R devel - Mar 2013 - numerics from a factor

A problem has been pointed out by a French user of the survival package and
I'm looking
for a pointer.

 > options(OutDec= ",")
 > fit <- survfit(Surv(1:6 /2) ~ 1)
 > fit$time
[1] NA  1 NA  2 NA  3

A year or two ago some test cases that broke survfit were presented to me. The
heart of
the problem was numbers that were almost identical, where table(x) and unique(x)
gave
different counts of distinct values.
The solution was to use "ftime <- factor(time)" at the top of the
code, and do all the
calulations using the integer levels of the factor as the unique time points. 
At the very
end the numeric component "time" of the result is created using 
as.numeric(levels(ftime)).  It's this last line that breaks.

I could set the OutDec option within survfit and reset when I leave using
on.exit.  Any
other simple solutions?  Any other ways I could get caught by this issue?

Terry Therneau

Hi Terry,

you can use type.convert instead of as.numeric for numbers with decimals:

type.convert(levels(factor(1:6/2)), dec=unlist(options("OutDec")))

Best,
Ulrike


Am 16.03.2013 12:00, schrieb r-devel-request at
r-project.org:
> Date: Fri, 15 Mar 2013 12:47:01 -0500
> From: Terry Therneau<therneau at mayo.edu>
> To:"r-devel at r-project.org"  <r-devel at r-project.org>
> Subject: [Rd] numerics from a factor
> Message-ID:<51435E95.304 at mayo.edu>
> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
>
> A problem has been pointed out by a French user of the survival package and
I'm looking
> for a pointer.
>
>   > options(OutDec= ",")
>   > fit <- survfit(Surv(1:6 /2) ~ 1)
>   > fit$time
> [1] NA  1 NA  2 NA  3
>
> A year or two ago some test cases that broke survfit were presented to me.
The heart of
> the problem was numbers that were almost identical, where table(x) and
unique(x) gave
> different counts of distinct values.
> The solution was to use "ftime <- factor(time)" at the top of
the code, and do all the
> calulations using the integer levels of the factor as the unique time
points.  At the very
> end the numeric component "time" of the result is created using
> as.numeric(levels(ftime)).  It's this last line that breaks.
>
> I could set the OutDec option within survfit and reset when I leave using
on.exit.  Any
> other simple solutions?  Any other ways I could get caught by this issue?
>
> Terry Therneau