Assignment 9

[Frank Harrell]

I am puzzled why model approximation did not fare well for estimating the true linear predictor X*beta in resimulation.  Can anyone postulate a reason?  One exploratory analysis that would be nice for someone to do is to model the mean absolute difference as a function of which variables are retained in the approximation.

A question for group 1: What is the purpose of approx.mod <- lrm(new.formula, data=support.new) and how was this used in your analysis?

A question for group 2: how could eliminated <- names(r2)[1 : which(abs(r2 - 0.95) <= ...)))] be correct?  which returns a vector, not a scalar.