trouble automating formula edits when log or * are present; update trouble

Michael:

m2 is a model fit, not a formula. So I don't think what you suggested will work.

However, I think your idea is a good one. The trick is to protect the
model specification from evaluation via quote(). e.g.

z <- ?deparse(quote(lm(y~x1)))
z

[1] "lm(y ~ x1)"

Then you can apply your suggestion:

w <-gsub("x1","log(x1)",z)
w

[1] "lm(y ~ log(x1))"

eval(parse(text=w))

Call:
lm(formula = y ~ log(x1))

Coefficients:
(Intercept) ? ? ?log(x1)
? -0.04894 ? ? ?0.36484


The gsub() would make the substitution wherever "x1" appeared in the
model formula, thus fulfilling the OP's request.

Two comments:

1. update() behaves as documented. It is a formula update method, not
a macro substitution procedure.

2. I believe this illustrates a legitimate violation of the "avoid the
eval(parse)) construction" precept. However, I may be wrong about this
and would welcome being corrected and shown a better alternative.

Cheers,
Bert





On Tue, May 29, 2012 at 9:31 AM, R. Michael Weylandt
<michael.weylandt at gmail.com> wrote:

Hi Paul,

I haven't quite thought through this yet, but might it not be easier
to convert your formula to a character and then use gsub et al on it
directly?

Something like this

# Using m2 as you set up below
m2 <- lm(y ~ log(x1) + x2*x3, data=dat)

f2 <- formula(m2)

as.formula(paste(f2[2], f2[1],gsub("x1", "x1c", as.character(f2[3]))))

It's admittedly unwieldy, but it seems pretty robust.

Something like:

changeFormula <- function(form, xIn, xOut){
? ?as.formula(paste(form[2], form[1], gsub(xIn, xOut, as.character(form[3]))))
}

changeForm(formula(m2), "x1", "x1c")

I'm not sure if this will play nice with environments and what not so
you might need to change those manually.

Hope this gets you started,
Michael

On Tue, May 29, 2012 at 11:43 AM, Paul Johnson <pauljohn32 at gmail.com> wrote:

Greetings

I want to take a fitted regression and replace all uses of a variable
in a formula. For example, I'd like to take

m1 <- lm(y ~ x1, data=dat)

and replace x1 with something else, say x1c, so the formula would become

m1 <- lm(y ~ x1c, data=dat)

I have working code to finish that part of the problem, but it fails
when the formula is more complicated. If the formula has log(x1) or
x1:x2, the update code I'm testing doesn't get right.

Here's the test code:

##PJ
## 2012-05-29
dat <- data.frame(x1=rnorm(100,m=50), x2=rnorm(100,m=50),
x3=rnorm(100,m=50), y=rnorm(100))

m1 <- lm(y ~ log(x1) + x1 + sin(x2) + x2 + exp(x3), data=dat)
m2 <- lm(y ~ log(x1) + x2*x3, data=dat)

suffixX <- function(fmla, x, s){
? ?upform <- as.formula(paste0(". ~ .", "-", x, "+", paste0(x, s)))
? ?update.formula(fmla, upform)
}

newFmla <- formula(m2)
newFmla
suffixX(newFmla, "x2", "c")
suffixX(newFmla, "x1", "c")

The last few lines of the output. See how the update misses x1 inside
log(x1) or in the interaction?

newFmla <- formula(m2)
newFmla

y ~ log(x1) + x2 * x3

suffixX(newFmla, "x2", "c")

y ~ log(x1) + x3 + x2c + x2:x3

suffixX(newFmla, "x1", "c")

y ~ log(x1) + x2 + x3 + x1c + x2:x3

It gets the target if the target is all by itself, but not otherwise.

After messing with this for quite a while, I conclude that update was
the wrong way to go because it is geared to replacement of individual
bits, not editing all instances of a thing.

So I started studying the structure of formula objects. ?I noticed
this really interesting thing. the newFmla object can be probed
recursively to eventually reveal all of the individual pieces:

newFmla

y ~ log(x1) + x2 * x3

newFmla[[3]]

log(x1) + x2 * x3

newFmla[[3]][[2]]

log(x1)

newFmla[[3]][[2]][[2]]

x1

So, if you could tell me of a general way to "walk" though a formula
object, couldn't I use "gsub" or something like that to recognize each
instance of "x1" and replace with "x1c"??

I just can't figure how to automate the checking of each possible
element in a formula, to get the right combination of [[]][[]][[]].
See what I mean? I need to avoid this:

newFmla[[3]][[2]][[3]]

Error in newFmla[[3]][[2]][[3]] : subscript out of bounds

pj

--
Paul E. Johnson
Professor, Political Science ? ?Assoc. Director
1541 Lilac Lane, Room 504 ? ? Center for Research Methods
University of Kansas ? ? ? ? ? ? ? University of Kansas
http://pj.freefaculty.org ? ? ? ? ? ?http://quant.ku.edu

--

Bert Gunter
Genentech Nonclinical Biostatistics

Internal Contact Info:
Phone: 467-7374
Website:
http://pharmadevelopment.roche.com/index/pdb/pdb-functional-groups/pdb-biostatistics/pdb-ncb-home.htm