[R] Fwd: Classification error rate increased by bagging - any ideas?

[Anthony Staines]

Hi,

I'm analysing some anthropometric data on fifty odd skull bases. We know the
gender of each skull,  and we are trying to develop a predictor to identify the
sex of unknown skulls.

Rpart with cross-validation produces two models - one of which predicts gender
for Males well, and Females poorly, and the other does the opposite (Females
well, and Males poorly). In both cases the error rate for the worse predicted
gender is close to 50%, and for the better predicted gender about 15%.

Bagging tree models produces a model which classifies both males and
females equally well (or equally poorly), but has an overall error rate
(just over 30%) higher than either of the rpart models (about 25%).

My instinct is to go for the bagging results, as they seem more reasonable,
but my colleagues really like the lower overall error rate. Any thoughts?

Ta,
Anthony Staines--
Dr. Anthony Staines, Senior Lecturer in Epidemiology.
School  of Public Health and Population Sciences, UCD, Earlsfort
Terrace, Dublin 2, Ireland.
Tel:- +353 1 716 7345. Fax:- +353 1 716 7407 Mobile:- +353 86 606 9713
Web:- http://phm.ucd.ie