eval in parent frame and current frame - r

Question

I am looking for an elegant (and safe!) way to evaluate an amended call in the parent frame. By "amended" I mean I modified the call in such a way that it refers to something not included in parent frame but in another frame. I guess one could also say: "send something up but only for evaluation".
It is clarified what I want by the example below which works in some circumstances, but not all. The update function (stats:::update.default) uses eval and I added the weights argument with something (res) that is not in the same environment as the evaluation takes place. So I used get("res", pos = -1L) and I hope it is a safe way to refer to the environment res lives in. For models estimated with a variable as formula, both defined methods fail:
mod <- lm(mpg ~ cyl, data = mtcars)
form <- mpg ~ cyl
mod2 <- lm(form, data = mtcars)
wls1 <- function(x) {
res <- residuals(x)^2 # example
result <- update(x, weights = 1/get("res", pos = -1L))
return(result)
}
wls2 <- function(x) {
res <- residuals(x)^2 # example
result <- update(x, weights = 1/res)
return(result)
}
wls3 <- function(x) {
data(ChickWeight)
ChickWeight$cyl <- ChickWeight$weight
ChickWeight$mpg <- ChickWeight$Time
result <- update(x, data = ChickWeight)
return(result)
}
wls1(mod) # works
wls1(mod2) # errors
wls2(mod) # works
wls2(mod2) # erros
wls3(mod) # works
wls3(mod2) # works
How can this be solved in general in a safe way?
I was looking for a function that gives the current environment (something like a fictious this.environment() function) so avoid the pos argument and use the envir of get (I know I can create my own temporary environment and have res associated to it to use something like envir = my.eny).

Answer 1

We can do this by creating a quoted 'language' object for the formula and then update the call of the model
form <- quote(mpg ~ cyl)
wlsN <- function(x, formula) {
x$call$formula <- formula
res <- residuals(x)^2
update(x, weights = 1/res) # it is in the same environment. No need for get
}
wlsN(mod2, form)
#Call:
#lm(formula = mpg ~ cyl, data = mtcars, weights = 1/res)
#Coefficients:
#(Intercept) cyl
# 37.705 -2.841
-checking with other formula
form1 <- quote(disp ~ cyl + vs)
form2 <- quote(mpq ~ gear + carb)
mod1 <- lm(form1, data = mtcars)
mod2 <- lm(form2, data = mtcars)
wlsN(mod1, form1) # works
wlsN(mod2, form2) # works

Answer 2

It's hard to work around the fact that R looks for the value of weights in either data or the environment of the formula - which in the case of the variable named form in your example, is the global environment.
An alternative that riffs on the same theme as akrun's answer:
wls3 <- function(x) {
environment(x$call$formula) <- environment()
res <- residuals(x)^2
result <- update(x, weights=1/res)
}
I can see how this could get ugly in less trivial uses of this workaround such as when the formula of x already has an environment that does not enclose (potentially wrong use of the term) the environment in the call to wls3().
Another alternative (not recommended) is to use assign, e.g.
wls4 <- function(x) {
assign('res', residuals(x)^2, envir=environment(formula(x)))
result <- update(x, weights=1/res)
}
however this has the unintended consequence of leaving the variable res in the global environment.