Conditional perform sum by group

[lpy0521]

Hi,

I have question that how to conditional perform vertical sum as following: 

Data have:

ID	VAR1	VAR2	COE1
A	2	4	0.2
A	NA/.	NA/.	0.3
A	1	3	0.5
B	4	6	0.1
B	7	9	0.2
B	1	3	0.7
C	5	6	0.5
C	2	3	0.5

I want to do: if there is NA or dot  in var1 or var2, then sum the rest vertically by ID, otherwise calculate the sum product between Var1, Var2 and COE 1.

Data want:

ID	VAR1	VAR2
A	3	7
B	2.5	4.5
C	3.5	4.5

where in data want,

Var1(A) = 2+1, Var2(A) =4+3;

Var1(B) =4*0.1+7*0.2+1*0.7=2.5 Var2(B) = 6*0.1+9*0.2+3*0.7=4.5;

Var1(C) = 5*0.5+2*0.5 = 3.5, Var2(C) = 6*0.5+3*0.5 = 4.5

Thanks!

[VDD]

here is a data step that gives your requested results

data have;
input ID $ var1 var2 coe1;
cards;

A 2 4 0.2
A NA/. NA/. 0.3
A 1 3 0.5
B 4 6 0.1
B 7 9 0.2
B 1 3 0.7
C 5 6 0.5
C 2 3 0.5
;
proc sort data=have out=have2;
	by ID;
run;

data want(drop = var1_total var2_total coe1);
	set have2;
	by id;
	if first.ID then
		do;
			var1_total = 0;
			var2_total = 0;
		end;
	var1_total + var1;
	var2_total + var2;
	if last.ID then
		do;
			var1 = var1_total;
			var2 = var2_total;
			output;
		end;
run;

[lpy0521]

Thanks for your suggestion. But it gives the sum for B C, instead of the sum product. By sum product I mean:

Var1(B)*COE(B) + VAR2(B)*COE(B).

Thanks!

[andreas_lds]

Lets solve the issue with the input data first: the string "NA/." can't be stored in a numeric variable. var1 and var2 must be numeric to be used in calculations. So please post the example data as data-step using the data-types you have in your original dataset.

Except for the special treatment when var1 or var2 are missing in a group, this looks like weighted sums.

[lpy0521]

Sorry for the confusion here. What I mean by "NA/." is actually missing values and there are almost half of my var1 and var2 are missing. So basically I want to perform:

For each ID, if there is missing value in Var1 or Var2, then sum those non-missing value up, if there are no missing value, then calculate the weighted sum with corresponding COE1.

Thanks for your comment!

[andreas_lds]

Not 100% that this works, but it seems to provide the requested results:

proc sql;
  /* find all IDs having var1 and/or var2 missing */
   create table work.WithMissing as
      select distinct id
         from work.have2
            where missing(var1) or missing(var2)
   ;
  /* set coe1 to 1 for the IDs */
   update work.have2 
      set coe1 = 1 
         where id in (select id from work.WithMissing)
   ;
quit;

proc summary data=work.have2 nway;
   class id;
   var var1 var2;
   weight coe1;
   output out=work.want(drop= _type_ _freq_) sum=;
run;

[s_lassen]

It can be done using SQL:

proc sql;
  create table want as select
    ID,
    case nmiss(var1)+nmiss(var2)
      when 0 then sum(var1*coe1)
      else sum(var1)
    end as var1,  
    case nmiss(var1)+nmiss(var2)
      when 0 then sum(var2*coe1)
      else sum(var2)
    end as var2
  from have
  group by id;
quit; 

 Or you can do it with datasteps like this:

data miss;
  set have(where=(var1 is null or var2 is null))
  by id;
  if first.id;
  keep id;
run;

data want;
  sum1=0;
  sum2=0;
  do until(last.id);
    merge miss(in=miss) have;
    by id; 
    if miss then do;
      sum1+var1;
      sum2+var2;
      end;
    else do; 
      sum1+var1*coe1;
      sum2+var2*coe1;
      end;
    end;
  drop sum1 sum2;
  var1=sum1;
  var2=sum2;
run;

[Astounding]

Assuming that your variables are actually numeric, and that your data is sorted ...

Why not just compute both and track which one you want to use.  For example:

data want;

var1_sum=0;

var1_weighted=0;

var2_sum=0;

var2_weighted=0;

do until (last.id);

   set have;

   by id;

   var1_sum + var1;

   var2_sum + var2;

   if var1 > . then var1_weighted + var1*coe1;

   else var1_missing=1;

   if var2 > . then var2_weighted + var2*coe1;

   else var2_missing=1;

end;

if var1_missing then var1=var1_sum;

else var1= var1_weighted;

if var2_missing then var2=var2_sum;

else var2=var2_weighted;

drop var1_: var2_: ;

run;