Group-updated `bcoxph` models fail when cross-validation with `cv.bh` about bhglm HOT 4 CLOSED

The root cause of the problem is the artificial offset induced when fitting group-updated models, which will trigger the wrong data pull with in cv.bh at line 136.

To examine, we follow the previous example code and show group-updated models introduce artificial offset with the follows

is.null(mdl_joint$offset) #TRUE
#> Error in eval(expr, envir, enclos): object 'mdl_joint' not found
is.null(mdl_group$offset) #FALSE
#> Error in eval(expr, envir, enclos): object 'mdl_group' not found

^{Created on 2022-01-03 by the reprex package (v2.0.1)}

from bhglm.

Suggested solution:
We can re-assign the correct offset after fitting coxph internally. For example, we can add a line of code

fit$offset<- offset

at bcoxph.r line 436

Technically, there is a correction of offset implemented in bcoxph with on Line 266

However, it doesn't work when offsets are 0.

from bhglm.

This bug does not have an impact on prediction via predict function. Specifically, predict.coxph in the survival package, check attributes to decide if offset should be included in prediction,

 has.offset <- !is.null(attr(Terms, 'offset'))
 if (is.null(object[['x']]) || has.weights || has.offset || # the remaining code is omited

In our case, the Terms are corrected via

To demonstrate there is no impact on prediction, here is the example (following previous code)

is.null(attr(mdl_joint, "offset")) #TRUE
is.null(attr(mdl_group, "offset")) #TRUE

is.null(attr(mdl_joint$terms, "offset")) #TRUE
is.null(attr(mdl_group$terms, "offset")) #TRUE


## Examination of prediction
cor(
  data.frame(
  joint_mdl = predict(mdl_joint, type="expected"),
  group_mdl = predict(mdl_group, type="expected")
  )
) # Correlation is 0.99

from bhglm.

The problem resolved via f0f80cc

from bhglm.