Cutpoint analysis for models with binary outcomes: a case study on branch mortality

Models of binary outcomes are commonly used in forestry, but the predictions errors of these types of models are difficult to present effectively. In addition, most studies generally use a fixed value of 0.5 as the separation between events and non-events. The use of cutpoint analysis has been widely utilized in the health sciences and other fields, while it is relatively uncommon in the forestry literature. Cutpoint analysis involves locating the optimal value that minimizes prediction errors associated with binary outcomes. This case study illustrates the use of cutpoint analysis to improve a dynamic model of individual branch mortality. In this study, the use of cutpoint analysis increased the model specificity (prediction of events) from 77.8% (standard cutpoint of 0.5) to 90.3% (optimal cutpoint of 0.672). At the same time, the sensitivity of the model decreased only slightly and the false positive rate (non-event predicted as an event) was greatly decreased from 22.2 to 9.7%. In addition, the use of receiver operating characteristics (ROC) curves was an effective approach for evaluating prediction errors of models of binary outcomes. Cutpoint analysis is a simple yet effective method for improving predictions of binary outcomes and should be used more regularly, particularly when modelling the binary outcome of rare events such as mortality.