Course Level: Intermediate
Format Requirements: Attendees should have experience with regression models, especially logistic and Cox proportional hazards regression models. Familiarity with R or S-PLUS software would also be helpful, but is not required. The course will be a mix of didactics and demonstrations.
Background: This course will cover the importance of prediction models in medical decision making, traditional and novel methods for their evaluation, and cost effective deployment methods. Attendees are expected to leave the course with improved skills in each of these domains.
Description and Objectives: Prediction models are proliferating and gaining increasing acceptance in clinical decision making. Physicians look to us, as analysts, to guide them on which model should be used in the clinic. At the same time that models have proliferated, several different error measures have been introduced. Traditional biostatistical measures include indicators of overall predictive performance, such as explained variation, calibration and discrimination. Novel measures include reclassification and integrated discrimination, as well as some measures explicitly based on decision-analytic principles, such as net benefit and decision curve analysis. The proliferation of measures has made it more difficult to evaluate prediction models, compare rival prediction models, or assess the value of an additional predictor, such as a molecular marker.
The objectives of this course are:
- Increase appreciation and utilization of medical prediction models.
- Learn more insightful ways of analyzing the accuracy of prediction models, both from a traditional biostatistical point of view and a decision analytic point of view. There will be a focus on novel decision-analytic tools.
- Learn a new tool for rapid web-based deployment of prediction models.