Solved
Contributor
Posts: 45

# Proc ttest: class continuous variable

Hi all,

I have a simple mission yet with a complication I have not found a way to solve. I want to do a two-sample t-test by a continuous variable.

Say I have variables con1 and con2, both are continuous. I want to see if in group 1 (con1>=median of con1 across whole sample) vs. in group 2 (con1<median of con1 across whole  sample), values of con2 are different. Below is what I have right now:

proc univariate data=have; var con1; run; /* manually get con1 median, say the median is 1.523*/
data a1; set have;

if con1>= 1.523 then dummy=1; if con1< 1.523 then dummy=0; if con1=. then dummy=.;

run;
proc ttest data=a1; class dummy; var con2; run;

Essentially I manually get the median of con1, and make a categorical variable based on the median ("dummy" in the above example) . The problem with this is: due to rounding, though the median really is 1.522823104, SAS output would show median of 1.523. And this makes the sample sizes of group1 and group2 vastly different. I want to make sure that group1 and group2 have the same size (or different by 1 if whole sample size is an odd number). Is there an easy way of doing this?

I appreciate any help!

JOL

Accepted Solutions
Solution
‎08-22-2016 07:16 PM
Super User
Posts: 13,583

## Re: Proc ttest: class continuous variable

You can have Univariate output percentiles using the OUTPUT statment. The values will have more precision than the output really designed for people to look at which will likely be rounded as you say.

But to create groups Proc Rank may be much easier

``````proc rank data=have groups=2;
var con1;
ranks dummy;
run;``````

The dummy will have 0 for the values below the median and 1 for above. Note no need for univariate. You can use Proc rank to generate different numbers of groups: 4 would be quartiles, 10 deciles, 100 percentiles.

All Replies
Solution
‎08-22-2016 07:16 PM
Super User
Posts: 13,583

## Re: Proc ttest: class continuous variable

You can have Univariate output percentiles using the OUTPUT statment. The values will have more precision than the output really designed for people to look at which will likely be rounded as you say.

But to create groups Proc Rank may be much easier

``````proc rank data=have groups=2;
var con1;
ranks dummy;
run;``````

The dummy will have 0 for the values below the median and 1 for above. Note no need for univariate. You can use Proc rank to generate different numbers of groups: 4 would be quartiles, 10 deciles, 100 percentiles.

Contributor
Posts: 45

## Re: Proc ttest: class continuous variable

Hi ballardw,

Thanks for the response! I did what you suggested, and it worked much better (though still not what I had hoped for, but I suppose I can settle with the much smaller sample size difference).

For future reference, below is what I had in the end:

``````proc rank data=have groups=2 out=want;
var con1;
ranks dummy;
run;``````

Thanks for the help!

☑ This topic is solved.