URGENT: Dynamic panel regression GMM (EViews 6)

[nadja123]

Hi there,

in a panel workfile with 10 cross-sections x 20 periods I need to estimate something like this:
:equation: d(pbal) = b0 + b1 * d(pbal(-1)) + b2 * d(gdp) + b3 * prenorm + b4 * elenorm + error,
where prenorm and elenorm are dummies. Instead of error you can imagine d(error) too. By d(.) I denote first differences.

So far I obtained LS estimates of the equation. However, I saw in some papers they use GMM (Arellano Bond) estimation and I need to use it too. I tried Quick / Estimate Equation / GMM - DPD regression, in the Dynamic Panel Wizard I specified:
Step 1: dependent variable: pbal, lags from one up to: 1
Step 2: other regressors: (blank), include period dummies (blank)
Step 3: transformation: differences
Step 4: @dyn(PBAL,-2) (leave unchanged)
Step 5: other instruments: gdp (differences), prenorm elenorm (no transformation)
Step 6: leave the options unchanged

After I click OK it reports error messages, mostly "the number of instruments is greater than number of observations", after other trials "insufficient number of instruments".

I checked out the EViews GMM example (abond_pan.wf1), it worked just fine. Of course the number of observations there is much higher than mine, 1031 vs just <200 (some are missing). Note that I need to run the regression even for a smaller number of observations.

I attach my workfile. Please can somebody help??? Thanks !!!

[EViews Gareth]

Not really clear what your question is here.

[nadja123]

Hi QMS Gareth,

sorry for unclarity. I'll try to put it more clearly. In fact, this problem can be turned into several question:

1. Is GMM/ Dynamic Panel Data option in EV suitable to estimate an equation as this one? Does the estimate give different results than if running LS?
:equation: d(pbal) = b0 + b1 * d(pbal(-1)) + b2 * d(gdp) + b3 * prenorm + b4 * elenorm + error,
where d(.) denotes first differences, prenorm and elenorm are dummies and instead of error you can imagine d(error) too.

2. I have a panel data set with 10 cross-sections x 20 periods and estimated the equation by LS in EV, no problem. The workfile is attached together with this post. However, I encountered problems to run an equivalent regression using the GMM / DPD option as in fact, GMM estimates is what I need to obtain. So, I am trying to figure out what is the reason of the error messages?
This is what I did, using the Dynamic Panel Wizard:
Step 1: dependent variable: pbal, lags from one up to: 1
Step 2: other regressors: (blank), include period dummies (blank)
Step 3: transformation: differences
Step 4: @dyn(PBAL,-2) (leave unchanged)
Step 5: other instruments: gdp (differences), prenorm elenorm (no transformation)
Step 6: leave the options unchanged
... and the error messages I obtained: "the number of instruments is greater than number of observations", or after other trials "insufficient number of instruments"

3. The possible reasons I imagine: Have I filled out the DPD dialog wrongly (I mean have I translated the equation wrongly into that dialog)? I am not sure anymore.

4. Or, is the reason perhaps that my sample size is insufficient? I checked out the EViews GMM example (abond_pan.wf1), it worked just fine. But the number of observations there is 1031 vs just <200 of mine (some are missing). I saw in the Wizard that the option is suitable for samples with a large number of cross-sections and short time series. Maybe my sample just doesn't fit into that.

Sorry, I just try to figure out how to run the GMM / DPD regression correctly and I just got confused by this... Thanks a lot for feedback !!!

[EViews Gareth]

I can't really answer the first question, I think you'll have to consult your econometric text books for that one.

As for the other questions, I think the error messages are somewhat self explanatory. In the base case it would appear that you have too many instruments - i.e. more instruments than observations. Obviously it is mathematically impossible to estimate such a model, so EViews errors. In your other cases you've gone the other way and have an insufficient number of instruments (you need more instruments than regressors).

Again I can't really tell you what the "correct" model should be. I don't think your problem lies with running DPD in EViews - you appear to have a pretty good understanding of how to fill in the boxes etc...

[nadja123]

Thanks QMS Gareth,

for a promt reply on Sunday... You found I have an understanding of how to fill in the Dynamic Panel Wizard dialog. Well, I have to assure you, not at all! I have now tried and indeed produced some output, however, it does not do what I expect it to do. Especially, I see no way to include the PRENORM and ELENORM dummies in levels so that they are recognized as regressors (and are not transformed at the same time).

I am afraid I have to stick to my "simple" LS estimates as the paper urges. If I find out more on this problem later, I will post that info to this forum. Thanks anyway for reply and have a nice Sunday !!! [Of course, working on Sundays is no too good...]