Summary: Bioinformatics Advance Access published June 6, 2007
Bøvelstad et al
is important in our context. This ensures a fair comparison bet-
ween methods and also that the methods are tuned to predict well
on novel data rather than on the training data themselves. In this
paper, all tuning parameters are determined by partial likelihood
cross-validation (Verweij and van Houwelingen, 1993).
In assessing the performance of a prediction rule, we need to train
and test the method on separate data. Failure to do so would favor
methods that fit well to the specific data at hand rather than to novel
data from the same population. We also need to take into account the
variation in prediction performance that would result from choosing
a different split of the data into a training data set and a test data
set. More generally, performance assessment for the prediction rules
must account for all model selections and training decisions made.
In this paper, we compare the various strategies on three well
known data sets of which two are on breast cancer and one is on
diffuse large-B-cell lymphoma, see van de Vijver et al. (2002), van't
Veer et al. (2002), Chang et al. (2005), van Houwelingen et al.
(2006), Rosenwald et al. (2002), Sørlie et al. (2003). We find that