[R] Trying to understand the magic of lm (Still trying)

Therneau, Terry M., Ph.D. therne@u @end|ng |rom m@yo@edu
Mon May 13 15:29:13 CEST 2019


  The text below is cut out of a "how to write a package" course I gave at the R 
conference in Vanderbilt.   I need to find a home for the course notes, because it had a 
lot of tidbits that are not well explained in the R documentation.
Terry T.


Model frames:
One of the first tasks of any modeling routine is to construct a special data frame 
containing the covariates that will be used, via a call to the model.frame function. The 
code to do this is found in many routines, and can be a little opaque on first view. The 
obvious code would be
coxph <- function(formula, data, weights, subset, na.action,
         init, control, ties= c("efron", "breslow", "exact"),
         singular.ok =TRUE, robust=FALSE,
         model=FALSE, x=FALSE, y=TRUE,  tt, method=ties, ...) {

      mf <- model.frame(formula, data, subset, weights, na.action)
since those are the coxph arguments that are passed forward to the model.frame routine.  
However, this simple approach will fail with a ``not found'' error message if any of the 
data, subset, weights, etc. arguments are missing. Programs have to take the slightly more 
complicated approach of constructing a call.
Call <- match.call()
indx <- match(c("formula", "data", "weights", "subset", "na.action"),
                   names(Call), nomatch=0)
if (indx[1] ==0) stop("A formula argument is required")
temp <- Call[c(1,indx)]  # only keep the arguments we wanted
temp[[1]] <- as.name('model.frame')  # change the function called
mf <- eval(temp, parent.frame())

Y <- model.response(mf)

We start with a copy of the call to the program, which we want to save anyway as 
documentation in the output object. Then subscripting is used to extract only the portions 
of the call that we want, saving the result in a temporary. This is based on the fact that 
a call object can be viewed as a list whose first element is the name of the function to 
call, followed by the arguments to the call. Note the use of \code{nomatch=0}; if any 
arguments on the list are missing they will then be missing in \code{temp}, without 
generating an error message. The \mycode{temp} variable will contain a object of type 
``call'', which is an unevaluated call to a routine.  Finally, the name of the function to 
be called is changed from ``coxph'' to ``model.frame'' and the call is evaluated.  In many 
of the core routines the result is stored in a variable ``m''.  This is a horribly short 
and non-descriptive name. (The above used mf which isn't a much better.)  Many routines 
also use ``m'' for the temporary variable leading to \code{m <- eval(m, parent.frame())}, 
but I think that is unnecessarily confusing.

The list of names in the match call will include all arguments that should be evaluated 
within context of the named dataframe. This can include more than the list above, the 
survfit routine for instance has an optional argument ``id'' that names an identifying 
variable (several rows of the data may represent a single subject), and this is included 
along with ``formula'' etc in the list of choices in the match function.  The order of 
names in the list makes no difference.  The id is later retrieved with 
\code{model.extract(m, 'id')}, which will be NULL if the argument was not supplied. At the 
time that coxph was written I had not caught on to this fact and thought that all 
variables that came from a data frame had to be represented in the formula somehow, thus 
the use of \code{cluster(id)} as part of the formula, in order to denote a grouping variable.

On 5/11/19 5:00 AM, r-help-request using r-project.org wrote:
> A number of people have helped me in my mission to understand how lm (and other fucntions) are able to pass a dataframe and then refer to a specific column in the dataframe. I thank everyone who has responded. I now know a bit about deparse(substitute(xx)), but I still don't fully understand how it works. The program below attempts to print a column of a dataframe from a function whose parameters include the dataframe (df) and the column requested (col). The program works fine until the last print statement were I receive an error,  Error in `[.data.frame`(df, , col) : object 'y' not found . I hope someone can explain to me (1) why my code does not work, and (2) what I can do to fix it.

	[[alternative HTML version deleted]]

More information about the R-help mailing list