Prediction of survival probabilities

[berithunsdieck]

Lets assume I have a survival model surv_model and a joint model joint_mod, where the surv_model is given as input. Shouldn't the prediction of the risks using the joint model be at least as good as the prediction of the risks using only the survival model?

For predicting these I use this code:

# for base survival model:
surv_probs <- pec::predictSurvProb(surv_model, newdata=input_survival_data,times=t0+12)

# for joint model: 
predSurv <- predict(model_of_interest, 
                    newdata = long_data%>%mutate(status2=0,total_months=t0), 
                    process = "event",
                    times = t0+12,
                    return_newdata = TRUE,type="subject_specific")

but looking e.g. at the AUC/ROC-curve, the prediction is way worse than the one using the survival model. How can I improve it? Am I using the right funtions for predicting resp. comparing?

[drizopoulos]

This question does not relate to an issue with the underlying code of the package.

[berithunsdieck]

@drizopoulos The code of the package is used for generating the joint model. Shouldn't the joint model be at least as good as the survival model you put in for fitting the joint model?

[drizopoulos]

Not if the longitudinal outcome is not a good predictor of the survival outcome.

[berithunsdieck]

@drizopoulos But shouldn't then be the estimate of the longitudinal outcome be set to 0 during the estimating process?

[drizopoulos]

The new version should resolve this.