@s_manoj: Your code runs a de facto Cartesian product reading every obs from one copy of NEW for every obs in another, and it doesn't even take advantage of the sorted order. Hence, it will scale extremely poorly as the input size grows. At nobs=100000, an efficient solution would run in under a second, yet your program would take good 30 minutes. And at nobs=1000000 (which is a very small data set by today's standards) it would take upward of 50 hours. The point= approach could be viable (though still slower than anything else) if you take advantage of the sorted order BY VISIT BMI and (a) use point= to read only the observations between the endpoints of the current BY group and (b) leave the point= loop as soon as the 10% condition is false. Paul D.
... View more
