[Rd] model.matrix evaluation challenges

Felix Andrews felix at nfrac.org
Wed Aug 12 02:50:42 CEST 2009


Hi,

OK, yes, it does require some explanation...

What is the difference between your examples 2 and 3, where the first
has the formula given explicitly, and the second has the formula
assigned to the object? As you said, case 2 works while case 3 fails.

2. with(c(newdata,object at data),model.matrix(~f))
3. z <- ~f; with(c(newdata,object at data),model.matrix(z))

The difference is that the environment attached to the formula is
different. Its environment defaults to the evaluation frame in which
the formula was defined. This evaluation frame is different inside the
with() statement: in that case the evaluation frame includes the
elements of c(newdata, object at data).

If the 'data' argument is missing, model.matrix() will only look in
the environment of z, which in your example is external to the with()
frame.

So one approach is to redefine the environment of the formula z:

z <- ~f
with(c(newdata,object at data), {
    environment(z) <- environment()
    model.matrix(z)
})

Or, we can force model.matrix() to look in a given environment by
using its 'data' argument:

with(c(newdata,object at data),model.matrix(~f, data = environment()))

Hope that helps
-Felix


2009/8/12 Ben Bolker <bolker at ufl.edu>:
>  that seems to work, thank you.
>  would anyone care to explain *why* it works ... ? or where
> I could go to read more about why it works ... ?
>
>  cheers
>    Ben Bolker
>
>
> Felix Andrews wrote:
>> how about...
>>
>> nrow(with(cc, model.matrix(params, data=environment())))
>>
>> cheers
>> -Felix
>>
>> 2009/8/10 Ben Bolker <bolker at ufl.edu>:
>>>  I am having difficulty with evaluation/environment construction
>>> for a formula to be evaluated by model.matrix().  Basically, I
>>> want to construct a model matrix that first looks in "newdata"
>>> for the values of the model parameters, then in "object at data".
>>> Here's what I've tried:
>>>
>>>  1. model.matrix(~f,data=c(newdata,object at data)) -- fails because
>>> something (terms()?) within model.matrix() tries to coerce "data"
>>> to a data frame -- but newdata and object at data have different
>>> numbers of rows.
>>>
>>>  2. with(c(newdata,object at data),model.matrix(~f))   works --
>>> BUT not when ~f is assigned to another object (as it will be
>>> when it is being dug out of a previously fit model).
>>>
>>>  3. If "f" exists as a column within newdata and/or object at data,
>>> but the formula is assigned to another object (see below), then
>>> with(c(newdata,object at data),model.matrix(z))  fails -- because
>>> model.matrix() is explicitly evaluating the variables in the formula
>>> in the environment of z (i.e., ignoring the first argument of "with" ...)
>>>
>>>
>>>  Any advice on how to solve this without making a bigger mess?
>>>
>>>  sincerely
>>>    Ben Bolker
>>>
>>>
>>> ## set up a data frame for prediction
>>>
>>> set.seed(1001)
>>> f = factor(rep(letters[1:4],each=20))
>>> x = runif(80)
>>> u = rnorm(4)
>>> y = rnorm(80,mean=2+x*(3+u[f]),sd=0.1)
>>> dat = data.frame(f,x,y)
>>>
>>> ## fit a model ... could easily do by lm() but want to
>>> ##   demonstrate the problem
>>>
>>> library(bbmle)
>>> m1 = mle2(y~dnorm(a+b*x,sd=exp(logs)),parameters=list(b~f),data=dat,
>>>  start=list(a=0,b=2,logs=-3))
>>>
>>> ## data frame for prediction
>>> pp0 = expand.grid(x=seq(0,1,length=11),
>>>  f=levels(dat$f))
>>>
>>> ## combine frame and model data: have to keep the model data
>>> ##  around, because it contain other information needed for
>>> ##  prediction.
>>>
>>> cc = c(pp0,m1 at data)
>>> try(mm <- model.matrix(~f,cc))  ## fails -- different number of rows
>>>  ## (at some point R tries to coerce cc to a data frame)
>>> nrow(with(cc,model.matrix(~f))) ## works ... but ...
>>>
>>> ## here's what happens within predict.mle2()
>>> ## extract information about parameter models
>>> params <- eval(m1 at call$parameters)
>>> params <- params[[1]]
>>> params[[2]] <- NULL
>>>
>>> nrow(with(cc,model.matrix(params)))
>>> rm(f)
>>> ## object "f" not found -- because model.matrix
>>> ##  evaluates in the parent environment of params ...
>>> try(nrow(with(cc,model.matrix(params))))
>>>
>>>
>>>
>>> --
>>> Ben Bolker
>>> Associate professor, Biology Dep't, Univ. of Florida
>>> bolker at ufl.edu / www.zoology.ufl.edu/bolker
>>> GPG key: www.zoology.ufl.edu/bolker/benbolker-publickey.asc
>>>
>>>
>>> ______________________________________________
>>> R-devel at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-devel
>>>
>>>
>>
>>
>>
>
>
> --
> Ben Bolker
> Associate professor, Biology Dep't, Univ. of Florida
> bolker at ufl.edu / www.zoology.ufl.edu/bolker
> GPG key: www.zoology.ufl.edu/bolker/benbolker-publickey.asc
>
>



-- 
Felix Andrews / 安福立
Postdoctoral Fellow
Integrated Catchment Assessment and Management (iCAM) Centre
Fenner School of Environment and Society [Bldg 48a]
The Australian National University
Canberra ACT 0200 Australia
M: +61 410 400 963
T: + 61 2 6125 1670
E: felix.andrews at anu.edu.au
CRICOS Provider No. 00120C
-- 
http://www.neurofractal.org/felix/



More information about the R-devel mailing list