model.matrix.default chokes on backquote (PR#7202) - R-devel

Sat, Aug 28, 2004 10:15 AM #

From:   	Peter Dalgaard <p.dalgaard@biostat.ku.dk>

"Gabor Grothendieck" <ggrothendieck@myway.com> writes:

ggrothendieck@myway.com writes:

The following gives an error:

`a(b)` <- 1:4
`c(d)` <- (1:4)^2
lm(`a(b)` ~ `c(d)`)

Error in model.matrix.default(mt, mf, contrasts) :
model frame and formula mismatch in model.matrix()

To fix it replace this line in model.matrix.default:

reorder <- match(attr(t, "variables")[-1], names(data))

with these two lines:

strip.backquote <- function(x) gsub("^`(.*)`", "\\1", x)
reorder <- match(strip.backquote(attr(t, "variables"))[-1],
strip.backquote(names(data)))

Hmm.. Yes, there's a bug (and it's likely not the only one we have
relating to odd variable names in model formulas), but I suspect that
the fix is wrong.

The backquotes are not part of the variable names, but get added by
deparsing -- sometimes! Other times they do not: Try for instance
as.character(quote(`a(b)`)). (Which is as it should be. Other pieces
of logic relating to nonsyntactical names represent some rather
awkward compromises.)

When backquotes have found their way into names(data) or the
"variables" attribute, I would rather suspect that they were created
by the wrong tool and fix that, not cure the symptom by stripping them
off at a later stage.

In model.frame.default there is a line:

varnames <- as.character(vars[-1])

that turns part of a call object, vars, into a character string.
We could change that to:

varnames <- strip.backquote(as.character(as.list(vars[-1])))

or perhaps as.character should not return the backquotes in the
first place in which case the fix would be to fix as.character.

Or not use it in this way. I forget what the reasoning was behind the
current behaviour of as.character, but the point is that

as.character(attr(terms(`a(b)`~`c(d)`),"variables"))

[1] "list" "`a(b)`" "`c(d)`"

whereas for instance

sapply(attr(terms(`a(b)`~`c(d)`),"variables")[-1],as.character)

[1] "a(b)" "c(d)"

1. That is quite subtle but a fix based on that would appear to 
   solve it.

2. Your example and possibly some verbiage should be added to 
   ?as.character .  

3. In looking for the offending spot, I seem to remember (though I did 
   not keep track of it) that one or more of lm, model.frame.default, 
   terms.formula, etc. had additional applications of as.character 
   directly to a list as in your first example and these should 
   probably be changed to correspond to your second example, as 
   well, where as.character is applied to the elements of the 
   list rather than the lsit itself.

Peter Dalgaard

Sun, Aug 29, 2004 6:53 PM #

"Gabor Grothendieck" <ggrothendieck@myway.com> writes:

Hmm, not quite. I tried, and terms like offset(foo) gets me in
trouble. Probably, I was fixing the wrong end of the original problem:
In the comparisons, we can't have one side with backquotes and the
other without them. That doesn't have to mean that they should be
removed from both sides, and indeed it would get us in trouble if
someone was perverse enough to do things like

 y ~ `offset(foo)` + offset(foo)

I.e. perhaps the real issue is that names(data) gets generated without
backquotes.

Anyways, this is a real can of worms and I'm not sure we're not too
close to 2.0.0 to start tampering with it...

O__  ---- Peter Dalgaard             Blegdamsvej 3  
  c/ /'_ --- Dept. of Biostatistics     2200 Cph. N   
 (*) \(*) -- University of Copenhagen   Denmark      Ph: (+45) 35327918
~~~~~~~~~~ - (p.dalgaard@biostat.ku.dk)             FAX: (+45) 35327907