topic Re: Exact logistic in SAS Procedures

Exact logistic

deleted_user — Wed, 02 Dec 2009 03:23:05 GMT

I am peforming proc logistic with the exact option. I keep getting an error message that reads....

The exact distribution contains frequencies larger than 9.0071993E15; accuracy was lost.

The model runs and get odds ratios and confidence intervals (which are comparable to logistic without the exact option). I am just confused as to why I am getting this error and what it means. I only have 528 people in my sample.

Any insight would be much appreciated.

Re: Exact logistic

Doc_Duke — Wed, 02 Dec 2009 16:57:56 GMT

From the 9.2 documentation "If a frequency in the exact distribution is larger than the largest integer that can be held in double precision, a warning is printed to the SAS log. But since inaccuracies due to adding small numbers to these large frequencies might have little or no effect on the statistics, the exact computations continue."

Re: Exact logistic

deleted_user — Thu, 03 Dec 2009 01:19:48 GMT

Thanks for the response. I don't really understand that either lol. Maybe I'm just dense. I called SAS Tech Support and one of their logistic gurus is supposed to be getting back to me. I'll post what they tell me.

Re: Exact logistic

Doc_Duke — Thu, 03 Dec 2009 15:27:43 GMT

Not dense, I heard a different question. The underlying reason for the message is that the counters have reached a floating point overflow condition (remember that SAS uses floating point for all arithmetic, even addition of whole numbers). So SAS is just telling you that it can on longer exactly represent the whole number in it's floating point system. Now, why does that happen with the EXACT statement when you have "only" 528 observations.

Actually, 500 is a LOT for an exact computation. The computation is described in
http://support.sas.com/documentation/cdl/en/statug/63033/HTML/default/statug_logistic_sect045.htm
It's very nasty arithmetic. This statement is where the action is ".. created by summing over all binary sequences... ". It is basically saying that it has to look at all possible combinations in your data in order to get the estimate. If you did it with brute force, you would have 2 raised to the 528 (2^528) power possibilities to compute (my calculator just generated an error message when I tried to get the total). There are computational efficiencies, but it is still a LOT of computation to do.

Re: Exact logistic

deleted_user — Thu, 10 Dec 2009 05:35:09 GMT

Thanks! That's what I have come to find out. But it was suggested that I use it anyway.

Here is the feedback I got from SAS:

The "accuracy was lost" warning means that in compiling the exact conditional distribution, some frequencies became extremely large. See "Exact Conditional Logistic Regression" in the PROC LOGISTIC documentation:

http://support.sas.com/91doc/getDoc/statug.hlp/logistic_sect42.htm

See also the comments toward the end of the "Computational Resources for Exact Conditional Logistic Regression":

http://support.sas.com/91doc/getDoc/statug.hlp/logistic_sect49.htm

As noted there, this message isn't necessarily problematic. It's just letting you know that the exact counts could not be maintained. Given the complexity of the exact algorithm, it's not really possible to track the cause of this message back to some feature in the data or model, but this problem is increasingly likely as the data set size increases. For data sets which are too large for the exact method, the Monte Carlo method (available in SAS 9.1 by specifying METHOD=NETWORKMC in EXACTOPTIONS) produces estimates of the exact p-values.