There are at least two issues here.
1. Whether your data need a non-identity transformation or link
function and which would be good, or best.
I can't say. But alternatives include logit and cloglog.
2. Whether logs are _inherently_ problematic for arguments in (0,1).
My answer is No. But for your data, plot log(p) against p and
see if outliers pop out. Look at the distribution of p and log(p),
etc. Consult theory, and use subject-matter knowledge.
Alessia Matano, or Ilaria Tucci, or anyone else on the team there:
zero is not present in my estimation. My variable varies between 0 and
1 with extremes excluded, since my share is the relative employment of
white collar workers in any firm and
there are no firms with only non white collar workers. However the
fact that the logarithmic variable is bounded at 0 worries me. It's
like I am estimating an impact on a variable which is "truncated", but
not in a classical way. It is in this respect that I would like to get
suggestions.
2007/11/6, Nick Cox <[email protected]>:
> The fact that a logarithm may be negative should not worry you one
bit.
> The fact that log 0 is indeterminate should worry you a lot if you
have
> exact zeros in your data. What to do when you have 0 and wish to take
> logarithms is one of the most frequently discussed problems on this
> list.
>
> Nick
> [email protected]
>
> Ilaria Tucci
>
> Another question: Is that right that is not possible to transform the
> dependent variable in logarithmic form and then apply the related
panel
> data models, since the logarithm of a variable between 0 and 1, varies
> between -infinite and 0, and thus it does no make sense to perform
> analysis on such a variable?
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
