R devel - Sep 2024 - Possible update to survival

I got good feedback from the list about a scope issue, so I am coming back for
more.
Prior issue:   users who type survival::coxph(survival::Surv(time, status) ~ x1
+ x2 + surv ival::strata(group), data=mydata)
This messes up the character string matching for strata, done via   tt <-
terms(formula, specials= ?strata?).  The code runs, and gives the wrong answer
(group is treated as an ordinary covariate).

The solution

  1.  Add code very early that rewrites the formula without the offendinge
survival:: string
  2.  Before calling model.frame(), insert my copy of strata into the calling
chain
    coxenv <- new.env(parent= environment(formula))
    assign("strata", survival::strata, envir= coxenv)
    environment(formula) <- coxenv

The second part of this deals with those who want to use survival functions but
not use library(survival).


Now to the question.
I had always (naively) assumed that the Surv(), strata(), etc calls within a
coxph formula would be resolved first within the survival namespace.  I realize
now that this is not true: they get resoloved within model.frame.  If a user has
their own Surv, it will get used.
For ultimate safety, I am thinking of extending the above to all of the internal
survival functions that might be used in a formula: Surv, strata, pspline,
cluster, ratetable  (I think that?s all).   An intitial limited test looks okay,
but before anything migrates to CRAN I am looking for any feedback.   The
current test version also prints a warning when removing the extraneous
?survival::? modifier, which I will retain for at least a little while although
I personally detest receiving such messages.

Terry T.

	[[alternative HTML version deleted]]

? Sun, 15 Sep 2024 00:43:31 +0000
"Therneau, Terry M., Ph.D. via R-devel" <r-devel at
r-project.org> ?????:
>   2.  Before calling model.frame(), insert my copy of strata into the
> calling chain coxenv <- new.env(parent= environment(formula))
>     assign("strata", survival::strata, envir= coxenv)
>     environment(formula) <- coxenv
<...>
> For ultimate safety, I am thinking of extending the above to all of
> the internal survival functions that might be used in a formula:
> Surv, strata, pspline, cluster, ratetable  (I think that_s all).   An
> intitial limited test looks okay, but before anything migrates to
> CRAN I am looking for any feedback.
What do you think of the following approach?

When changing the environment of the formula, construct the following
environment chain:

1. Top: your 'coxenv' environment with the special survival functions
2. Enclosing environment: list2env(data)
3. Enclosing^2 environment: original environment(formula)

Since the environment chain constructed by eval() when called by
model.frame() looks different (top: data -> enclosure:
environment(formula)), someone truly determined to shoot themselves in
the foot could still sneak a 'strata' function inside their
'data'
argument.

Having said that, what you are planning to implement may be already
reliable enough.

-- 
Best regards,
Ivan