Hi @ErikLund_Jensen, The initial dataset contains just a little under 1.4 billion records with 59 variables. 50 are character with an average length of 7 and 9 are numeric with average length of 8. The filtering is based on a single variable unrelated to the ones used in the code discussed here. The dataset is about 400GB. However, after filtering, the dataset is reduced to just a little under 2.6% of the total observations (~36 million) and 4 variables. To this subset the codes we discussed here would be applied to reduce it further to ~13 million in the final dataset. Of course, this all happens in a single step in PROC SUMMARY and PROC SQL. I hope that makes sense. I would also be curious to see what you find. I think the I/O operations speed is going to be important here. ... View more

> Since it's not the default, is there a reason not to be in unicode mode?  None, except most historical data is written in ASCII, and therefore there will be a performance penalty reading ASCII tables in a UTF8 SAS session.   I didn't notice differences in data set size, even when storing double-byte values like Russian characters (and I can't explain why. Why oh why?).   Also, once you allow Unicode data in the data sets, a warning in the log will inform that these characters are lost whenever non-ASCII characters are retrieved from a non-Unicode SAS session.   ... View more

For legends (keylegend, gradlegend, etc), you should refer to the plot's NAME value -- not the statement name. In Reeza's example, the heatmap has the name of "nope1". So, in this case, you would say:   gradlegend "nope1";   Hope this helps! Dan ... View more

I think this posting is also relevant to your inquiry:   https://communities.sas.com/t5/SAS-Statistical-Procedures/Proc-Mixed/m-p/296000#U296000 ... View more