03.01.2012
11:00 am
Thomas Alexander Gerds, PH.D.
Department of Biostatistics
University of Copenhagen
THURSDAY, MARCH 1, 2012
11:00 a.m.– 12:00 pm
Clinical Research Building CRB 988
1120 NW 14th Street
Miami, Florida
Traditional likelihood methods and
machine learning tools provide many alternative strategies for building a
risk prediction model based on training data. The estimation of
prediction performance, however, is a challenging task in the absence of
independent validation data. Two relevant questions are:
(1) What is the best prediction model based on my training data?
(2) What statistical strategy finds the best prediction model?
This talk discusses popular
cross-validation estimates with regard to these questions. The main tool
is a decomposition of the expected Brier score into model accuracy and
model uncertainty. Repeated splits of the training data can be used to
estimate these terms and to compare and test alternative statistical
modeling strategies. The talk is illustrated with examples from medical
statistics.