Re: st: Adjustment to likehood value due to dependence of data observations

["Abdul Q Memon" <a.memon@ucl.ac.uk>]

Dear Nick

Thanks for a quick reply. Just to clearify few points.

1. As GEE takes a long time to run (quarter of million observations) so I
decided to run GLM which is quite fast and compared the likelihood values
for model selection. Based on those likelihood values I selected the model
and used the GEE for that. Now the question has arised that since it is
panel data, so all my likelihood values (obtained by using GLM) are wrong
due to dependent obervations. I am looking for a way to correct
(correction factor) for the likelihood to adjust the likelihood for glm
which is used for model selection.

2. After running the GEE which is suitable for model (Panel data) it seems
there is still some trend in residuals (Plot of residuals and fitted
values possibly becuase of homoscadicity). Am i right in using robust
after gee command?? I have looked at adding some more variables but afraid
that have no new ones.

Some more input from you might clear up what I am missing.

Regards

Memon



> #1 is easier. The answer is No in general. If you have time series for
> panels, you need to fit appropriate models at the outset. That is the
> principle. If you ignore that, then in broad terms, your parameter
> estimates may sometimes be about right, but standard errors are likely
> to be wrong and P-values are likely to be very very wrong. However, if
> the model really is wrong, it is best to fit another.
>
> #2 I don't understand. If there is a trend in residuals then your
> model sounds misspecified. Getting a better idea of what the standard
> errors are, or should be, for a model you have fitted that is
> evidently wrong doesn't sound very useful. You may be misunderstanding
> what -robust- options do, which is much less than people often think.
>
> On Thu, Sep 22, 2011 at 3:10 PM, Abdul Q Memon <a.memon@ucl.ac.uk> wrote:
>
>> I would really appreciate your reply on this.
>>
>> I have run several models using glm (possion and negative binomial)
>> command in STATA. Based on the log-likelihood and BIC values I have
>> selected the most appropriate models (with smallest BIC values). After
>> this I have run GEE with AR1 structure for only the preferred model to
>> account for serial correlation in data. I have these two questions.
>>
>> 1. Since my model seclection is based on (log-likelihood and BIC values)
>> and in this case data is not independent (time series and panel data),
%

*
*   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/