predict.coxph.penal ignores newdata when the model has only a sparse term

[camdenlo]

This is closely related to #70 and also probably has minimal impact.

In predict.coxph.penal, when there is only a sparse /penalty term in the model, the newdata argument (if supplied) gets ignored in calculating the predictions (except for the assignment line 45, which gets overwritten).

survival/R/predict.coxph.penal.S

Lines 41 to 54 in a676f5e

	if (nvar==1) {
	# Only the sparse term!
	if (!missing(newdata)) {
	n <- nrow(as.data.frame(newdata))
	pred <- rep(0,n)
	}
	se <- sqrt(object$fvar[indx])
	pred <- object$linear.predictor
	if (type=='risk') pred <- exp(pred)
	if (type=='expected') {
	pred <- object$y[,ncol(object$y)] -object$residuals
	se.fit=FALSE
	}
	}

Currently, if newdata has been supplied, an error stops execution at line 47 unless "indx" has been defined outside the function (issue #70), but when that is fixed, the resulting predictions ignore newdata.

[therneau]

What do you want the prediction to be from a frailty model with new data?
I'm not trying to be difficult -- I have struggled with this question myself. Say the model has frailty(inst) to have a random effect per institution. If the new data contains a new institution -- give NA or the mean of the random effects distribution? Can I count on random effects having a consistent label?

[camdenlo]

I was thinking that the behavior would be the same as when there are fixed covariates in addition to the frailty term — that the predictions would be for a random effect of zero. In the case of a model with only a frailty term, the only part of newdata that would matter for predictions (and only if type="expected") would be the follow-up time. Rows of newdata that differed only in the value of the frailty term would get the same prediction.