Dynamic Panel / GMM (EV 6): Update

[nadja123]

Hi there,

I have two questions concerning Dynamic panel regression / GMM:

1. Is it possible and how to display the regressed coefficients for instrumental variables?
2. Is it is possible to use certain "regressors" without letting them to be transformed (by differences or orthogonalization) by typing @lev(regressor) in the list of regressors? Does EV consider them then in levels?

Please can somebody help with this? Thanks!!!

[EViews Gareth]

I don't understand your first question, but the answer to the second is "yes".

[nadja123]

Dear QMS Gareth,

this is strange. Have you modified your answer while I have been writing this message? I would swear on that I first saw a "no" to the 2nd Q. This is what I wanted to write in a reply to "no":

2. no = so you mean no regressors will be transformed? This puzzles me, as I have tried to write the regressors as @lev(regressor) (eq01_levregressors), and alternatively as such (eq02_justregressors). It gives just different results and it does not report any error in specification. I attach a .wf1 with the alternative equations. Why is the output different, and what does EV actually estimate under each of them?

Sorry for repeating my question, just to be sure, I just re-ask: Is it okey to use @lev(regressor) in the DPD/GMM estimation? Do you mean that EV considers these regressors then in levels?

As for the 1st question: By displaying estimated coefficients for the instrumental variables I just mean whether I can see something like estimation output, but including the coeffs. for instrumental variables.

Many thanks.

[EViews Gareth]

Yes, my first response was "no", but then I realised my mistake and changed it to a "yes".

You can see that EViews is transforming them correctly with the following example:

Code: Select all

create a 1990 2000 10 series y=nrnd series x=nrnd series dx=d(x) equation e1.GMM(CX=FD,COV=PERWHITE,GMM=PERWHITE,ITER=ONEB) Y Y(-1) @lev(DX) C @ @DYN(Y,-2) equation e2.gmm(CX=FD,COV=PERWHITE,GMM=PERWHITE,ITER=ONEB) Y Y(-1) x C @ @DYN(Y,-2)

You'll note that the two equations give identical answers, the first using @lev of the differences Xs, and the second using the raw Xs.


I still don't understand what you mean by "estimated coefficients for the instruments". What coefficients?

[hannibal]

Hi,

I have panel data and would like to use the method by Arrelano and Bond (1991. In Eviews versions 5.1 and 6.0 there is a tool called wizard which is supposed to enable me to do so. Unfortunately, I did not quite understand how it works and whether the requirements of the estimation strategy are fulfilled.
According to Arellano and Bond, the growth rate of income can be regressed on its own lags and the growth rates of other variables.
My question is: If I insert the levels of the variables into the Wizard, will the first differences appear automatically or should I directly fill them in? Another question is where to introduce the instruments and how to test for the endogeneity problem.

Thank you for your help!

[EViews Gareth]

The help system / User Guide II has a walk through of Arrelano Bond estimation (replicating the original AB (1991) paper, I believe) in the panel estimation examples section.

[nadja123]

Hi hannibal,

I think QMS Gareth refers to the GMM example Layard Nickel / Arrelano Bond, based on the workfile abond_pan.wf1 you should be able to find in your EViews example files under Working with panel data. The trick is to realize (which took me quite a while, to big astonishment of QMS Gareth) that ideally you need to match the number of instruments with the number of regressors.

So, first decide on your regressors. It is expected that one of the regressors are lagged values of your dependent variable (one or more periods) (STEP 1). In the EV example, the first sort of regressors are 2 period lags of N. With regressors, you 've got 2 options: either let them transform or plug them in as levels, by using the tag @lev(). In the wizard, STEP 2 and 3. The disadvantage with using the wizard is that it transforms all regressors either by first differences or orthogonalization. If you want, write regressors in levels directly in the Equation dialog. You may see the proof of QMS Gareth that levels are supported. Now, the remaining issue is to take care of the instruments so that you have (at least) as many as regressors, STEP 4 and 5 in the wizard. The wizard supports entering instruments in levels.

You understand the EV GMM example as soon you are you able to anwer these 2 questions: What are the regressors? What are the instruments? The regressors are: lagged values n(-1) n(-2) and control variables w w(-1) k ys ys(-1), trasformed by 1st dif. The instruments are @dyn(N,-2) and the set of the control variables w w(-1) k ys ys(-1), again tranformed by 1st dif. So in fact the same. From econometric textbooks, I originally thought that I need other variables than regressors as instruments. But in fact, if you read Woolridge, it may be just one or two variables of interest that need special instruments, the remaining regressors are instrumented by themselves.

I hope this helps a bit, but it'd be good if some EV expert has a look on this post, I hope I was correct.

[nadja123]

Sorry, a typo:
@dyn(N,-2) takes lagged values as from the 2nd period backwards I think. So it is not the same as taking lags of two periods, which I think you would write @dyn(N,-1,-2). In the EV GMM example taking lags <= -2 could be afforded as there are just 9 periods and 140 cross-sections.

As for endogeneity, try to read on the Hausman test in Help (I have not used it myself).

[hannibal]

hi nadja 123,

thanks for your help. I have trouble to interprate the p-value from the J-statistics. Do you have an idee.

[nadja123]

Hi hannibal,
have a look on the help notes under the GMM example. The J-statistic is in fact the Sargan statistics, but I'd need to have a look in some textbook myself.

[hannibal]

thanks