Re: st: Normally distributed error term & testing normality of residuals

["Seed, Paul" <paul.seed@kcl.ac.uk>]

I regularly use -qnorm- for this purpose when dealing with 
biomarkers in concentrations that are too small to detect.

The trick is to know which deviations from the diagonal are due to censoring (and ignorable)
and which to non-normality.

With uncensored data, I expect to see a straight diagonal line of points, matching the 
reference line (if the distribution is normal), or a curve away from the reference line (if it is not).  
With censored data, I expect to see much the same thing in either case, except for a horizontal section 
corresponding to the censored values.

Other plots (-qladder-, -qqreg-, -ppreg-) may also help, but I prefer to know one plot well than 
several badly.

I might add that I generally work on the raw data, not the residuals, as it is easier to 
understand the qnorm plot and the transformation needed; and I'm not interested in testing the 
residuals formally. 

BW

Paul Seed

On Sun, Oct 14, 2012 at 9:54 AM, Ebru Ozturk <ebru_0512@hotmail.com> wrote:
> Thank you. So, there is no possibility to check heteroscedasticity graphically?
>
> Ebru
>
> ----------------------------------------
>> Date: Sat, 13 Oct 2012 18:25:14 -0400
>> Subject: Re: st: Normally distributed error term & testing normality of residuals
>> From: jvverkuilen@gmail.com
>> To: statalist@hsphsun2.harvard.edu
>>
>> On Sat, Oct 13, 2012 at 1:39 PM, Ebru Ozturk <ebru_0512@hotmail.com> wrote:
>> > With default do you mean without robust standard error or something else?
>> >
>> Yes, the default is OIM.
>>
>>
>> > ----------------------------------------
>> >> Date: Sat, 13 Oct 2012 11:57:59 -0400
>> >> Subject: Re: st: Normally distributed error term & testing normality of residuals
>> >> From: jvverkuilen@gmail.com
>> >> To: statalist@hsphsun2.harvard.edu
>> >>
>> >> On Sat, Oct 13, 2012 at 11:41 AM, Ebru Ozturk <ebru_0512@hotmail.com> wrote:
>> >> > Thank you very much. One last question although it is slightly different from the main issue:
>> >> >
>> >> > In order to test heteroscedasticity assumption can I use scatter plots etc? or do I need to use a formal test like for testing normality issue?
>> >>
>> >> I think I'd probably use both, as the test just says "there appears to
>> >> be heteroscedasticity" but doesn't tell you much about what it might
>> >> be. One other quick and cheap diagnoser of heteroscedasticity is to
>> >> run with robust standard errors and with default. If they appear to
>> >> differ much, that's a sign you have a problem.
>> >> *
*
*   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/



Paul T Seed, Senior Lecturer in Medical Statistics, 

Division of Women's Health, King's College London
Women's Health Academic Centre KHP
020 7188 3642, 
 paul.seed@kcl.ac.uk, 
http://www.kcl.ac.uk/medicine/research/divisions/wh/about/people/seedp.aspx

Please do not send unencrypted un-anonymised data to this address.



*
*   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/