SAS loses decimal accuracy

[Dontik]

Hi

I have a weird problem. I am operating on numbers which add up to around 800 000 000. I have a dataset which I need to split into minidatasets with 1000 000 records. The problem is that when I split my dataset, I loose decimal precision, for example:

Befote splitting the sum of my costs was 810268311.8536920

After splitting - 810268311.8401380

Splitted datasets have the total number of records equal to number of records which was in the initial table.

What can I do about that?

Thank you in advance!

[TomKari]

Your numbers never had the accuracy. The 8 byte floating point numeric format used by SAS can only hold up to fifteen significant digits accurately. Your numbers have sixteen digits; you will lose precision on the sixteenth and above digits.

Tom

[Dontik]

Hmm so what is there to do about that? Can I at least keep the number at 0.01 decimal precision? 810268311.85 - so that I could get this exact number after splitting, I do not need anything more precise. I"ve tried adding round(cost,0.01) to my data step which splits the data but that didn't help.

[ChrisHemedinger]

There is an exhausting treatise on this topic here:

• Numerical precision in SAS

The behavior of SAS is the same as any software that relies on floating point math, defined by an IEEE standard.

There is a newer DECIMALCONV= option that might help.  From the doc:

DECIMALCONV=STDIEEE

specifies to convert and format decimal values by using the IEEE Standard for Floating-Point Arithmetic 754-2008. Using the STDIEEE argument improves the accuracy and readability of floating-point numbers. In some cases, more significant digits can be displayed in the same field width.

[Dontik]

Unfortunately, my SAS says "unrecognized SAS option name, DECIMALCONV" 😕

[ChrisHemedinger]

DECIMALCONV= was added in SAS 9.4. 

[Astounding]

To the extent you can work with integers instead of decimal fractions, the results will improve.  If it would be practical, try:

• Multiply costs * 1000
• Sum up the new costs
• Divide the totals by 1000

Those totals should be closer.