R help - Mar 2014 - Survfit error

>> The help page for the survfit function says it expects a formula as its
>> first argument so try:
>>
>> > sleepfit <- survfit(Surv(timeb, death)~1, data = sleep)
>>
>> David
>> Sent from my iPhone ... so unable to test.
>>
>>
> This was a recent (well, 2007) change in behaviour. Previously the function
> did some tricks to make either approach work, which could be described as
> 'clever' or 'too clever by half'.
>
>   -thomas
  Certainly the latter, a design mistake that I finally admitted and corrected. 
The
trouble is this usage
      fit <- survfit(Surv(time, status), data=mine)

The data, subset, weights, and na.action arguments are all meant to work with
formulas.
The first argument above isn't a formula.  This left survfit with the
quandry of how to
recognize that the first argument is a "Surv" object and not a
formula, and do something
differently.  Which leads to a stalemate
	a. One doesn't know what type of object the first argument is until it is
evaluated
	b. Evaluation will fail, however, unless it is evaluated in the
"data=mine" context
	c. The function that does that process (model.frame) expects a formula as its
first arg

My program needed to know the answer to a) in order to transform the first arg
into a
formula and correctly execute c), but needed the answer from c) to know if that
was necessary.

The old code peeked back at the actual string that was typed by the user,
looking for the
letters "Surv", but was easily fooled.

On Fri, Mar 7, 2014 at 6:46 AM, Therneau, Terry M., Ph.D.
<therneau at mayo.edu> wrote:
>>> The help page for the survfit function says it expects a formula as
its
>>> first argument so try:
>>>
>>> > sleepfit <- survfit(Surv(timeb, death)~1, data = sleep)
>>>
>>> David
>>> Sent from my iPhone ... so unable to test.
>>>
>>>
>> This was a recent (well, 2007) change in behaviour. Previously the
>> function
>> did some tricks to make either approach work, which could be described
as
>> 'clever' or 'too clever by half'.
>>
>>   -thomas
>
>
>  Certainly the latter, a design mistake that I finally admitted and
> corrected.  The trouble is this usage
>      fit <- survfit(Surv(time, status), data=mine)
>
> The data, subset, weights, and na.action arguments are all meant to work
> with formulas. The first argument above isn't a formula.  This left
survfit
> with the quandry of how to recognize that the first argument is a
"Surv"
> object and not a formula, and do something differently.  Which leads to a
> stalemate
>         a. One doesn't know what type of object the first argument is
until
> it is evaluated
>         b. Evaluation will fail, however, unless it is evaluated in the
> "data=mine" context
>         c. The function that does that process (model.frame) expects a
> formula as its first arg
>
> My program needed to know the answer to a) in order to transform the first
> arg into a formula and correctly execute c), but needed the answer from c)
> to know if that was necessary.
>
> The old code peeked back at the actual string that was typed by the user,
> looking for the letters "Surv", but was easily fooled.
It's not possible to be perfect in this scenario, but you could use
some heuristics:


survfit <- function(x, ...) {
  xq <- substitute(x)
  if (!is.call(xq)) {
    stop("First argument to survfit must be a function call")
  }

  if (identical(xq[[1]], quote(Surv))) {
    "surv"
  } else if (identical(xq[[1]], quote(`~`))) {
    "formula"
  } else {
    "don't know"
  }
}

survfit(Surv(time, status), data=mine)
survfit(Surv(timeb, death)~1, data = sleep)


You'd need a different heuristic if is.name(xq) is TRUE, indicating
that the user passed in the name of an existing variable.

Hadley


-- 
http://had.co.nz/