Re: st: QREG Question

[Nick Cox <njcoxstata@gmail.com>]

There's some confusion here between what to do when a response  
variable has a highly skewed distribution and what functional form is  
appropriate to express the relationship between that response and a  
bundle of predictors.

By implication Arnold has a response that is expected to have positive  
mean but is often zero in practice. I'd expect, with no further  
information, that -qreg- might perform a bit better than -regress- in  
this situation, but I'd also expect that -poisson- would do better  
than either. See also  -glm, link(log)-. Seek out Bill Gould's blog  
posting on Poisson at blog.stata.com

However, if some or all of the zeros come from a distinct group you  
may need a two-part model. For example, zero hours spent watching the  
Olympics on TV may reflect zero interest and/or zero access to a TV.

Either way no linear combination Xb has "correcting for skewness" as  
its direct  purpose.

What -qreg- does is well described in the documentation but by default  
it models the conditional median.

Nick

On 27 Jul 2012, at 20:33, Arnold Behrer <behrer@post.harvard.edu> wrote:

> Hello,
>
> I have two questions, one related to stats generally and the other to
> Stata specifically.
>
> The first is about how useful quantile regression is in correcting for
> skewed data.  I have a data set heavily skewed to the right.  I have
> considered log transforming it but I would rather not because the
> skewness is due in large part to zero values that are meaningful in
> the context of the regression.  I thought that perhaps quantile
> regression might be a way of correcting for this skew without dropping
> the zero values.  Is that true?
>
> The second question is simply whether the Qreg command in Stata runs a
> normal quantile regression - splitting the
*
*   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/