Output from Glimmix procedure, how to interpretate between estimate of ">999.999" vs "."

[chouj]

Hi,

I ran the multinominal regression model with GLIMMIX procedure. The outcome variable has 7 categories and some of the categories contain small frequency counts or no event at all. The data is very sparse. On the Odds Ratio estimates output, there are estimates labelled as either "<0.001" or ">999.999". On the 95% confidence limits, some of the estimates are printed as "."

What are the differences between ">999.999" or "<0.0001" vs "."

In the case where I obtain OR='<0.001', the beta estimate is a very large negative number (example -9.6)

In the case where I obtain OR='>999.999", the beta estimate is a very large positive number (example, 12.87)

I believe ">999.999' and "<0.001" are SAS formats to label extremely large or small values, respectively . What about in place where I got "." Does "." means that the value is not estimable or exp(beta)=infinity? 

Is there a citation or documentation on these? 

Thank you

Thank you so much!

[jiltao]

This might be due to the quasi-complete separation.

I see that you opened a case with SAS Technical Support. I will provide more information in the ticket if needed.

Thanks,

Jill

[ballardw]

For some details you can try looking at the TEMPLATE definitions used to create the output.

For example the  Oddsratio template for Glimmix has elements such as

define Estimate;                                                        
         header = "Estimate";                                                 
         translate _val_=. into " Non-est";                                   
         format = oddsr8.3;                                                   
      end; 

And you can look up the Oddsrw. format to find:

The ODDSRw.d format follows the rules for the w.d format, except in these conditions:

• values from –1e-12 to 10^-d –eps are displayed as less than “0.0...01”
• values that are greater than or equal to 10^w-d-2 are displayed as greater than “999.999”

Special values are shown in the documentation but .  is still . (missing). So is suspect something was not able to be calculated for your data and options chosen. Ratios for example require two values. If one is missing no ratio can be calculated.

[chouj]

Thank you so much! This is very helpful

[Ksharp]

Maybe that is a value scale problem.
Make variable have more scale unit:
x=x*100;
x=x*1000;