I have searched extensively but have been unable to find any definitive guidance for whether, and how, I can test the INequality of coefficients across regressions.
For example, I run the resgressions:
reg y x1 x2 if Negative==1
estimates store neg
reg y x1 x2 if Negative==0
estimates store pos
Then I do:
suest neg pos
The test for x1 (in the negative partition) = x2 (in the positive partition) is:
test [neg_mean]x1 = [pos_mean]x1
But I am not sure how to test whether x1(neg)>x1(pos) or x1(neg)<x1(pos)
The closest I can find (or at least I think it may be getting me close) came from http://www.stata.com/support/faqs/st...-coefficients/ but it is for coefficients within a regression, not across.
Following the above, I could test x1(neg)>=x1(pos) by doing:
local sign_negpos = sign([neg_mean]x1-[pos_mean]x1)
display "H_: neg coef >= pos coef. p-value = " normal(`sign_negpos'*sqrt(r(chi2)))
This gives me a p-value that is half of the p-value given by the equality hypothesis test, which makes me think the test may be good. However, I don't know if this is a valid test or if it is reliable. If it is a valid test, can I just divide the p-value for the equality Chi-square test in half as I typically would for directional hypotheses? And what to do about the Chi-squared coefficient...double it, or leave it?
Thanks in advance!
For example, I run the resgressions:
reg y x1 x2 if Negative==1
estimates store neg
reg y x1 x2 if Negative==0
estimates store pos
Then I do:
suest neg pos
The test for x1 (in the negative partition) = x2 (in the positive partition) is:
test [neg_mean]x1 = [pos_mean]x1
But I am not sure how to test whether x1(neg)>x1(pos) or x1(neg)<x1(pos)
The closest I can find (or at least I think it may be getting me close) came from http://www.stata.com/support/faqs/st...-coefficients/ but it is for coefficients within a regression, not across.
Following the above, I could test x1(neg)>=x1(pos) by doing:
local sign_negpos = sign([neg_mean]x1-[pos_mean]x1)
display "H_: neg coef >= pos coef. p-value = " normal(`sign_negpos'*sqrt(r(chi2)))
This gives me a p-value that is half of the p-value given by the equality hypothesis test, which makes me think the test may be good. However, I don't know if this is a valid test or if it is reliable. If it is a valid test, can I just divide the p-value for the equality Chi-square test in half as I typically would for directional hypotheses? And what to do about the Chi-squared coefficient...double it, or leave it?
Thanks in advance!
Comment