F test after regression using factor variables

[Karl Yesler]

Hi Stata list,

I am trying to perform a post estimation partial-F (Chow) test after a regression specified using a factor variable.

I have a model of the effect of price on the probability of purchase of different goods. Since the level of purchase and the effect of price may both differ by the good, I estimate a different intercept and slope for each good (a specific type of fixed effect model, I think). Here's the code:

glm yvar_ind c.price##i.good, family(binomial) link(probit)

I'd like to test whether the slopes for the different goods really differ. Per http://www.jblumenstock.com/files/co...4/FEModels.pdf, I am performing the
partial-F (Chow).

To do this after running the regression what is the appropriate code. Is it just:

test c.price#i.good

Thanks!

[Clyde Schechter]

Assuming the levels of good are 1, 2, 3, 4, ..., 12, and that the reference category was 1, I think the simplest test of the null hypothesis that all the slopes are equal would be

testparm price#i(2/12).good, equal

If the levels of good are an irregular set of numbers and too large to conveniently enumerate:

levelsof good, local(goods)
testparm price#i(`goods').good, equal

[Richard Williams]

I agree with the testparm but I am not sure about adding the -equal- option. It seems to me that this would allow good #1 to have a different slope than goods 2-12. Instead I think you want to test whether all the interactions are 0, which would be written as

testparm c.price#i.good

I may be misunderstanding the problem but give it a try and see how it looks.

[Karl Yesler]

Ok, thank you both. I will give both a try. The dataset is large, so the regressions will take over a few hours.

[Richard Williams]

Why aren't you just using the probit command? I believe it is more efficient than glm when estimating the same model.

If the dependent variable is a proportion and you were using Stata 14, you could use the fracreg command.

[Clyde Schechter]

Re #4. No, don't do it both ways. Richard is right; I had misunderstood your original post, but on re-reading it I see that the use of the -equal- option would not be correct here.