We refer to these estimates of association as
synthetic values, since they involve only
the linear predictor, and not the outcome.
They exploit mathematical associations which hold for certain models,
e.g. between R-squared and a certain chiquare test of association in the
linear model, and assume that the same holds in a Cox model where said
test is readily available but not a simple R-square computation.
R.D is the value that corresponsds the Royston and Sauerbrei
\(D\) statistic. R.KO is the value proposed by Kent and
O'Quigley, R.N is the value proposed by Nagelkerke, and
C.GH corresponds to Goen and Heller's concordance measure.
An adjustment for D is based on the ratio
r= (number of events)/(number of coefficients). For models which have
sufficient sample size (r>20) the adjustment will be small.
The Nagelkerke value is the Cox-Snell R-squared divided by a scaling
constant. The two separate values are present in the result of
summary.coxph as a 2 element vector rsq, and were listed as
"Rsquare" and "max possible" in older versions of the print routine.
(Since superseded in the default printout by the concordance.)
The Nagelkerke estimate is not returned when newdata is present.