How much time does the -meanonly- option of -summarize- save?

[Svend Juul]

Nick and I had a mini-discussion about the benefits of the -meanonly- option of summarize. I did an experiment on my 4-year old 32-bit PC, Stata SE 13.1:

clear
set obs 1000000
gen x=runiform()
set rmsg on
quietly summarize x
summarize x, meanonly

quietly summarizing a million observations took 0.08 seconds.
With the meanonly option it took 0.05 seconds.
I think I can do without the meanonly option.

[ben earnhart]

I dunno. I ran your code on a 64-bit-but-older Core-2 Duo, and got .12 seconds vs .03 seconds. Did it with cases bumped up by a factor of 10, and got .69 vs .31. Bumped by a factor of 100, got 5.87 seconds vs 3.04 seconds. So 50-75% improvement, in my environment. Like many things we try to benchmark, there are so many unknown things impacting performance that it's nearly impossible to have a true comparison.

Hmm. Ran it again (we weren't setting seeds) and got .06 vs .05 seconds when I took it back to a million cases (I was surprised your 32-bit was faster than my 64). So there is a lot of variability in what happens here. Need charts of sample size, with numerous seeds, across platforms to draw any conclusions.

[Nick Cox]

Every argument can be correct here.

In Svend's example I doubt that, working interactively, anyone could type , meanonly in 0.03 s, to say nothing about the display or return list necessary to see results.

But I see heavily-used programs in which summarize is used again and again, just for a mean or count or sum, and I hear of many people to whom a million observations is tiny. So, there are cases in writing a program where making it faster is a simple and good idea, and people here often benefit from such care, albeit unknowingly.