Poisson Regression

[57078]

Hi,

I am trying to fit a poisson regression in GENMOD and don't think the data fit the model very well after viewing the residuals, I am already using the DSCALE option for overdispersion as well. I think I should be using a zero-inflated model but I don't have a variable that predicts the zeros really. I am using income and categorical race to predict green space. I just read somewhere that negative binomial is better for data with lots of zeros maybe I should try this?

Also should the residual plots be a random cloud around zero for count data? What other diagnostics should I use? What should they look like?

Thanks! I just ran negative binomial and it did not converge.


Message was edited by: 57078

[plf515]

> Hi,
>
> I am trying to fit a poisson regression in GENMOD and
> don't think the data fit the model very well after
> viewing the residuals, I am already using the DSCALE
> option for overdispersion as well. I think I should
> be using a zero-inflated model but I don't have a
> variable that predicts the zeros really. I am using
> income and categorical race to predict green space. I
> just read somewhere that negative binomial is better
> for data with lots of zeros maybe I should try this?
>
>
> Also should the residual plots be a random cloud
> around zero for count data? What other diagnostics
> should I use? What should they look like?
>
> Thanks!
>
> I just ran negative binomial and it did not converge.
>
>
>
> age was edited by: 57078

What data are you collecting, and on what geographic area?

I'm guessing "greenspace" is something like parks and so on. But income and race are individual variables ....

Why are there a lot of 0's?

Peter

[StatDave]

A plot of the standardized deviance residuals (STDRESDEV= option) or the likelihood residuals (RESLIK= option) can be useful. A plot of either against the predicted values should be randomly scattered about zero. Points beyond about 2 or -2 are not being fit well by the model. Plots of these residuals against the predictors can help you determine the proper form of the predictors in the model.

Note the apparent overdispersion may really mean that you need additional effects in your model. If there is overdispersion, you could use the REPEATED statement to fit the model using GEE (even if you have no repeated data). GEE corrects for overdispersion. For more, see this usage note:

http://support.sas.com/kb/22630