Date:         Tue, 21 Mar 2006 12:20:24 -0800
Reply-To:     David L Cassell <davidlcassell@MSN.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         David L Cassell <davidlcassell@MSN.COM>
Subject:      Re: Statistical Question--PROC LOGISTIC
In-Reply-To:  <OF944F7851.4083BF19-ONC2257138.004551A3-C2257138.0045DD7C@hsbc.com.tr>
Content-Type: text/plain; format=flowed

BoraYavuz@hsbc.com.tr wrote:
>David,
>
>Can you expand a bit more on "certainty sampling" which you mentioned.
>
>We frequently find ourselves in pretty much the same situation Nick
>described and deal with it - mostly - using weights manually calculated in
>Excel (after deciding on the stratification variables through common
>sense).  And I believe it would be great if you could provide some more
>info (examples, your comments, etc.) and pointers on the subject matter.
>
>I also didn't understand the "size multipliers" that you mentioned.  :-(
>I'd be grateful if you could expand on that one too.

When we build a sample that has different weights (within a stratum or
with no strata at all) we do that by picking a 'multiplier' so that we pick
some records with a higher likelihood than others.  The variable that we
use for this is the variable we list in the SIZE statement.

If our boss comes to us and says:

"Okay Bora, here's what I need and I need it last week.  So step on it!
I want a sample of 40,000 from the database.  Yeah, yeah, I know you
pulled one yesterday, but this time I need it different.  I need the people
with incomes under $10,000 sampled at only one-tenth the rate we use
on the people with incomes over $10,000 .  And I need every single one
of the people with an income over $100,000 .  I expect to see this in
my inbox by close of business today!"

Okay, maybe our boss isn't that nice.  :-)

But now we have certainty sampling (we have to get all the high-income
people) and we have PPS sampling.  PPS= Probability Proportional to Size.

Let's do this now.  The SIZE variable gets used in the certainty sampling
part too, so we need to think about this.  We want a multiplier which is
10 times larger for the medium class than the low class:

  if income > 10000 then mult = 10;
    else mult = 1;

Or we could use a Boolean and write it as:

  mult = 1 + 9*(income > 10000);

But we also need that certainty sample.  The SIZE variable works with the
certainty option CERTSIZE like this: we give the largest values of the
multiplier to the records to be sampled for certain, and we use the
CERTSIZE option to tell the system what that cut-off will be.  So let's
tack that extra bit on:

  if income > 100000 then mult = 20;
    else if income > 10000 then mult = 10;
    else mult = 1;

Or we use our little Boolean trick again:

  mult = 1 + 9*(income > 10000) + 10*(income > 100000);

Now we can do both the certainty sampling *and* the weighted sampling
together:


proc surveyselect data=YourBigData out=YourSample
     seed=40589584
     method=pps
     certsize=20
     sampsize=40000;
  size mult;
  run;


Does that make more sense now?

As for what I wrote before about certainty sampling, I'm copying the
URL so you can find it in the SAS-L archives.

http://listserv.uga.edu/cgi-bin/wa?A2=ind0603C&L=sas-l&P=R28796

HTH,
David
--
David L. Cassell
mathematical statistician
Design Pathways
3115 NW Norwood Pl.
Corvallis OR 97330

_________________________________________________________________
Is your PC infected? Get a FREE online computer virus scan from McAfeeŽ
Security. http://clinic.mcafee.com/clinic/ibuy/campaign.asp?cid=3963