topic Re: Proc Univariate: max and min in Statistical Procedures

Proc Univariate: max and min

Csands — Wed, 30 Oct 2013 16:38:45 GMT

Hi,

I did a proc univariate for a variable, and got in the extreme value tables and in the quantiles table different values for the maximum and minimum of the variable. As shown below.

Extreme Values
Lowest			Highest
Order	Value	Freq	Order	Value	Freq
1	-778.998.059	1	.	.	.
. . . .	. . . .	. . . .	. . . .	. . . .	. . . .
.	.	.	270756	3.990.158.000	1

Quantiles
Quantile	Estimate
100% Max	3.990.160.000
99%	1.190.450
95%	210.926
90%	97.597
75% Q3	29.182
50% Median	6.771
25% Q1	-2.523
10%	-17.403
5%	-36.202
1%	-174.758
0% Min	-778.998.000

Why are this values different?

Thanks in advance for the help.

Re: Proc Univariate: max and min

PGStats — Wed, 30 Oct 2013 18:38:35 GMT

That's a very good question. How many observations are in your dataset? Did you try playing with UNIVARIATE option PCTLDEF= ?

Re: Proc Univariate: max and min

Reeza — Wed, 30 Oct 2013 18:39:47 GMT

Given that FREQ=1 for each of them would it matter?

Which ones correct by the way?

Re: Proc Univariate: max and min

PGStats — Wed, 30 Oct 2013 21:47:17 GMT

I would bet the extreme values are exact. They are easy to get in a single pass of the dataset. Not the same for quantile estimates (other than Min and Max). They look like rounded values that most likely result from some binning algorithm. - PG

Re: Proc Univariate: max and min

ballardw — Wed, 30 Oct 2013 23:57:38 GMT

I'm willing to bet that the difference is the default table format for the two tables, probably a difference between best9 or best 10 and best 12.

Re: Proc Univariate: max and min

PGStats — Thu, 31 Oct 2013 02:37:47 GMT

Well, that hypothesis could be checked by exporting the tables to ODS datasets. - PG

Re: Proc Univariate: max and min

Csands — Thu, 31 Oct 2013 09:49:10 GMT

So....

The code I use is very simple.

PROC UNIVARIATE DATA = table NEXTRVAL=30;

BY byvar;

VAR var_to_study;

RUN;

I run this code for more than 30 continuous variables (as the var are continuos is ok to have a freq=1). For almost all the variables with values of more than 6 digits I have this issue. My variables have no formats, just a length 8 associated. The dataset I use as more than 500.000 rows, the results I showed before belong to a by group of more than 200.000 rows, and I can have missing values.

I already calculated a max and min with proc sql, and the results where the same as the extreme values table.

Of course I have several ways of getting the max and min .... but I was expecting to get the same values in all the methodologies.

Re: Proc Univariate: max and min

Reeza — Thu, 31 Oct 2013 15:06:24 GMT

What version of SAS are you on? I can't replicate the issue on SAS 9.3

I think I agree with BallardW that it's probably just a format issue with the template.

Add in this line before your code and see if the values in the dataset xvalues are what you expect:

ods table extremevalues=xvalues;

Re: Proc Univariate: max and min

PGStats — Thu, 31 Oct 2013 16:15:36 GMT

I can't reproduce the "Format Issue" with this simple test:

data _null_;
do x = -778998059, 3990158000;
     put x :best6.;
     put x :best7.;
     put x :best8.;
     put x :best9.;
     put x :best10.;
     put x :best11.;
     put x :best12.;
     end;
run;

-779E6
-7.79E8
-7.79E8
-7.79E8
-778998059
-778998059
-778998059
3.99E9
3.99E9
3.9902E9
3.99016E9
3990158000
3990158000
3990158000

I'm using version 9.3

Re: Proc Univariate: max and min

Reeza — Thu, 31 Oct 2013 16:38:24 GMT

Time for Tech Support

See the very fine print at the bottom of the page, I suggest submitting a ticket.

Re: Proc Univariate: max and min

Reeza — Thu, 31 Oct 2013 16:47:31 GMT

I assume there's actually a comma or european comma format on the data, because of the periods rather than commas?

Re: Proc Univariate: max and min

Csands — Thu, 31 Oct 2013 16:53:58 GMT

Reeza , I think you are right. It's time for SAS tech support.

I just posted the question here, because I notice some users with great knowledge of SAS, and probably someone already had the same issue, it turns out that I am the first .

Thanks all for your time.

(By the way my version is 9.3)