Odd behavior doing join and - sql

Question

create table umd2
as select a.permno, a.date, a.realdate, exp(sum(log(1+b.ret))) - 1 as cum_return
from msex2 (obs=50 keep=permno date realdate) as a, msex2 (obs=50 keep=permno date ret) as b
where a.permno=b.permno and 0<=intck('month', b.date, a.date)<3
group by a.permno, a.date
having count(b.ret)=3;
This query is to calculate the momentum (cumulative return in the past 3 month of a stock). However, this gives me duplicate rows. I thought group by would not return duplicate rows?
When I added the realdate column to my group by statement,
create table umd2
as select a.permno, a.date, a.realdate, exp(sum(log(1+b.ret))) - 1 as cum_return
from msex2 (obs=50 keep=permno date realdate) as a, msex2 (obs=50 keep=permno date ret) as b
where a.permno=b.permno and 0<=intck('month', b.date, a.date)<3
group by a.permno, a.date, a.realdate
having count(b.ret)=3;
those duplicate rows disapear. Why is this?

Answer 1

This is the way that SAS behaves. SAS recognizes the following query:
select a.permno, a.date, a.realdate, count(*)
from <whatever>
group by a.permno, a.date, a.realdate;
as being an aggregation query. That means that the rows are aggregated and reduced, with one result row per combination of the three columns. In particular, the non-aggregated columns in the select match (or are a subset) of the columns in the group by.
When you do this:
select a.permno, a.date, a.realdate, count(*)
from <whatever>
group by a.permno, a.date;
You are now using non-standard SQL. Most databases would generate an error. MySQL would accept this syntax, and assign an arbitrary value to a.read_date from the matching values. SAS does something different. SAS says: "Well, you clearly don't intend for this to be an aggregation query." So, it doesn't aggregate the rows, but it appends the aggregated values onto each row. In other databases, you would do this using window functions.
Technically, SAS calls this remerging summary data, which is documented here.