You can work this out! Generate some normal random deviates. Look at
the -qnorm-. Introduce some skewness. Look at the -qnorm-.
. clear
. set obs 100
obs was 0, now 100
. gen normal = rnormal()
. qnorm normal
. gen lognormal = exp(normal)
Use a sample size that is realistic for your project.
The bottom line is that skewness introduces curvature to the plot.
Nick
On Tue, Feb 12, 2013 at 3:36 PM, Xixi Lin <winnielxx@gmail.com> wrote:
> Thanks for your help. I have a question about the skewness. How to
> tell the skewness from the qnorm plot? Or is there any test that I can
> test the skewness of the residuals?
On Tue, Feb 12, 2013 at 9:41 AM, JVerkuilen (Gmail)
> <jvverkuilen@gmail.com> wrote:
>> Lots depends on details you've not indicated but I largely agree with
>> Maarten. Non-Gaussianity is in itself not a huge problem, but skew or
>> other clear patterns in the residuals may be a sign that the linear
>> regression is not correct. You may have an omitted variable, for
>> instance. A transformation may help if it makes sense, but I wouldn't
>> do it just to fix up residuals.
>>
>> If you have a reasonable regression model and still have mild
>> non-Gaussianity (which is best diagnosed using -qnorm-) then
>> bootstrapping or robust standard errors is a good fix.
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/faqs/resources/statalist-faq/
* http://www.ats.ucla.edu/stat/stata/
