Bias Plus Variance Decomposition for Survival Analysis Problems

For classification problems, it is well known that bias and variance components of the estimation prediction error combine to influence classification in a very different way, and have different importance depending on the sample size. For small and for high-dimensional datasets, variance of the prediction caused by variations in the training samples makes largest contribution into the expected prediction error. For large datasets, bias becomes more important component of the error [1]. Thus, the decomposition of expected error into bias and variance parts is an important tool to understand differences between the algorithms, to find areas of the optimal application. To the best of author's knowledge, such decomposition was not proposed for survival analysis problem. Here we describe an approach to define this decomposition for this class of problem. On two real life datasets we study how bias and variance of we show how regularization and size of the training sample affect bias, variance and overall errors of the methods. 1 2 Bias - Variance Decomposition