It occurs to me that this behavior may be caused by omitted variable bias. I suggest you see what happens if you add higher powers and interaction terms for your independent variables, and see what happens to the serial correlation of the residuals. David Greenberg, Sociology Dept., New York University
----- Original Message -----
From: Antonio Silva <[email protected]>
Date: Thursday, August 14, 2008 2:59 pm
Subject: st: More on autocorrelation in Poisson, some diagnostic results
To: [email protected]
> Hello All: I fear now that I run the risk of alienating the people who
> have helped me with my question, but I am going to ask one more
> question nonetheless. In response to helpful comments I received, (see
> previous posts), I ran a Poisson model, following the advice of
> several posters. If you recall, I was concerned about autocorrelation
> in a Poisson model.
>
> Here is the model I ran:
>
> glm Y X1 X2 X3, family(poisson) link(log)
>
> The actual results of the model are good, and they confirm the theory.
> But next, to look at the residuals, I did this:
>
> predict dev
> then this:
> corrgram dev
> Here are the results of that exercise:
>
>
> LAG AC PAC Q Prob>Q [Autocorrelation] [Partial Autocor]
>
> -------------------------------------------------------------------------------
>
> 1 0.9403 0.9449 38.972 0.0000 |------- |-------
>
> 2 0.8620 -0.2916 72.558 0.0000 |------ --|
>
> 3 0.7664 -0.4628 99.808 0.0000 |------ ---|
>
> 4 0.6625 -0.1638 120.72 0.0000 |----- -|
>
> 5 0.5586 -0.1633 136 0.0000 |---- -|
>
> 6 0.4575 -0.3101 146.54 0.0000 |--- --|
>
> 7 0.3369 -0.5216 152.43 0.0000 |-- ----|
>
> 8 0.2092 -0.5005 154.77 0.0000 |- ----|
>
> 9 0.0897 -0.4995 155.21 0.0000 | ---|
>
> 10 0.0007 0.2802 155.21 0.0000 | |--
>
> 11 -0.0657 0.5618 155.46 0.0000 | |----
>
> 12 -0.1220 0.4781 156.37 0.0000 | |---
>
> 13 -0.1667 0.0302 158.12 0.0000 -| |
>
> 14 -0.2122 0.5885 161.06 0.0000 -| |----
>
> 15 -0.2525 0.1082 165.38 0.0000 --| |
>
> 16 -0.2761 0.1019 170.75 0.0000 --| |
>
> 17 -0.2791 -0.3027 176.48 0.0000 --| --|
>
> 18 -0.2681 1.3390 181.98 0.0000 --| |--------
>
>
>
> This looks pretty bad to me, as these results seem to suggest serious
> AC problems. Am I correct in this conclusion? My first thought was to
> use arpois, with an ar(1) variable in the model. Does this sound
> reasonable? Even when I do this, however, the results of the corrgram
> show big-time AC of the residuals. Moreover, there continues to be AC
> in the residuals, even when I use higher order ar terms in the model.
> I am really not sure what to do next. Give up? Run arpois with more ar
> parameters?
>
>
> As an aside, I should also note that if I include a lagged version of
> the dependent variable of the model to see if there was some sort of
> correlation in the Poisson counts themselves, and the variable did not
> turn up significant.
>
>
> Again, any thoughts are helpful.
> Antonio
>
> _________________________________________________________________
> Get Windows Live and get whatever you need, wherever you are. Start here.
> http://www.windowslive.com/default.html?ocid=TXT_TAGLM_WL_Home_082008
> *
> * 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/
*
* 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/
