use of "rcorr.cens" with binary response?
Tom Vanwalleghem wrote:
Dear R-helpers, I recently switched from SAS to R, in order to model the occurrence of rare events through logistic regression. Is there a package available in R to calculate the Goodman-Kruskal Gamma? After searching a bit I found a function "rcorr.cens" which should do the job, but it is not clear to me how to define the input vectors? Is "x" a vector with the fitted probabilities and "s" a vector containing the observed response variable? Or does anybody know an alternative? Any help would be greatly appreciated, Thanks, Tom -- Tom Vanwalleghem Physical and Regional Geography, K.U.Leuven Redingenstraat 16 B-3000 LEUVEN +32(0)16/326414
If dealing with a binary response, Somers' Dxy rank correlation may have a slight advantage over gamma. You can get Dxy from somers2 in Hmisc and from that you can easily compute ROC area (Dxy = 2*(ROC - .5)). rcorr.cens thought will also give standard errors. You are right about the inputs except that x can be probability or log odds - anything that ranks the same as probability.
Frank E Harrell Jr Professor and Chair School of Medicine
Department of Biostatistics Vanderbilt University