Hello @cuan,
Here's another, independently developed solution, so you can validate one against the other.
/* Create larger test dataset */
data have(drop=phase);
call streaminit(27182818);
do id=1 to 50000;
date='01JAN2000'd+rand('integer',10);
do phase=1 to rand('integer',50);
treat=rand('table',0.5);
output;
date+rand('integer',1,50); /* "('integer',0,50)" would cause ambiguities */
end;
end;
format date date9.;
run;
/* Create dataset with periods of treatment 1 */
data want(drop=date _: rename=(id=subjid));
array _t[%sysevalf('01JAN2000'd):%eval(%sysfunc(today())+100)] _temporary_;
do until(last.id);
set have;
by id date;
if first.id then _first_trt1=date;
if treat~=1 then do;
if _trt2_b=. then _trt2_b=date;
_trt2_e=date;
end;
if _trt2_e>. & (treat=1 | last.id) then do;
do _d=_trt2_b to _trt2_e;
_t[_d]=2;
end;
call missing(of _trt2:);
end;
end;
do _d=_first_trt1 to date;
if _t[_d]=. then _t[_d]=1;
end;
do _d=lbound(_t)+1 to hbound(_t)-1;
if _t[_d]=1 then do;
if _t[_d-1]~=1 then start=_d;
if _t[_d+1]~=1 then stop=_d;
end;
if n(start, stop)=2 then do;
treat=1;
output;
call missing(start, stop);
end;
end;
call missing(of _t[*]);
format start stop date9.;
run;
It's less elegant than @Shmuel's and also much slower on the large HAVE dataset created above (with more than one million observations). If your real dataset is much smaller (say, 1,000 patients, not 50,000), we're talking about fractions of a second, though. You can also improve performance by choosing more realistic lower and upper bounds (based on your data) for the temporary array _t, which just needs to cover all of your treatment dates +/- 1 day (which I assumed to fall between January 1, 2000 and 100 days from today).
Note that the resulting WANT datasets do not match for the large HAVE input dataset. Maybe they do with your real data because certain rare situations (like a patient treated only one day, with treat=1) don't occur in your data. You may want to check a few of the discrepancies (e.g. the single date difference for ID=6 of my simulated HAVE dataset) to get an idea of whether they are relevant for your application. The "ambiguities" mentioned in the comment above would arise if duplicate dates were allowed within an ID in dataset HAVE (e.g., two observations with the same ID and date, one with treat=1 and one with treat=2; it's not obvious what this would imply for the treatment on the preceding days, after an observation with treat=2). This would also cause more discrepancies between the WANT datasets of the two solutions.
Click image to register for webinar
