Ranking of the residual mean calculated by regression

[Choi_Jay]

I calculated the residual through the following regression analysis(with computer A and B).

proc reg data=data;
model m1=m2 m3 m4 ROA W1-W56 /*countries*/ I1-I66 Y1-Y9/ white VIF;
output out=d1 R=DA;
run; quit;

And I calculated the average of the calculated residuals for each country (W1-W56, W57), and ordered them by "proc sort" the average. I used the same data-set and sas-code except for computer (A and B).

The problem is that the ranking of the residual mean for each computer is different. The values of the residuals calculated by each computer are the same (visible, 10 decimal places), but differ from each other when the mean is calculated. For example, "2.398665E-17" and "1.878665E-17" etc.

The ttest result of the two residuals (PADA1 and PADA2) calculated by each computer is as follows.

Your advice is desperately needed. please.

[Rick_SAS · Accepted solution]

Yes. What you describe is well-known in numerical analysis and often happens. Different chip sets use different caches sizes, which results in round-off-sized differences (1e-16), even in simple operations such as addition.

The two analyses are equally valid, and the numbers are said to be "equal to machine precision."