how to suppress the intercept in an lm()-like formula method?

On 1/29/2013 9:25 AM, Duncan Murdoch wrote:

On 29/01/2013 9:14 AM, Michael Friendly wrote:

To partly answer my own question:  It wasn't that hard to hack the
result of model.matrix() to remove the intercept,

remove.intercept <- function(x) {
    if (colnames(x)[1] == "(Intercept)") {
        x <- x[,-1]
        attr(x, "assign") <- attr(x, "assign")[-1]
    }
    x
}

I think you need to do some of the low level calculations yourself,
specifically the "terms" calculation.

For example, this forces no intercept, regardless of what the user
specifies.  You might prefer just to change the default to no
intercept and allow the user to use "+1" to add one, which looks
harder...

# .... set formula to the user's formula ...
# Now modify it to suppress the intercept:

class(formula) <- c("nointercept", class(formula))
terms.nointercept <- function(x, ...) {
  result <- NextMethod(x, ...)
  attr(result, "intercept") <- 0
  result
}

That's very clever.  I think I can use this locally in my function.
However, it is still a mystery to me *where* in the process terms()
gets called.  The main incantation for model formulae in lm() is below,
so who calls terms()?

      cl <- match.call()
      mf <- match.call(expand.dots = FALSE)
      m <- match(c("formula", "data", "subset", "weights", "na.action"),
names(mf), 0L)
      mf <- mf[c(1L, m)]

      mf[[1L]] <- as.name("model.frame")
      mf <- eval(mf, parent.frame())

      mt <- attr(mf, "terms")

It is also a mystery to me where na.action takes its effect. Presumably
this is somewhere before lm.fit(x, y, ...)
gets called, but I'd like to take account of this in my cancor.default
method.

-Michael

Now lm(formula) does a fit with no intercept.

Duncan Murdoch

However, the model frame and therefore the model terms stored in the
object are wrong, still including the intercept:

Browse[1]> str(mt)
Classes 'terms', 'formula' length 3 cbind(SAT, PPVT, Raven) ~ n + s + ns
+ na + ss
    ..- attr(*, "variables")= language list(cbind(SAT, PPVT, Raven), n,
s, ns, na, ss)
    ..- attr(*, "factors")= int [1:6, 1:5] 0 1 0 0 0 0 0 0 1 0 ...
    .. ..- attr(*, "dimnames")=List of 2
    .. .. ..$ : chr [1:6] "cbind(SAT, PPVT, Raven)" "n" "s" "ns" ...
    .. .. ..$ : chr [1:5] "n" "s" "ns" "na" ...
    ..- attr(*, "term.labels")= chr [1:5] "n" "s" "ns" "na" ...
    ..- attr(*, "order")= int [1:5] 1 1 1 1 1
    ..- attr(*, "intercept")= int 1
    ..- attr(*, "response")= int 1
    ..- attr(*, ".Environment")=<environment: R_GlobalEnv>
    ..- attr(*, "predvars")= language list(cbind(SAT, PPVT, Raven),
n, s,
ns, na, ss)
    ..- attr(*, "dataClasses")= Named chr [1:6] "nmatrix.3" "numeric"
"numeric" "numeric" ...
    .. ..- attr(*, "names")= chr [1:6] "cbind(SAT, PPVT, Raven)" "n" "s"
"ns" ...
Browse[1]>


On 1/29/2013 8:44 AM, Michael Friendly wrote:

I'm trying to write a formula method for canonical correlation

analysis,

that could be called similarly to lm() for
a multivariate response:

cancor(cbind(y1,y2,y3) ~ x1+x2+x3+x4, data=, ...)
or perhaps more naturally,
cancor(cbind(y1,y2,y3) ~ cbind(x1,x2,x3,x4), data=, ...)

I've adapted the code from lm() to my case, but in this situation, it
doesn't make sense to
include an intercept, since X & Y are mean centered by default in the
computation.

In the code below, I can't see where the intercept gets included in

the

model matrix and therefore
how to suppress it.  There is a test case at the end, showing that the
method fails when called
normally, but works if I explicitly use -1 in the formula. I could

hack

the result of model.matrix(),
but maybe there's an easier way?

cancor <- function(x, ...) {
     UseMethod("cancor", x)
}

cancor.default <- candisc:::cancor

# TODO: make cancisc::cancor() use x, y, not X, Y
cancor.formula <- function(formula, data, subset, weights,
         na.action,
         method = "qr",
     model = TRUE,
     x = FALSE, y = FALSE, qr = TRUE,
     contrasts = NULL, ...) {

     cl <- match.call()
     mf <- match.call(expand.dots = FALSE)
     m <- match(c("formula", "data", "subset", "weights",

"na.action"),

names(mf), 0L)
     mf <- mf[c(1L, m)]

     mf[[1L]] <- as.name("model.frame")
     mf <- eval(mf, parent.frame())

     mt <- attr(mf, "terms")
     y <- model.response(mf, "numeric")
     w <- as.vector(model.weights(mf))
     if (!is.null(w) && !is.numeric(w))
         stop("'weights' must be a numeric vector")

     x <- model.matrix(mt, mf, contrasts)
     # fixup to remove intercept???
     z <- if (is.null(w))
         cancor.default(x, y,  ...)
     else stop("weights are not yet implemented")  # lm.wfit(x, y,

w,  ...)

     z$call <- cl
     z$terms <- mt
         z
}

TESTME <- FALSE
if (TESTME) {

# need to get latest version, 0.6-1 from R-Forge
install.packages("candisc", repo="http://R-Forge.R-project.org")
library(candisc)

data(Rohwer)

# this bombs: needs intercept removed
cc <- cancor.formula(cbind(SAT, PPVT, Raven) ~  n + s + ns + na + ss,
data=Rohwer)
## Error in chol.default(Rxx) :
##  the leading minor of order 1 is not positive definite

#this works as is
cc <- cancor.formula(cbind(SAT, PPVT, Raven) ~  -1 + n + s + ns + na +
ss, data=Rohwer)
cc
## Canonical correlation analysis of:
##       5   X  variables:  n, s, ns, na, ss
##   with        3   Y  variables:  SAT, PPVT, Raven
##
##     CanR  CanRSQ   Eigen percent    cum
## 1 0.6703 0.44934 0.81599   77.30  77.30
## 2 0.3837 0.14719 0.17260   16.35  93.65
## 3 0.2506 0.06282 0.06704    6.35 100.00
##
## Test of H0: The canonical correlations in the
## current row and all that follow are zero
##
   ...
}