Re: Re: st: Automatic fit of distribution

[Richard Williams <richardwilliams.ndu@gmail.com>]

Changing the subject slightly -- it is often recommended that you 
examine your data, e.g. do graphs or whatever, run various 
diagnostics. I am inclined to agree; indeed I always tell people to 
start with assorted descriptive statistics before launching into 
their high tech models. However, things like stepwise regression are 
widely condemned. Again I am inclined to agree, but I have a hard 
time explaining what exactly the difference is. In both cases, aren't 
you looking at the data first and using that information to guide 
your model building? By graphing the data first, couldn't that lead 
to over-fitting, and run the risk that analysis with different data 
would lead to different results? If, say, my visual examination or 
diagnostics have led me to add squared terms or even use a different 
statistical method, aren't my p values misleading? It seems like a 
lot of the cautions and concerns raised with stepwise could also be 
raised for approaches that are considered much more acceptable. My 
instincts go with the conventional wisdom but I am not sure how I 
would respond if pressed on these matters.

At 11:29 AM 7/11/2013, David Hoaglin wrote:
> Diagnostics are fine, but there is no sustitute for looking at the
> data (e.g., in well-chosen histograms and quantile-quantile plots).
> Programs that rely on the sample skewness and kurtosis will be blind
> to mixtures that show more than one mode, and the sensitivity of
> sample moments to outliers makes those measures unsuitable for
> diagnosing distribution shape.
>
> Also, the process should take into account whether the data are
> continuous or discrete.
>
> David Hoaglin
>
> On Thu, Jul 11, 2013 at 11:45 AM, Ariel Linden. DrPH
> <ariel.linden@gmail.com> wrote:
> > I completely agree with Nick and Maarten that the user should do the work
> > required to determine what type of distribution they are dealing 
> with and go
> > from there. However, it seems to me that there could be a program that
> > "points the user in the right direction" after running a few simple
> > diagnostics. For example, there are several programs already available to
> > test for normality (ie., -sktest-, -swilk-, -ksmirnov-). It would be rather
> > straightforward to test for a Poisson distribution based on the variance =
> > mean. It would get harder as we go to other distributions, or fall between
> > choices...
> *
> *   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/

-------------------------------------------
Richard Williams, Notre Dame Dept of Sociology
OFFICE: (574)631-6668, (574)631-6463
HOME:   (574)289-5227
EMAIL:  Richard.A.Williams.5@ND.Edu
WWW:    http://www.nd.edu/~rwilliam

*
*   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/