How can I do a distinct sum? - sql

Question

I am trying to create a "score" statistic which is derived from the value of a certain column, calculated as the sum of a case expression. Unfortunately, the query structure needs to be a full outer join (this is simplified from the actual query, and the join structure survives from the original code), and thus the sum is incorrect, since each row may occur many times. I could group by the unique key; however, that breaks other aggregate functions that are in the same query.
What I really want to do is sum (case when ... distinct claim_id) which of course does not exist; is there an approach that will do what I need? Or does this have to be two queries?
This is on redshift, in case it matters.
create table t1 (id int, proc_date date, claim_id int, proc_code char(1));
create table t2 (id int, diag_date date, claim_id int);
insert into t1 (id, proc_date, claim_id, proc_code)
values (1, '2012-01-01', 0, 'a'),
(2, '2009-02-01', 1, 'b'),
(2, '2019-02-01', 2, 'c'),
(2, '2029-02-01', 3, 'd'),
(3, '2016-04-02', 4, 'e'),
(4, '2005-01-03', 5, 'f'),
(5, '2008-02-03', 6, 'g');
insert into t2 (id, diag_date, claim_id)
values (4, '2004-01-01', 20),
(5, '2010-02-01', 21),
(6, '2007-04-02', 22),
(5, '2011-02-01', 23),
(6, '2008-04-02', 24),
(5, '2012-02-01', 25),
(6, '2009-04-02', 26),
(7, '2002-01-03', 27),
(8, '2001-02-03', 28);
select id, sum(case when proc_code='a' then 5
when proc_code='b' then 10
when proc_code='c' then 15
when proc_code='d' then 20
when proc_code='e' then 25
when proc_code='f' then 30
when proc_code='g' then 35 end), count(distinct t1.claim_id) as proc_count, min(proc_date) as min_proc_date
from t1 full outer join t2 using (id) group by id order by id;

Answer 1

You can separate out your conditional aggregates into a cte or subquery and use OVER(PARTITION BY id) to get an id level aggregate without grouping, something like this:
with cte AS (SELECT *,sum(case when proc_code='a' then 5
when proc_code='b' then 10
when proc_code='c' then 15
when proc_code='d' then 20
when proc_code='e' then 25
when proc_code='f' then 30
when proc_code='g' then 35 end) OVER(PARTITION BY id) AS Some_Sum
, min(proc_date) OVER(PARTITION BY id) as min_proc_date
FROM t1
)
select id
, Some_Sum
, count(distinct cte.claim_id) as proc_count
, min_proc_date
from cte
full outer join t2 using (id)
group by id,Some_Sum,min_proc_Date
order by id;
Demo: SQL Fiddle
Note that you'll have to add these aggregates to the GROUP BY in the outer query, and the fields in your PARTITION BY should match the t1 fields you previously used in GROUP BY, in this case just id, but if your full query had other t1 fields in the GROUP BY be sure to add them to the PARTITION BY

Answer 2

You can use a subquery (by id and id_claim) and then regroup:
with base as (
select id, avg(case when proc_code='a' then 5
when proc_code='b' then 10
when proc_code='c' then 15
when proc_code='d' then 20
when proc_code='e' then 25
when proc_code='f' then 30
when proc_code='g' then 35 end) as value_proc,
t1.claim_id , min(proc_date) as min_proc_date
from t1 full outer join t2 using (id) group by id, t1.claim_id order by id, t1.claim_id)
select id, sum(value_proc), count(distinct claim_id) as proc_count, min(min_proc_date) as min_proc_date
from base
group by id
order by id;
See that I sugest avg for the internal subquery, but if you are sure that the same claim_id have the same letter you can use max or min and that was integer. If not is prefer this.