Date:         Sun, 16 Jan 2005 14:44:06 -0500
Reply-To:     "Zack, Matthew M." <MMZ1@CDC.GOV>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "Zack, Matthew M." <MMZ1@CDC.GOV>
Subject:      Re: surveyselect question
Content-Type: text/plain; charset="us-ascii"

What if you randomly select patients and all their visits without PROC
SURVEYSELECT?

  * Sort patient visits;
  *  by patient ID;

  proc sort;
    by pt;
  run;

  * Generate a uniform random number for each patient;

  data two(drop=rnseed);
     retain rnseed 6093141 rn;
     set;
     by pt;
     if (first.pt eq 1)
        then rn=uniform(rnseed);
     output two;
  run;

  * Sort patient visits;
  *  by patient ID, visit ID, and ascending uniform random number;

  proc sort data=two;
     by pt visit rn;
  run;

  * Select about 50 [+- 2 visits so that range=48 to 52] total patient
visits;
  *  Add five visits (possibly from different patients) after the 50
above are selected;
  *   where SBP=. or DBP=.;

  data visit50(drop=rn lstvisit nmissbp);
     retain lstvisit nmissbp 0;
     set two;
     by pt visit;
     select;
        when (lstvisit eq 0) do;
           if ((ABS(50-_n_) le 2) and
               (last.pt eq 1))
              then lstvisit=1;
           output visit50;
        end;
        when (lstvisit eq 1) do;
           if ((sbp eq .) or
               (dbp eq .)) then do;
                nmissbp=nmissbp+1;
              if (nmissbp le 5)
                 then output visit50;
                   else lstvisit=2;
           end;
        end;
        otherwise stop;
     end;
  run;

  * Select about 20% of the input data set;
  *  Add five visits (possibly from different patients) after the above
20% are selected;
  *   where SBP=. or DBP=.;

  data visit20p(drop=rn nmissbp);
     retain nmissbp 0;
     set two;
     by pt visit;
     select;
        when (rn le 0.20) output visit20p;
        otherwise do;
           if ((sbp eq .) or
               (dbp eq .)) then do;
                nmissbp=nmissbp+1;
              if (nmissbp le 5)
                 then output visit20p;
                   else stop;
           end;
        end;
     end;
  run;

Matthew Zack

-----Original Message-----
From: SAS(r) Discussion [mailto:SAS-L@LISTSERV.UGA.EDU] On Behalf Of
Scott
Sent: Sunday, January 16, 2005 1:12 AM
To: SAS-L@LISTSERV.UGA.EDU
Subject: surveyselect question


Hi,

I've read various posts about SURVEYSELECT and random samples in the
archives, but couldn't find the answer to my problem, thus this post...

Say I have a dataset:

PT VISIT SBP DBP, where

PT     = patient
VISIT  = visit number, say 1 - 4, which may be incomplete for a given
PT, i.e. could be 1; 1,2,4; 1,2,3; 1,3; etc.
SBP    = systolic blood pressure
DBP    = diastolic blood pressure  (both BP's could have missing values)

I'd like to sample this dataset as follows:

1.  Sample has "around" say 50 observations in total.

2.  Sample has say 20% of observations from input data set.

In both of these samples, *** ALL observations for a given PT are
included ***, i.e. if PT 7 is one of the patients randomly selected,
then all visits for that PT are included in the random sample.

3.  #1 and #2 above, augmented by say 5 random observations where either
SBP, DBP, or both have a missing value.

For #3, I don't care if I make two passes over the data, but one pass
would be nice.

IOW, in "pseudocode":

1.  If each PT had 4 visit records, I would have either 12 (48) or 13
(52) observations in the sample dataset, since I specified a sample size
of around 50.

2.  If each PT had 4 visit records, and the total input dataset is 1000
observations, I would have 200 observations in the sample dataset,
comprised of 50 PTs with 4 visits each.

3.(1) 12 or 13 random patients, plus 5 observations where SBP, DBP, or
both were missing.

3.(2) 50 random patients, plus 5 observations where SBP, DBP, or both
were missing.

I've played with SURVEYSELECT, but can't figure out how to get all
records for a given PT to be included in the output.

Note that this sampling is for QC tests of code algorithms, not for
further statistical analyses of the resulting sample.

Thanks,
Scott