Statistics Seminar - Samantha-Jo Caetano - Concordance measures for validating time-to-event models.
Speaker: Samantha-Jo Caetano (McMaster University)
Title: Concordance measures for validating time-to-event models.
In cancer research, time-to-event prediction models are often used to assess a patient’s risk of experiencing some event. Thus, prediction models are imperative to identifying risk factors and informing treatment decisions. When building a prediction model, statisticians have a multitude of models to choose from. Thus, the model validation process is crucial in ensuring that an accurate and valid prediction model is selected. The concordance (c) statistic is a commonly used model validation tool. Originally developed for logistic regression models, the c-statistic has become increasingly popular in validating time-to-event prediction models.Unlike the logistic regression setting, time-to-event outcomes have a "censor" component which must be accounted for when measuring concordance. Due to the censoring artefact of time-to-event data there is no unique definition of concordance in this setting. There have been multiple proposals of how to define and measure concordance in a time-to-event framework, but no consensus on what to use in practice. During this talk I will describe three proposed definitions of time-to-event concordance and provide the results of a comparative analysis of five time-to-event c-statistics(including two that I developed).