topic Re: CAS Bootstrapping in SAS Data Science

CAS Bootstrapping

phopkinson — Fri, 07 Jun 2024 10:03:58 GMT

Is there any way of replicating proc surveyselect in CAS for bootstrapping datasets - I have found sampling.srs action but can't seem to get the code working and as a second point proc surveyselect returns the number of times a row has been selected can we do the same with sampling.srs ?

Re: CAS Bootstrapping

sbxkoenk — Fri, 07 Jun 2024 20:58:03 GMT

Hello,

It's important to know which Viya version you are using , because in Viya 3.5 it is difficult (but possible) to get always the same bootstrap re-samples (across memory purges). In Viya 4, it's easy as there is a "reproducibility button" (plastically expressed).

Please submit :

%put &=sysvlong4;
%put &=SYSVIYARELEASE;
%put &=SYSVIYAVERSION;

... and tell us about the results (see log-screen)!

Also, in PROC SURVEYSELECT you probably use method=URS (sampling with equal probability and with replacement) , while the sampling.srs action corresponds to method=SRS.

Go here for what you need (proper bootstrap re-sampling in CAS), but we may need to assist you more to ensure reproducibility and repeatability (first tell us about your Viya release before I go into details here).

Bootstrap Resampling At Scale: Part 1 (of 3) dd. 03 March 2020
https://statmike.com/blog/sgf2020p1
Bootstrap Resampling At Scale: Part 2 (of 3) dd. 04 March 2020
https://statmike.com/blog/sgf2020p2
Bootstrap Resampling At Scale: Part 3 (of 3) dd. 05 March 2020
https://statmike.com/blog/sgf2020p3

BR, Koen

Re: CAS Bootstrapping

sbxkoenk — Fri, 07 Jun 2024 21:10:15 GMT

A note on the above (see my earlier reply).

I understand that eternal reproducibility and repeatability may be desired, but we should not -- in the case of bootstrapping -- exaggerate its importance either. After all, if you get completely different results with slightly different re-samples, that is a good signal that something is wrong with the bootstrapping. Then you probably need more and/or bigger re-samples. As soon as you set up reliable bootstrapping, the actual samples (almost) don't matter anymore. The results will be identical to several decimal places.

Koen