XTDPDGMM Diff-GMM Nocons option

[Jaspal Channa]

Hi all,

I am running a diff-GMM on Stata. Here is my code:

xtdpdgmm L(0/1).pcm unemployment marketshare hh union msxhh msxu1 msxlp, model(diff) collapse gmm(pcm, lag(2 2)) gmm(marketshare msxhh msxu1 msxlp, lag(1 1)) gmm(unemployment, lag(1 1)) gmm(hh union, lag(1 1)) teffects two vce(r) small

In my results, the constant has always been omitted and I thought that would be fine. But then I realised "cons" is often added as an option. When I add it, it ruins my results. Is this supposed to happen? Do I even need nocons?

Many thanks,
Jaspal

[Sebastian Kripfganz]

But then I realised "cons" is often added as an option.

What do you mean by that?

Without seeing you estimation output, there is not much I can say here.

[Jaspal Channa]

BRB. Editing.

[Jaspal Channa]

What do you mean by that?

Without seeing you estimation output, there is not much I can say here.

I misspoke - I meant to say that "nocons" is added as an option for diff-GMM using xtdpdgmm, but I cannot workout what difference it makes.

Here is my estimation output. For some reason the output changes depending on whether or not "nocons" is selected, despite it's omission in both cases. What is the reason for this?

Above is without "nocons"
Below is with "nocons"

[Jaspal Channa]

Upon further examination the discrepancy in results seems to be related to the time periods estimated. In the bottom screenshot, I used "nocons" and the co-efficients for timeids 44, 49 and 54 were omitted. When "nocons" was not used, the estimates for timeids 42, 49 and 54 were omitted.

[Sebastian Kripfganz]

That is very weird. With a full set of time dummies, the noconstant option should not make a difference. I am even more puzzled that some of your time effects are shown as "empty". I am afraid I cannot figure out what is going on here without actually seeing the data. If you feel comfortable sending me your data set by e-mail, I would be happy to have a look into it.

[Sebastian Kripfganz]

Here is what happened: The variables unemployment hh union are perfectly collinear with the set of time dummies. This is because these are global variables that are the same for all firms. You cannot include such variables together with time dummies. The time dummies already capture all of the variables that are constant across firms.

[Jaspal Channa]

Here is what happened: The variables unemployment hh union are perfectly collinear with the set of time dummies. This is because these are global variables that are the same for all firms. You cannot include such variables together with time dummies. The time dummies already capture all of the variables that are constant across firms.

Ahh, very interesting. Thanks for taking a look! Why does this happen with GMM, but not with fixed effects? I had thought that with standard fixed effects regressions, you can include both time dummies and global variables. Am I wrong?

Best,
Jaspal

[Sebastian Kripfganz]

The same issue arises in a conventional fixed-effects regression with time dummies.