I have 2 models (negbin log and poi log) created using proc genmod in SAS. Both have overdispersion. Do I correct for overdispersion and then compare models again or do I compare right away choosing model with smallest deviance ignoring the overdispersion?
What is best remedy to be used with overdispersion? What does the remedy do to correct the situation? In re to former question, different remedies may affect things?
What you do depends on several things and is usually specific to your model. Redo the attached search for full details.. Best wishes David Booth
Further model comparison depends on more than you indicated. For an example see the attached paper. Best wishes David Booth
I have found two basic ways to address (remedy) overdispersion in the model
1.change distributional assumption to the negative binomial in which the variance is larger than the mean.
2.Run the model 2 times. The first time do not specify a scale parameter. It will be set equal to one. The second run should include a scale=pearson or scale =deviance statement. SAS will perform the overdispersion adjustment accordingly by scaling the covariance matrix. The model parameter standard errors will increase with the adjustment. The parameters themselves will not change.
Seeing the change is very helpful in understanding the process.
- Badges
- Science topic
- Similar topics
- Correction