we have provided for your benefit the first 3 sentences of the full text.

Dr Gilbert and colleagues make an important point about the difficulties in predicting suicide, but their conclusions do not follow as strongly from their data as they suppose. They report negative findings based on logistic regressions using clinical, demographic, and cognitive predictor variables. However, their data set is relatively small, with 28 events (suicide attempts) in 67 subjects; their regressions use 12 clinical and demographic predictors and, separately, 7 cognitive and demographic predictors.

See reply by Gilbert, et al, related article by Gilbert, et al, and correction for article by Gilbert, et al.

Predicting Suicide Attempt Risk: Logistic Regression Requires Large Sample Sizes

To the Editor: Dr Gilbert and colleagues make an important point about the difficulties in predicting suicide,1 but their conclusions do not follow as strongly from their data as they suppose. They report negative findings based on logistic regressions using clinical, demographic, and cognitive predictor variables. However, their data set is relatively small, with 28 events (suicide attempts) in 67 subjects; their regressions use 12 clinical and demographic predictors and, separately, 7 cognitive and demographic predictors. Simulation experiments2 have shown that logistic regression requires roughly 10 events per predictor, which would limit its use to 2—or, if stretched, 3—predictors for their dataset. The effect for the study in question is not entirely clear, but performing regressions below the advisory event per predictor threshold can bias the coefficients and distort the standard errors and could have been responsible, at least in part, for failure to reach statistical significance. The take-home message is that logistic regression requires relatively large sample sizes for proper statistical inference.3

In passing, I note that the odds ratios listed in their Table 3 are, incorrectly, copies of the Î² coefficients; the proper odds ratios can be computed by raising e, the base of the natural logarithm, to the power Î².

References

1. Gilbert AM, Garno JL, Braga RJ, et al. Clinical and cognitive correlates of suicide attempts in bipolar disorder: is suicide predictable? J Clin Psychiatry. 2011;72(8):1027-1033. PubMed doi:10.4088/JCP.10m06410

2. Peduzzi P, Concato J, Kemper E, et al. A simulation study of the number of events per variable in logistic regression analysis. J Clin Epidemiol. 1996;49(12):1373-1379. PubMed doi:10.1016/S0895-4356(96)00236-3

3. Bagley SC, White H, Golomb BA. Logistic regression in the medical literature: standards for use and reporting, with particular attention to one medical domain. J Clin Epidemiol. 2001;54(10):979-985. PubMed doi:10.1016/S0895-4356(01)00372-9

Steven C. Bagley, MS, MD

[email protected]

Author affiliation: Consultant, Palo Alto, California. Potential conflicts of interest: None reported. Funding/support: None reported.

doi:10.4088/JCP.11lr07393

Download Free PDF

Save Save