Predictive accuracy measures for binary outcomes: Impact of incidence rate and optimization techniques.

摘要

Evaluating the performance of models predicting a binary outcome can be done using a variety of measures. While some measures intend to describe the model's overall fit, others more accurately describe the model's ability to discriminate between the two outcomes. If a model fits well but doesn't discriminate well, what does that tell us? Given two models, if one discriminates well but has poor fit while the other fits well but discriminates poorly, which of the two should we choose? The measures of interest for our research include the area under the ROC curve, Brier Score, discrimination slope, Log-Loss, R2 and Fbeta. To examine the underlying relationships among all of the measures, real data and simulation studies are used.;The real data comes from multiple cardiovascular research studies and the simulation studies are run under general conditions and also for incidence rates ranging from 2% to 50%. The results of these analyses provide insight into the relationships among the measures and raise concern for scenarios when the measures may yield different conclusions. The impact of incidence rate on the relationships provides a basis for exploring alternative maximization routines to logistic regression.;While most of the measures are easily optimized using the Newton-Raphson algorithm, the maximization of the area under the ROC curve requires optimization of a non-linear, non-differentiable function. Usage of the Nelder-Mead simplex algorithm and close connections to economics research yield unique parameter estimates and general asymptotic conditions. Using real and simulated data to compare optimizing the area under the ROC curve to logistic regression further reveals the impact of incidence rate on the relationships, significant increases in achievable areas under the ROC curve, and differences in conclusions about including a variable in a model.