R help - Apr 2007 - stringsAsFactor global option (was "character coerced to a factor")

--- Gabor Grothendieck <ggrothendieck at gmail.com>
wrote:
> Just one caveat.  I personally would try to avoid
> using
> global options since it can cause conflicts when
> two different programs assume two different settings
> of the same global option and need to interact.
  
   I see this argument often, and don't buy it.  In any case, for this
particular option, the Mayo biostatistics group (~120 users) has had 
stringsAsFactors=F as a global default for 15+ years now with no ill effects.
It is much less confusing for both new and old users.

   Johh Kane asked "Any idea what the rationale was for setting the
option to TRUE?"  When factors were first introduced, there was no option
to turn them off.  Reading between the lines of the white book (Statistical
Models in S) that introduced them, this is my guess: they made perfect sense for
the particular data sets that were being analysed by the authors at the time.
Many of the defaults in the survival package, which I wrote, have exactly the
same rationale --- so let us not be too harsh on an author for not forseeing
all the future consequences of a default!

  A place where factors really are a pain is when the patient id is a character
string.  When, for instance, you subset the data to do an analysis of only
the females, having the data set `remember' all of the male id's (the
original
levels) is non-productive in dozens of ways.  For other variables factors
work well and have some nice properties.  In general, I've found in my work
(medical research) that factors are beneficial for about 1/5 of the character
variables, a PITA for 1/4, and a wash for the rest; so prefer to do any
transformations myself.

For the historically curious: 
   In Splus, one originally fixed this with an override of the function
   	as.data.frame.character <- as.data.frame.vector
before they added the global option.  In R, unfortunately, this override
didn't work due to namespaces, and we had to wait for the option to be
added.  (Another dammed-if-you-do dammed-if-you-don't issue.  Normally you
don't want users to be able to override a base function, because 9 times out
of 10 they did it by accident and dont' want it either.  But when a user
really
does want to do so ...)  

	Terry Therneau

>   A place where factors really are a pain is when the patient id is a
character
> string.  When, for instance, you subset the data to do an analysis of only
> the females, having the data set `remember' all of the male id's
(the original
> levels) is non-productive in dozens of ways.  For other variables factors
> work well and have some nice properties.  In general, I've found in my
work
> (medical research) that factors are beneficial for about 1/5 of the
character
> variables, a PITA for 1/4, and a wash for the rest; so prefer to do any
> transformations myself.
It seems to me that the most importance difference between factors and
character vectors is that factors also store the range of the
variable.  You could imagine doing something similar for continuous
variables.  This would have the interesting property that plots of
subsets would have the same range as plots of the original data.  I'd
imagine, just as with factors, this would be useful and frustrating in
equal parts.

In terms of which should be the default, I can imagine two arguments:

 * keep to the original format of the data as closely as possible:
character vectors should be the default

 * maintain as much information about the original data as possible:
factors should be the default.

Hadley

On Mon, 23 Apr 2007, Terry Therneau wrote:
> --- Gabor Grothendieck <ggrothendieck at gmail.com>
> wrote:
>
>> Just one caveat.  I personally would try to avoid
>> using
>> global options since it can cause conflicts when
>> two different programs assume two different settings
>> of the same global option and need to interact.
>
>   I see this argument often, and don't buy it.  In any case, for this
> particular option, the Mayo biostatistics group (~120 users) has had
> stringsAsFactors=F as a global default for 15+ years now with no ill
effects.
> It is much less confusing for both new and old users.
>
>   Johh Kane asked "Any idea what the rationale was for setting the
> option to TRUE?"  When factors were first introduced, there was no
option
> to turn them off.  Reading between the lines of the white book (Statistical
> Models in S) that introduced them, this is my guess: they made perfect
sense for
> the particular data sets that were being analysed by the authors at the
time.
> Many of the defaults in the survival package, which I wrote, have exactly
the
> same rationale --- so let us not be too harsh on an author for not
forseeing
> all the future consequences of a default!
>
>  A place where factors really are a pain is when the patient id is a
character
> string.  When, for instance, you subset the data to do an analysis of only
> the females, having the data set `remember' all of the male id's
(the original
> levels) is non-productive in dozens of ways.  For other variables factors
> work well and have some nice properties.  In general, I've found in my
work
> (medical research) that factors are beneficial for about 1/5 of the
character
> variables, a PITA for 1/4, and a wash for the rest; so prefer to do any
> transformations myself.
>
> For the historically curious:
>   In Splus, one originally fixed this with an override of the function
>   	as.data.frame.character <- as.data.frame.vector
> before they added the global option.  In R, unfortunately, this override
> didn't work due to namespaces, and we had to wait for the option to be
> added.  (Another dammed-if-you-do dammed-if-you-don't issue.  Normally
you
> don't want users to be able to override a base function, because 9
times out
> of 10 they did it by accident and dont' want it either.  But when a
user really
> does want to do so ...)
That is what 'assignInNamespace' is for (and it came in with
namespaces).

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595