Get imputation values for survival model about jointai HOT 3 CLOSED

Thank you for your post. This is indeed a bug. The wrong elements of the data matrix are set to be monitored. I will look into fixing this. In the meantime, you can "solve" the issue by specifying the correct elements to be monitored via the "other" argument of monitor_params.

library("JointAI")

# modelspecification without actually running it
mod0 <- survreg_imp(Surv(futime, status != "alive") ~ age + sex +
                       copper + trig, models = c(copper = "lognorm", trig = "lognorm"),
                     data = subset(PBC, day == 0), n.adapt = 0)

# indices of the column and rows of the data matrix containing the variable "copper"
col <- which(colnames(mod0$data_list$M_lvlone) == "copper")
rows <- which(is.na(mod0$data_list$M_lvlone[, "copper"]))

# node to be monitored
imp_copper <- paste0("M_lvlone[", rows, ",", col, "]")

# run the model
mod6a <- update(mod0, n.adapt = 100, n.iter = 250,
                monitor_params = list(imps = TRUE, other = imp_copper))

# extract the imputed values
impDF <- get_MIdat(mod6a, m = 10, seed = 2019)

from jointai.

Thank you! Dr NErler. I am not sure of the meaning of this solution and I would like to use it on my own data with over 300000 rows and 80 columns. Is it possible to use survreg_imp/coxph_imp for multil-class (like in competing risk model) or multi-label classification? Is it possible to training survreg_imp/coxph_imp model and then use this model to impute the missing value for another dataset (like the testing set)? Thank you!

from jointai.

It is not yet possible to fit competing risk models in JointAI.
There is no out-of-the-box solution for imputing values in another dataset using the parameter estimates from the first dataset.
Missing values in covariates are imputed from their full-conditional posterior distributions, which are derived within JAGS from the joint distribution that JointAI specifies.
Imputing values in a new dataset based on the parameters from the original data would require you to do this sampling in R, using, for instance, a Metropolis-Hastings sampler because the imputation models will usually not have a closed-form.

from jointai.