model.frame(), model.matrix(), and derived predictor variables

Ben,

It works for me ...

x = rpois(100, 5) + 1
y = rnorm(100, x)
d = data.frame(x,y)
m <- lm(y~log(x),d)
update(m,data=model.frame(m))

Call:
lm(formula = y ~ log(x), data = model.frame(m))

Coefficients:
(Intercept)       log(x) 
     -4.010        5.817 

You can also re-fit using the model.matrix directly. In your example,
the model frame can be passed directly to lm.fit /lm.wfit.

~G

sessionInfo()

R version 3.0.1 (2013-05-16)
Platform: x86_64-pc-linux-gnu (64-bit)

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C             
 [3] LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8   
 [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8  
 [7] LC_PAPER=C                 LC_NAME=C                
 [9] LC_ADDRESS=C               LC_TELEPHONE=C           
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C      

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base    

loaded via a namespace (and not attached):
[1] tools_3.0.1




On Sat, Aug 24, 2013 at 7:40 PM, Ben Bolker <bbolker at gmail.com
<mailto:bbolker at gmail.com>> wrote:


      Bump: just trying one more time to see if anyone had thoughts on this
    (so far it's just <crickets> ...)


    -------- Original Message --------
    Subject: model.frame(), model.matrix(), and derived predictor variables
    Date: Sat, 17 Aug 2013 12:19:58 -0400
    From: Ben Bolker <bbolker at gmail.com <mailto:bbolker at gmail.com>>
    To: R-devel at stat.math.ethz.ch <mailto:R-devel at stat.math.ethz.ch>
    <R-devel at stat.math.ethz.ch <mailto:R-devel at stat.math.ethz.ch>>


      Dear r-developers:

      I am struggling with some fundamental aspects of model.frame().

      Conceptually, I think of a flow from data -> model.frame() ->
    model.matrix; the data contain _input variables_, while model.matrix
    contains _predictor variables_: data have been transformed, splines and
    polynomials have been expanded into their corresponding
    multi-dimensional bases, and factors have been expanded into appropriate
    sets of dummy variables depending on their contrasts.
      I originally thought of model.frame() as containing input variables as
    well (but with only the variables needed by the model, and with cases
    containing NAs handled according to the relevant na.action setting), but
    that's not quite true.  While factors are retained as-is, splines and
    polynomials and parameter transformations are evaluated. For example

    d <- data.frame(x=1:10,y=1:10)
    model.frame(y~log(x),d)

    produces a model frame with columns 'y', 'log(x)' (not 'y', 'x').

    This makes it hard (impossible?) to use the model frame to re-evaluate
    the existing formula in a model, e.g.

    m <- lm(y~log(x),d)
    update(m,data=model.frame(m))
    ## Error in eval(expr, envir, enclos) : object 'x' not found

    It seems to me that this is a reasonable thing to want to do
    (i.e. use the model frame as a stored copy of the data that
     can be used for additional model operations); otherwise, I
    either need to carry along an additional copy of the data in
    a slot, or hope that the model is still living in an environment
    where it can find a copy of the original data.

    Does anyone have any insights into the original design choices,
    or suggestions about how they have handled this within their own
    code? Do you just add an additional data slot to the model?  I've
    considered trying to write some kind of 'augmented' model frame, that
    would contain the equivalent of
    that appeared in the formula but not in the model frame ...].
    setdiff(all.vars(formula),model.frame(m)) [i.e.  all input variables
    that appeared in the formula but not in the model frame ...].

      thanks
       Ben Bolker