I'm trying to generalize the MASS function "stepAIC" to study the effects of hierarchy for both interactions and parabolic terms. I include a new argument 'hierarchy = c("include", "exclude", "ignore")' being a 2-vector, with the first component for interaction and the second for parabolic terms:
* "exclude" fails to consider A:B (or A^2) unless both A and B are currently in the model. This is the current default for interactions but not for parabolic terms.
* "include" forces A:B to be evaluated with 3 degrees of freedom if neither A nor B are currently in the model, 2 degrees of freedom if only 1 are in the model, and 1 degree of freedom if both are currently in the model (assuming A and B are both either numerical variables or 2-level factors). George Box says he invented response surface methods for a chemical plant that had only interactions. The plan had gone through several general managers, who had done what they could with one factor at a time experiments. I plan to analyze several Taguchi 3-level experiments in order to evaluate effect sparcity and the frequency with which both interactions and parabolic terms occur in industrial experimentation, and I need an automated way of doing this.
* "ignore" means include or exclude the interaction independent of whether the main effect is in the model. This is the current default for A^2.
I want to make "include" the default for both A:B and A^2. In my revision of stepAIC, I call match.arg as follows;
hierarchy <- rep(match.arg(hierarchy), length=2)
What do you think? I know that one example does not justify adding a rarely used option to a general tool. However, it occurred to me that there might be other applications
#################
Thanks for reminding me about the "seq(length= length(arg))" construction. After I sent that email, I discovered that problem, and added the following line right before the "for" loop:
if((length(arg)==length(choices)) && all(arg==choices))
return(choices[1])Since I had heard no reply, I decided not to issue a "correction" to an email that might go nowhere or might only irritate the recipients.
Thanks again for your comments and all your effort and creativity in the R project.
Best Wishes, Spencer Graves
Martin Maechler wrote:
"Spencer" == Spencer Graves <[EMAIL PROTECTED]> on Mon, 17 Feb 2003 17:43:23 -0800 writes:
Spencer> Hello: I'm not on the "r-devel" list, but I just Spencer> modified "match.arg" from R 1.6.2 for Windows to Spencer> accept a vector for "arg".
Spencer> Is it appropriate that I send such things to Spencer> this email address for consideration for inclusion Spencer> in a future release?
yes (using "R-devel" for ...). But why is this current proposal really useful?
Note that I can't see reason to match more than one argument to the
*same* list of choices. `choices' already is a vector typically...
Making a function more complicated makes it also more error-prone, so I think we'd need a good motivation for it. E.g., your proposal below quite badly fails when length(arg) == 0, since you've used the ``well-known to be unsafe'' 1:length(arg) idiom instead of the safe seq(length= length(arg)) one.
Regards, Martin
Spencer> I just compared this with "match.arg" in Spencer> S-Plus 6.1 for Windows. There, I got the Spencer> following:
>> match.arg(c("a","b"), c("aa", "bb")) Spencer> [1] "aa" "bb"
Spencer> However, match.arg(c("a", "b")) in a test Spencer> function with "default" = c("aa", "bb") returned Spencer> only "a"; the following returns c("aa", "bb").
Spencer> Thanks for all your hard work in developing this Spencer> marvelous product.
Spencer> Sincerely, Spencer Graves
match.arg <-
function (arg, choices)
{
if (missing(choices)) {
formal.args <- formals(sys.function(sys.parent()))
choices <- eval(formal.args[[deparse(substitute(arg))]])
}
# cat("choices =", choices, "\n")
for(j in 1:length(arg)){
if (all(arg[j] == choices))
arg[j] <- choices[1]
else{
i <- pmatch(arg[j], choices)
if (is.na(i))
stop(paste("ARG should be one of", paste(choices, collapse = ", "),
sep = " "))
if (length(i) > 1)
stop("there is more than one match in match.arg")
arg[j] <- choices[i]
}
}
arg
}
______________________________________________ [EMAIL PROTECTED] mailing list http://www.stat.math.ethz.ch/mailman/listinfo/r-devel
