Proc NPAR1WAY Wilcoxon Ranked Sum Test Continually Runs/Never Ends

[JackHP]

Hello,

When I run the below code SAS continually "runs" and I don't seem to ever get an output despite waiting over 10 minutes. My dataset has under 200 observations so I can't imagine that processing power is the issue. 

proc NPAR1WAY data=mvrx19.master wilcoxon correct=no;
title "compare PSS Medians by FI status";
class FI_1;
var PSS4NS_1 PSS10NS_1;
where psurvey1 ne '.'d;
exact wilcoxon;
run;

I tried taking out "Exact wilcoxon" which fixed this issue, but wondering why I couldn't get output when including it, and would like to have the option of generating exact p values.

Thank you!

I am running SAS University Edition with 6 GB RAM dedicated for the virtual Machine.

[FreelanceReinh · Accepted solution]

Hello @JackHP,

I wouldn't rule out that the computation of the two exact tests takes a long time. (There's a reason why the MAXTIME= option of the EXACT statement exists; see Computation Options.) To estimate the time needed I would restrict the input dataset using a suitable (extended) WHERE condition or an OBS= dataset option or by drawing random samples so as to use only, say, 10, 20, 40, ... observations and see how the run times increase. Of course, both FI_1 values must be represented in each of these smaller datasets.

Then, if the extrapolated run time for the full dataset was inacceptable, I would use the MC option of the EXACT statement (after a slash, see link above) and as long as the run times are tolerable add the N= option (MC is then redundant) with increasing values n>10000 (the default). You'll observe that the confidence limits of the exact p-values tend to get narrower as n increases. (As random variables they fluctuate even between runs of the same code, unless you use the SEED= option.) Finally, the Monte Carlo estimate with the largest tolerable n value might be the best estimate of the exact p-value that you can reasonably get.

[JackHP]

Hi @FreelanceReinh ,

Thank you for your detailed response and suggestions. I will take a look at restricting my sample to get an idea of processing time for smaller samples to extrapolate the run time for the whole sample - great idea! I appreciate your explanation and links to these resources, it is very helpful!