Multivariate prediction of smoking cessation following physician advice to quit smoking: a validation study

A multivariate predictive model was developed to classify patients with respiratory disease as to their smoking status following physician advice to quit (L.L. Pederson, J. C. Baskerville, and J. M. Wanklin, Prev. Med. 11, 536--549 (1982)). The purpose of this study was to validate this model on a new group of patients by comparing their predicted smoking behavior with their actual behavior. Using a probability of 0.50 as the cutoff for prediction, overall accuracy was 89.6%. However, the sensitivity for detecting those who would actually quit was low. By reducing the cutoff probability to 0.20, overall accuracy remained high and sensitivity was increased. A discussion of the implications of different types of classification errors is presented based on cost-effectiveness considerations. The clinical usefulness of prediction models is discussed.