Fluorite | Level 6

## RANGAM vs. RAND("Gamma", )

Hello,

I am trying to simulate data from a gamma distribution. I originally used the RAND function, from my understanding this is a newer function. However, I also tried the older function RANGAM. When using the same shape (and scale) parameters I get very different graphs when plotting the data.  I'm obviously not understanding something correctly, could anyone advise on why one would use RAND vs. RANDGAM? I assumed there really shouldn't be any major differences.

Sample Code I tried

%let shape =0.15;

%let scale = 0.25;

data test;

do i = 1 to 1000;

do until (x>=0.2);

xnew=&scale.*rand("Gamma", &shape.);

xold=&scale.*randgam(1,&shape.);

end;

output;

end;

run;

proc univariate data=test;

var xnew xold;

histogram/gamma;

inset gamma/pos=ne;

run;

The data I'm trying to simulate needs to have a lower bound no less than 0.2, which is why I have the do until loop above. I'm hoping this is appropriate.

1 ACCEPTED SOLUTION

Accepted Solutions

## Re: RANGAM vs. RAND("Gamma", )

Hello @jonesk48,

The major difference between the RAND function (with first argument "Gamma") and the RANGAM function is the underlying pseudo-random number generator: The RAND function uses the famous Mersenne Twister, whereas the older RANGAM and other RANxxx functions use an older generator with inferior statistical properties. So, for professional simulations the RAND function should be preferred. Please see the documentation for more details. You can use the CALL STREAMINIT routine to specify a seed value (as you did in the first argument of RANGAM).

5 REPLIES 5
Fluorite | Level 6

## Re: RANGAM vs. RAND("Gamma", ) -

Well.....Please ignore.... I realized I needed to debug my own code! I see there is no difference when I correct my code.

%let shape =0.15;
%let scale =0.25;

data test;
do i = 1 to 1000;
do until (xnew>=0.2);
xnew=&scale.*rand("Gamma", &shape.);
END;
OUTPUT;
end;
run;

data test1;
do i = 1 to 1000;
do until (xold>=0.2);
xnew=&scale.*rand("Gamma", &shape.);
xold=&scale.*rangam(100,&shape.);
END;
OUTPUT;
end;
run;

proc univariate data = test;
var xnew;
histogram/ gamma;
inset gamma/pos=ne;
run;

proc univariate data = test1;
var xold;
histogram/ gamma;
inset gamma/pos=ne;
run;

## Re: RANGAM vs. RAND("Gamma", )

Hello @jonesk48,

The major difference between the RAND function (with first argument "Gamma") and the RANGAM function is the underlying pseudo-random number generator: The RAND function uses the famous Mersenne Twister, whereas the older RANGAM and other RANxxx functions use an older generator with inferior statistical properties. So, for professional simulations the RAND function should be preferred. Please see the documentation for more details. You can use the CALL STREAMINIT routine to specify a seed value (as you did in the first argument of RANGAM).

Fluorite | Level 6

## Re: RANGAM vs. RAND("Gamma", )

@FreelanceReinh Thank you for the explanation! I will stick with the RAND statement and will consider using the STREAMINIT option. Thanks for taking time to reply!

Super User

## Re: RANGAM vs. RAND("Gamma", )

Check Rick's blog for the complete explanation  RANDXXX() v.s. RAND() .

http://blogs.sas.com/content/iml/2013/07/10/stop-using-ranuni.html

Fluorite | Level 6

## Re: RANGAM vs. RAND("Gamma", )

Hi @Ksharp,

Thank you for providing the link to Rick's article. I skimmed over it but will definitely read it thoroughly tomorrow. Appreciate you taking the time to respond to my question.

Sincerely,

Kendra

Discussion stats
• 5 replies
• 2874 views
• 11 likes
• 3 in conversation