Obsidian | Level 7

Using PROC GENMOD, in "Criteria for Assessing Goodness of Fit" I got :

``````Deviance = 0.6841
Scaled Deviance = 1.1002``````

What does it means ? Is my model good ?

I know that a model with too large a deviance doesn’t fit the data well, but I dont know which deviance should I take

1 ACCEPTED SOLUTION

Accepted Solutions
Barite | Level 11

Re: Question about the PROC GENMOD

Its quite complicated to explain exactly how the statistics are calculated and how they should be used. I will recommend you to read the documentation which are very will written for this topic. SAS documentation

Shortly speaken, the scaled deviance is the deviance/dispersion. This statistic is in many model approximately chi-square distributed with n-p degrees of freedom.

5 REPLIES 5
Super User

Re: Question about the PROC GENMOD

Scaled Deviance is near 1 which means your model is not overdisperse / oversparse.

Barite | Level 11

Re: Question about the PROC GENMOD

I disagree with @Ksharp. The scaled deviance is a fit-statistic that can be used for testing whether each cell has its own predicted value (one parameter per cell) to the model specified. Therefore, and as the documentation mention, this fit-statistic has (in some models) a chi-square distribution with n-p degree of freedom.

Whereas the deviance/df  (found in the collumn "value/DF") is a test statistic for no overdispersion or underdispersion.

But, it is not in all models it make sense to use this test statistic!

Super User

Re: Question about the PROC GENMOD

@JacobSimonsen

Opps. I forgot  ScaledDeviance/DF = 1

Obsidian | Level 7

Re: Question about the PROC GENMOD

I dont understand

Output values are :

``````Deviance = 0.6841
Scaled Deviance = 1.1002``````

With Df = 57000

Barite | Level 11

Re: Question about the PROC GENMOD

Its quite complicated to explain exactly how the statistics are calculated and how they should be used. I will recommend you to read the documentation which are very will written for this topic. SAS documentation

Shortly speaken, the scaled deviance is the deviance/dispersion. This statistic is in many model approximately chi-square distributed with n-p degrees of freedom.

Discussion stats
• 5 replies
• 1668 views
• 4 likes
• 3 in conversation