How to run industry*year fixed effect in proc glm?

[Phil_NZ]

Hi all,

I want to ask how to control firm and industry*year fixed effects in proc glm?

My regression when regressing controlling for firm and year fixed effect is as below:

proc glm data=merge_treat_con_6th_may;
    ABSORB TYPE;
    class yr / truncate;
    model y= x1 x2 x3 yr/solution ss3;
run;
quit;

While TYPE and yr are firms and years accordingly.

Now I want to control for firm and industry*year fixed effect while INDC3 is industry indicator and character variable. Could you please let me know how to adjust the code above to do so?

Thank you and warm regards.

[PaigeMiller]

If INDC3 is an industry indicate and TYPE is the firm name, is TYPE nested underneath INDC3? Is there only a single INDC3 for each level of type? Is there is ever a value of TYPE that can be in two INDEC3 levels?

[Phil_NZ]

Hi @PaigeMiller 

Thank you for your question helping me to clarify my concern

I hope that you can explain me more about this part" is TYPE nested underneath INDC3". 

Is there only a single INDC3 for each level of type? ->  what do you mean by "each level of type then"?

Is there is ever a value of TYPE that can be in two INDC3 levels? ->No

Thank you and warm regards

[PaigeMiller · Accepted solution]

"Each level of TYPE" ... the variable TYPE has a certain number of discrete of values, that number is the number of levels — for example, a much smaller data that I have just made up contains values of variable TYPE equal to "Eastman Kodak","ON Semiconductor" or "TrueSense Imaging", that's 3 levels.

If each TYPE can be in only one INDC3, then TYPE is nested within INDC3.

So if you want INDUSTRY*YEAR in the model, I think you have to not use the ABSORB statement. Using the ABSORB statement, I don't think it can be done.

[Phil_NZ]

@PaigeMiller  I need to use ABSORB because I need to deal with the limitation of distinct observation (>32k something). So, you mean, no hope for that so far by using SAS?

[PaigeMiller]

Not that I know of. But that certainly shouldn't stop you from experimenting to see if it can be done.

Again, I state my skepticism that >32K levels of a variable are really useful in modeling.

[Phil_NZ]

Hi @PaigeMiller ,

Thank you for your help, but could you please explain this phrase to me , please?

" I state my skepticism that >32K levels of a variable are really useful in modeling"

Do you mean I may check again the number of distinct observations?

Warm regards.

[PaigeMiller]

I explained it in your earlier thread here: https://communities.sas.com/t5/SAS-Programming/Getting-multicolinearity-dealing-with-Simpson-Paradox...