ATT Weights from PROC CAUSALTRT with METHOD=IPWR

blund — Wed, 19 Apr 2023 16:31:36 GMT

This documentation link gives the formula for ATT weights (Let ps=propensity score, t is treatment)

https://documentation.sas.com/doc/en/pgmsascdc/9.4_3.4/statug/statug_causaltrt_details06.htm

The formulas for the weights are:

Control: Let CTOTAL = Sum[ (1-t)*ps/(1-ps) ]. Then a control weight is ( (1-t)*ps/(1-ps) ) / CTOTAL

Treated: Let TTOTAL = Sum[ t ]. Then a treated weight is t / TTOTAL

I don’t see how the weights from the PROC CAUSALTRT example are in agreement with the formulas. See column ATTwgt2 in the example below.

In particular, I think the weights for the treated observations should all equal 0.2 … which is not in agreement with column ATTwgt2. Any help is appreciated!

data WORK;

input Y T X;

datalines;

1 0 1

2 1 1

3 0 1

1 1 1

1 0 3

1 1 3

4 0 4

2 0 4

3 1 4

2 0 4

3 0 4

2 1 4

;

PROC CAUSALTRT DATA= WORK ATT;

CLASS X;

PSMODEL T(descending)=X;

MODEL Y;

OUTPUT OUT=OUTwork2 PS=PS2 IPW=ATTwgt2;

proc print data = OUTwork2;

run;

Obs Y T X PS2 ATTwgt2

1 1 0 1 0.50000 2.00000

2 2 1 1 0.50000 2.00000

3 3 0 1 0.50000 2.00000

4 1 1 1 0.50000 2.00000

5 1 0 3 0.50000 2.00000

6 1 1 3 0.50000 2.00000

7 4 0 4 0.33333 1.50000

8 2 0 4 0.33333 1.50000

9 3 1 4 0.33333 3.00000

10 2 0 4 0.33333 1.50000

11 3 0 4 0.33333 1.50000

12 2 1 4 0.33333 3.00000

Re: ATT Weights from PROC CAUSALTRT with METHOD=IPWR

jiltao — Wed, 19 Apr 2023 18:06:18 GMT

The weight in the OUTPUT OUT= ipw= statement in PROC IRT is the weight values for ATE. For the weight values for ATT, you can use the following data step code to obtain it --

data outwork2;

set outwork2;

ipwATT = ps2 * ipw2;

run;

where the data set outwork2 is from the OUTPUT OUT= statement in PROC CAUSALTRT --

PROC CAUSALTRT DATA= WORK ATT;

CLASS X;

PSMODEL T(descending)=X;

MODEL Y;

OUTPUT OUT=OUTwork2 PS=PS2 IPW=ipw2;

run;

We will add the IPW values for ATT in the next release.

Thanks,

Jill

Re: ATT Weights from PROC CAUSALTRT with METHOD=IPWR

blund — Wed, 19 Apr 2023 18:19:06 GMT

OK ... thanks for the response.

Bruce Lund

topic Re: ATT Weights from PROC CAUSALTRT with METHOD=IPWR in Statistical Procedures

ATT Weights from PROC CAUSALTRT with METHOD=IPWR

Re: ATT Weights from PROC CAUSALTRT with METHOD=IPWR

Re: ATT Weights from PROC CAUSALTRT with METHOD=IPWR