predict.crr() method not working with crr2

[jquach410]

Hi there,

I was wondering if you have attempted predict() from the crr2() output? There doesn't seem to be a method for it!

Thanks!

[raredd]

Sorry, no I haven't thought much about that. There is a predict method from cmprsk::predict.crr, and there is an example for how to use it in cmprsk::crr; unfortunately, it doesn't seem to be very user-friendly like other things in the cmprsk package. The example there is only for numeric predictors, so if you are using categorical predictors there is a bit more nuance to getting it to work.

Namely, for predict.crr(object, cov1, cov1), cov1 has to be in the same order and format as the original call to crr. With crr2 this model matrix is done for you, so you just need to make sure you know how to input the cov1 into predict. For example,

library('cmprsk2')
tp <- within(transplant, {
  event_ind <- as.integer(event) - 1L
  year <- NULL
})
cr <- crr2(Surv(futime, event_ind(0) == 1) ~ age + sex + abo, tp)

pt <- head(attr(cr$`CRR: 1`, 'model.matrix'), 2)
pt
#   age sexf aboB aboAB aboO
# 1  47    0    1     0    0
# 2  55    0    0     0    0

This is how the model matrix was formed before fitting the model, so when you go to predict, your cov1 must be in that order (and a numeric matrix, it seems to have a problem with data frames, so just make sure cov1 is also a matrix).

So now you can pick one of the crr models and do:

pd <- predict(cr$`CRR: 1`, cov1 = pt)
plot(pd)

Or input some custom profile for cov1

Happy to hear thoughts on what a predict.crr2 method would do. Obviously it could make the user experience a bit nicer and give some better error messages/better documentation than crr. If you have one cov1, do you want it to be applied to each of the crr models and return a list of prediction matrices?

[jquach410]

Thanks for this example!

I think that a predict.crr2 method could mirror the predict.coxph method in terms of functionality. At the moment predict.crr seems to only have this output: Returns a matrix with the unique type 1 failure times in the first column, and the other columns giving the estimated subdistribution function corresponding to the covariate combinations in the rows of cov1 and cov2, at each failure time (the value that the estimate jumps to at that failure time.

It would be nice if predict.crr2 can also have other types of output such as linear predictors, risk scores, etc... in a similar type argument.

From predict.coxph documentation

• type: the type of predicted value. Choices are the linear predictor ("lp"), the risk score exp(lp) ("risk"), the expected number of events given the covariates and follow-up time ("expected"), and the terms of the linear predictor ("terms"). The survival probability for a subject is equal to exp(-expected).

In addition, if possible to not have to convert the dataframe to a matrix for the cov1 would be nice.

Your idea to return a list of prediction matrices for each crr models would be useful as well.

And yes, better documentation would be great. Would be happy to discuss further!