|Date: ||Thu, 15 Sep 2005 12:28:43 -0700|
|Reply-To: ||David L Cassell <davidlcassell@MSN.COM>|
|Sender: ||"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>|
|From: ||David L Cassell <davidlcassell@MSN.COM>|
|Subject: ||Re: modeling overdispersion|
|Content-Type: ||text/plain; format=flowed|
>If anyone can point me in the right direction I would really appreciate it.
>I am modeling count data (frequency of accidents at intersections) and am
>using a negative binomial error distribution. Instead of having SAS
>one value for the overdispersion parameter I want to model the
>overdispersion parameter as a function of the characteristics at the
>intersections. Thus I would not only get the parameter estimates for the
>model predicting accident frequencies but also the parameters for the model
>predicting the overdispersion parameter. Any ideas on how this can be done?
I see that Dale has already given his usual excellent advice. So let me lob
my $0.02 .
Do you really have enough data to be able to do this kind of modeling
Do you have data that would, say, let you plot out characteristics and see
plots that different categories have different dispersions? (Bear in mind
you'll have to be able to ignore outliers as you look at the plots. They
your view of the plots just as easily as they can distort the estimation
If you don't, then I would recommend caution in taking the above approach.
Even if you do, I would recommend caution. I would recommend that you
perform some manner of Monte Carlo simulations or bootstrapping to get
a handle on how variable your final estimates really are. I would also
that you do the same with the typical one-dispersion-parameter model and
see how stable that is.
Hopefully you have a nice, stable model with really interesting
David L. Cassell
3115 NW Norwood Pl.
Corvallis OR 97330
Express yourself instantly with MSN Messenger! Download today - it's FREE!