[R] Setting log(0) to 0

[Terji Petersen]

Thank you both of you, Kenneth and Ted:-). Log1p(x) is not what I asked 
for, but it is better:-D

And Ted, thanks for your thoughts on funcional form. I'm just starting 
out with R, and feel like I've barely scratched the surface of the 
program. I have never in my life done a non-linear regression, but that 
will soon change:-)

(Ted Harding) wrote:

>On 16-Feb-05 Terji Petersen wrote:
>  
>
>>Hi,
>>
>>I'm trying to do  a regression like this:
>>
>>wage.r = lm( log(WAGE) ~ log(EXPER)
>>
>>where EXPER is an integer that goes from 0 to about 50.
>>EXPER contains some zeros, so you can't take its log,
>>and the above regression therefore fails. I would like
>>to make R accept log(0) as 0, is that  possible?
>>Or do I have first have to turn the 0's into 1's to be
>>able to do the above regression?
>>    
>>
>
>If treating "log(0)" as 0 would do your business, then a
>preliminary pass to turn "0" into "1" would be the simplest
>method. It only takes 1 line.
>
>This is a bit of a Catch-22. It looks like you're trying
>to fit a power law
>
>  WAGE = A*(EXPER^B)
>
>(where I guess "EXPER" means experience) and you've got
>some cases with no experience. Whether your work-round
>is appropriate depends in part on the unit of "experience".
>If it's in years, then a case with 3 months experience
>would have log(EXPER) = -1.39, thereby weighing in with
>a lesser value than someone with zero experience, on your
>proposal.
>
>On the other hand, if it's in days, then
>
>  log(EXPER) = log(91) = 4.51
>
>and even someone with only a week has log(EXPER) = 1.95
>
>But your log(0) = 0 data would be sitting there all the
>time, whatever the scale of EXPER, and so would have an
>influence on your regression which depended on this scale.
>You might have to consider using log(0) --> const
>where the "const" is such as to give reasonable results,
>given what comes out of the rest of the data (where EXPER>0).
>
>The fundamental problem is that your power law predicts
>zero wage for zero experience, which is rarely the case.
>
>You might do better to try a non-linear fit
>
>  WAGE = W0 + A*(EXPER^B)
>
>for which sort of thing there are several resources in R,
>perhaps the simplest being 'nls'.
>
>For what you have in your installed packages, try a
>
>  help.search("nonlinear")
>
>Once you open this door, you can try perhaps more realistic
>non-linear models, including what can be found amongst the
>"SS....." (Self-Starting) models in "nls" --  have a look
>at what's listed by
>
>  library(help=nls)
>
>as well as what is allowed according to "?nls".
>
>Such models would allow an initial (zero-experience) wage,
>perhaps not changing much for some time, then rising more
>rapidly once an "experience threshold" is passed, then
>flattening out to a lower slope over a longer time (something
>which many of us have experience of). And even ultimately
>ending to decrease ...
>
>Hoping this helps,
>Ted.
>
>
>--------------------------------------------------------------------
>E-Mail: (Ted Harding) <Ted.Harding at nessie.mcc.ac.uk>
>Fax-to-email: +44 (0)870 094 0861
>Date: 16-Feb-05                                       Time: 13:26:53
>------------------------------ XFMail ------------------------------
>
>  
>