Test for differences in coefficients after SUR estimation

[Johann Beier]

Dear Stata Users,

I'm currently working with a SUR (Seemingly Unrelated Regression) Model, which I managed to implement properly. I estimated the SUR-Model for two different groups of companies in my sample (via a sample split). What I want to do now is to test the difference of those estimated coefficients in the SUR-Mdoel across these subsamples with a bootstrapping method.
What I've read so far is that I need to estimate another simultaneous regression that compares the coefficents of my two subsamples via suest.

My first question is if this the right way to compare the coefficents and how do I do this after the SUR estimation?
And secondly: How do I combine this with a bootstrapping method?

Many thanks for your help,

Johann

[Joseph Coveney]

I believe that suest doesn't support sureg.

This sounds like a so-called Chow test. Following by analogy the StataCorp FAQ, couldn't you just form interaction terms and then use either test or lrtest?

sureg (response1 i.group##c.(predictor1 predictor2)) (response2 i.group##c.(predictor1 predictor2))

test [response1]1.group, notest
test [response2]1.group, accumulate notest
test [response1]1.group#c.predictor1, accumulate notest
test [response2]1.group#c.predictor1, accumulate notest
test [response1]1.group#c.predictor2, accumulate notest
test [response2]1.group#c.predictor2, accumulate

// or

estimates store Full
sureg (response1 c.(predictor1 predictor2)) (response2 c.(predictor1 predictor2))
lrtest Full

[Johann Beier]

Dear Jospeh,

many thanks for your answer. Yes, I also figured out that suest does not support sem (Error: sem is not supported by suest r(322). As I setup the SUR model with sample split I'd like to stick on that. Is there any alternative to suest when I'm using a SUR model in that case?

EDIT: Further restriction is that I'm using the sem command to implement the multi-equation model (instead of sureg), which does not allow for factor variables.

many thanks again,

Johann

[Joseph Coveney]

. . .I'm using the sem command to implement the multi-equation model (instead of sureg), which does not allow for factor variables

1. Use -sureg- instead, which allows factor variables.

2. Use -gsem- instead, which allows factor variables.

3. Stick with -sem-, and -generate- the interaction terms manually as shown in the FAQ linked to in my post.

[Johann Beier]

Dear Joseph,

again many thanks for your help. I managed to translate my original sem model to a gsem model. To get my situation more clear:

1. I estimate the gsem model with nine dependent variables and three predicting variables which I interact with the factor variables (my group dummy)

quiet   gsem            (response1 <-  i.group##c.(predictor1 predictor2 predictor3 ))/*
                                    */   (response2 <-  i.fgroup##c.(predictor1 predictor2 predictor3 )) /*
                                       .
                                       .
                                       .
                                       .
                                      */  (response9 <-  i.group##c.(predictor1 predictor2 predictor3 ))/*
                                      */   , nocapslatent vce(ID) const( $constraints )

2. Now I just want to estimate whether the predicted coefficent of predictor1 is different for both groups (dummy = 1 vs. dummy = 0) for all the respose variables.


Many thanks again!

Regards,

Johannes

[Phil Bromiley]

If after any estimator you enter estimator, coefl [e.g., gsem, coefl] it will show you how you can refer to the parameters. Then a simple test may work.

[Johann Beier]

Dear Phil,

many thanks for your answer it worked with

gsem, coflegend

then a simple test of differences in coefficients. However refering to the start of my question I also want to do sample splits. Is it possible (maybe with estimates store?) to save the information of coeflegend (the coefficients and standard errors) after each gsem prediction and then use it later in a test on differences in coefficients?

regards and many thanks again,

Johann

[Johann Beier]

Hello,

sorry to bother again, but I still got no solution to my problem.

Any help is very much appreciated.

regards,

Johann