One plausible explanation for the serial autocorrelation is omitted variables that have some stability over time. This might suggest trying negative binomial regression. You can again look at the autocorrelations among the residuals. Something else to try is a lagged endogenous variable in the Poisson regression. David Greenberg, Sociology Department, New York u.
----- Original Message -----
From: Antonio Silva <[email protected]>
Date: Tuesday, August 26, 2008 1:43 pm
Subject: st: serial autocorrelation in residuals, what to do?
To: [email protected]
> ________________________________
>
> From: [email protected]
> To: [email protected]
> Subject: serial autocorrelation in residuals, what to do?
> Date: Mon, 25 Aug 2008 22:53:46 -0400
>
>
> Hello Statalist:
> Well, again I must resort to you for help. Here is the situation. I
> ran a Poisson model of the following sort:
>
> y = x1 + x1squared + X2 + X3 + X4
>
> The results were very good. I hypothesized that there was an inverse
> U-shaped relationship between X1 and the dependent variable, and the
> results supported this--in other words, X1 was positive, and X1squared
> was negative, and both coefficients were significant. Also, the
> inflection point was within the range of the data.
> However, I sent the paper out for review, and a reviewer suggested I
> test for autocorrelation. I asked for help, and members of this great
> list helped me. I found that there was no autocorrelation of the
> Poisson counts themselves (they are yearly counts). However, next,
> based on advice I received from members of this list, I tested for
> autocorrelation of the residuals (using GLM and the predict dev
> command) and found out that the residuals were highly correlated. In
> short, the residuals are not independent.
> I am stuck. I have tried everything to resolve this problem. First, I
> tried transforming X1 and X1squared (the variables of most interest)
> by using first differences for both. When I did this, it eliminated
> the autocorrelation among the residuals, but the model blew up and X1
> and X1squared became highly insignificant. Second, I tried adding
> omitted variables. This helped, but not enough.
> I am writing to see if anyone has any ideas about what I should do
> from here. I
> figure I have three options.
> 1. I can give up.
> 2. I can try a different transformation of the variables of interest
> (any ideas?). or
> 3. I can cite some work that says that serial autocorrelation of
> residuals in Poisson regression is not that big of a problem. (I
> figure this may be the case because including the squared term in the
> model virtually assures that the residuals will be correlated).
> Any advice is welcome and appreciated. Thanks again for all your help.
>
> Antonio Silva
>
>
> ________________________________
>
> Get thousands of games on your PC, your mobile phone, and the web with
> Windows®. Game with Windows
>
> _________________________________________________________________
> Be the filmmaker you always wanted to be—learn how to burn a DVD with
> Windows®.
> http://clk.atdmt.com/MRT/go/108588797/direct/01/
> *
> * 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/
