topic Re: GLM ABSORB for Two-Way Fixed-Effect Model? in SAS Programming

GLM ABSORB for Two-Way Fixed-Effect Model?

Junyong — Thu, 17 Sep 2020 04:50:09 GMT

I use class in proc glm to run a one-way fixed-effect model as follows.

data have;
do i=1 to 3;
do j=1 to 3;
do k=1 to 10;
x=rannor(1);
y=x+rannor(1);
output;
end;
end;
end;
run;

proc glm data=have;
class i;
model y=x i/noint solution;
ods output parameterestimates=oneway_class;
quit;

Or sometimes I use absorb because it is faster than class.

proc glm data=have;
absorb i;
model y=x/noint solution;
ods output parameterestimates=oneway_absorb;
quit;

And I found that absorb, unlike class, cannot estimate a two-way fixed-effect model in this way.

/*THE FOLLOWING ESTIMATES TWO-WAY FIXED-EFFECT MODELS CORRECTLY*/
proc glm data=have;
class i j;
model y=x i j/noint solution;
ods output parameterestimates=twoway_class;
quit;

/*THE FOLLOWING DOES NOT ESTIMATE TWO-WAY FIXED-EFFECT MODELS*/
proc glm data=have;
absorb i j;
model y=x/noint solution;
ods output parameterestimates=twoway_absorb;
quit;

Though the first glm with class is correct, it becomes too slow because the real data have so many observations. How can I estimate a two-way fixed-effect model with absorb in glm instead?

Re: GLM ABSORB for Two-Way Fixed-Effect Model?

PaigeMiller — Thu, 17 Sep 2020 10:39:00 GMT

Correct.

If you use the ABSORB statement with two variables, ABSORB assumes that the second variable in the list is nested within the first, and this is not the design you have.

Re: GLM ABSORB for Two-Way Fixed-Effect Model?

Junyong — Thu, 17 Sep 2020 18:49:02 GMT

Thanks. Then should I give up the absorb approach for two-way fixed effects?

Re: GLM ABSORB for Two-Way Fixed-Effect Model?

PaigeMiller — Thu, 17 Sep 2020 21:29:55 GMT

@Junyong wrote:

Thanks. Then should I give up the absorb approach for two-way fixed effects?

I would replace the word "should" with "must"

Re: GLM ABSORB for Two-Way Fixed-Effect Model?

Junyong — Fri, 18 Sep 2020 03:25:19 GMT

Sad news since LSDV-estimating multi-way fixed-effect models is computationally costly, but thanks anyway.