[R-pkg-devel] Determine subset from glm object

Duncan Murdoch murdoch@dunc@n @ending from gm@il@com
Sun Jul 8 19:10:54 CEST 2018

On 08/07/2018 11:48 AM, Charles Geyer wrote:
> I need to find out from an object returned by R function glm with argument
> x = TRUE
> what the subsetting was.  It appears that if gout is that object, then
> as.integer(rownames(gout$x))
> is a subset vector equivalent to the one actually used.

You don't want the "as.integer".  If the dataframe had rownames to start 
with, the x component of the fit will have row labels consisting of 
those labels, so as.integer may fail.  Even if it doesn't, the rownames 
aren't necessarily sequential integers.   You can index the dataframe by 
the character versions of the default numbers, so simply
rownames(gout$x) should always work.

More generally, I'm not sure your question is well posed.  What do you 
mean by "the subsetting"?  If you have something like

df <- data.frame(letters, x = 1:26, y = rbinom(26, 1, 0.5))

df1 <- subset(df, letters > "b" & letters < "y")

gout <- glm(y ~ x, data = df1, subset = letters < "q", x = TRUE)

the rownames(gout$x) are going to be numbers for rows of df, because df1 
will get a subset of those as row labels.

> I do also have the call to glm (as a call object) so can determine the
> actual subset argument, but this seems to be not so useful because I don't
> know the length of the original variables before subsetting.

You should be able to evaluate the subset expression in the environment 
of the formula, i.e.

eval(gout$call$subset, envir = environment(gout$formula))

This may give incorrect results if the variables used in subsetting 
aren't in the dataframe and have changed since glm() was called.

> So now my questions.  Is this idea above (using rownames) OK even though I
> cannot find where (if anywhere) it is documented?  Is there a better way?
> One more guaranteed to be correct in the future?

I would trust evaluating the subset more than grabbing row labels from 
gout$x, but I don't know for sure it is likely to be more robust.

Duncan Murdoch

More information about the R-package-devel mailing list