Events and Seminars : 2013 Seminars

RISK ESTIMATION AND PROBABILITY MACHINES

TUESDAY, OCTOBER 22, 2013
2:00 p.m.– 3:00 p.m, CRB 692

RISK ESTIMATION AND PROBABILITY MACHINES

Logistic regression is a standard technique in biomedical research for classification and probability estimation, given a list of features (predictors) and a binary outcome (tumor, not tumor). However it assumes the model connecting the features with the outcome is known, and all necessary interaction terms are included. Recently it has been shown that entirely model-free optimal probability estimation is possible, using any consistent nonparametric regression scheme.
These are called probability machines. More recently, probability machines have been deployed for estimating the risk effects for any features or exposure terms. These are called risk machines. If the logistic model is in fact correct the risk estimates for exposure terms match the coefficients in the model, while the risk machine is still optimal with no model specification. Examples are given and many other applications using probability machines are discussed.