ERROR: Integer overflow on computing amount of memory required for confidence limit computation.

[Katya1517]

Hi, experts!

I have a trouble with  proc Loess:

my code is 

Proc LOESS data=work.pre_loess; 
ods output OutputStatistics=RESULTS_1;
model Pre_Bl=Date / smooth=0.2
residual alpha=0.05
dfmethod=exact;
by Code;
run;

And it throws an error: ERROR: Integer overflow on computing amount of memory required for confidence limit computation.

Do you have any ideas?

[SteveDenham · Accepted solution]

How many points are in your regression?  The documentation says that the work is proportional to n cubed, so the EXACT option might be what is causing this.  Have you tried the default DFMETHOD=NONE or the other method DFMETHOD=APPROX? 

If you still run into the issue you may have to sample your data, and devise some method of model averaging.  One way would be to generate samples with replacement, fit PROC LOESS by sample, and then apply some bootstrapping to average the models.  And this is where my knowledge of PROC LOESS hits a wall - I don't know which of the output tables generated by the DETAILS option will give what you need.

SteveDenham

[Katya1517]

Many thanks!

[Rick_SAS]

Since you are not outputting the CLM, I don't think the exact computation is required. You can probably use DFMETHOD=NONE since you are only computing the predicted and residual values.