Different estimates based only on the order of the instrumen

[EconGrp]

Hi.

I have this weird problem:
I am estimating a panel tsls.
The estimation of the model is done encountering no errors or anything of the sort. But when I try to place the instruments in a different order, something really weird happens: my estimates change!

That is,

the case where the instruments are placed in this order
Instr(1) Instr(2) Instr(3)

and the case where they are placed like this (for example)
Instr(1) Instr(3) Instr(2)

, differ in terms of the estimates.

Does anyone have a clue about what is going on?

[EViews Gareth]

Could you let us know your EViews version number, your build date (available from Help->About EViews), and, if possible, post your workfile.

[EconGrp]

Eviews 7.1
Sep 10 2010 build

Can't post the workfile atm... but there is some missing data for some of the cross sections, for some of the periods - could that be the problem?

[EViews Gareth]

Please update to the latest version of EViews 7. You can download the latest update patch from the help menu, or from here:
http://www.eviews.com/downloads/

[EconGrp]

Ok, thanks - I'll do that.
But do you think that will solve the problem?

[EViews Gareth]

Hopefully.

[EconGrp]

Doesn't work :/

[EViews Gareth]

How many instruments did you add, and what is the reported instrument rank?

[EconGrp]

I have 7 explanatory variables and a constant. Only one of the explanatory variables is endogenous, and I only got 1 instrument for this. The remaining exogenous explanatory variables I use as instruments for themselves.
I have an instrument rank on 209.
I have 217 individuals in the cross section.

[EViews Gareth]

When you get the different results, do you have the same number of observations reported?

[EconGrp]

Yes - I do.

[startz]

I have 7 explanatory variables and a constant. Only one of the explanatory variables is endogenous, and I only got 1 instrument for this. The remaining exogenous explanatory variables I use as instruments for themselves.
I have an instrument rank on 209.
I have 217 individuals in the cross section.

I'm not sure I'm reading this right, but do you have 209 instruments with 217 observations???? Otherwise something is very screwy, because the instrument rank can't be higher than the number of instruments.

Even if you can't post your workfile, you might want to post your output.

[EViews Glenn]

It's a panel. I assume he's doing fixed effects and we're counting the implicit dummy variables as instruments.Though it sounds as though he has fewer than 217 in his final estimation sample.

[EconGrp]

I have tested the problem using different data - and even though the difference in estimates isn't big, the problem seems to endure over different workfiles.
So I ignore it for now.
I'm now facing a different problem: How do Eviews manage unbalanced panels? Does the program automatically change the estimators, so they'll fit the partly missing data? And is there an easy way to exclude cross sections with missing data, so I only use the balanced panel?

Thanks in advance

[EViews Gareth]

I cannot replicate what you're seeing at all. The following program creates random data on an unbalanced panel, and always gets identical results for the three equations (even though the order of instruments is different):

Code: Select all

create a 1990 2020 20 series y=nrnd series x1=nrnd series x2=nrnd series z1=nrnd series z2=nrnd series z3=nrnd series z4=nrnd 'make unbalanced smpl if nrnd>1.1 y = na pagecontract if y<>na smpl @all equation eq1.tsls(cx=f) y c x1 x2 @ c z1 z2 z3 z4 equation eq2.tsls(cx=f) y c x1 x2 @ c z2 z4 z1 z3 equation eq3.tsls(cx=f) y c x1 x2 @ c z3 z1 z4 z2