In a study on seclusion and restraint in psychiatry in the Netherlands, Georgieva c.s. aimed to find an empirically and clinically relevant prediction model for identifying patients at risk. The authors conducted a forward stepwise logistic regression analysis, which will result in the smallest model, not necessarily a relevant model. But they claim to have found "a simple, accurate and highly predictive model". However, the data suggest a simpler, more accurate and better predictive model than the one reported.

The authors' model, including psychological impairment, uncooperative behavior and involuntary commitment, is not very accurate as the Nagelkerke R-square is .35, indicating a medium model-fit. And the model is not highly predictive. The paper states that the model predicted 72% of the cases correctly, based on the SPSS classification table in logistic regression which is supposed to show how many cases are correctly predicted. However, that model is not the best model that fits the data ...

A 72% correct classification means that we would be better of with the very simple intercept-only model, because most patients (86%) are not secluded or restrained. Predicting that a patient will not be secluded will be right not in 72% but 86% of the cases. In stead of improving the ability of professionals to predict seclusion and restraint, an open mind or positive attitude of professionals and good working conditions for staff members would be more helpful in reducing the number of seclusions.

Not familiar with the classification table? Read more here >>