Re: st: xtgee for skewed data

[Nick Cox <njcoxstata@gmail.com>]

No, I don't think so. Your problem as I understand it is that your
response variable has granularity near 0 because the reportable values
start with  0 or 0.04.  This isn't a censoring or truncation problem.

The bigger questions for you are whether 0 is a plausible value and
also whether you want to entertain models that might predict negative
values for some combination of predictors.

Also, your variable sounds like a concentration so there is an upper
limit too, merely that it does not bite, i.e whatever it is will never
be solid beta carothen. whatever that is.

Some kind of -glm, f(binomial) vce(robust) link(logit)- as very often
discussed on this list might work better.

Nick

2011/8/25 José Maria Pacheco de Souza <jmpsouza@usp.br>:

> I am not sure whether I can include this question in the thread above, but
> the subjet is related. If the interest is to run a linear regression  of a
> continuous variable as the response, say level of beta carothen, and for
> very small values the results are zero because the equipment can only show
> values equal or greater than .04, the use of -tobit- can be an statistical
> alternative?
> Let´s assume there is no money to buy a better device.
>
> Em 25/08/2011 08:15, Nick Cox escreveu:
>>
>> The discussion started by William Gould at
>>
>>
>> http://blog.stata.com/2011/08/22/use-poisson-rather-than-regress-tell-a-friend/
>>
>> seems relevant. You have a massive spike in the distribution. No
>> transformation will much affect that, as by Murphy's theorem a spike
>> maps to a spike, and in any case there would be the usual argument
>> about what to do with zeros. However, (importantly different here)

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/