GHASTLY TSLS

[cocoabutter]

Hi again, I've got a couple more questions please

1) So I have a panel ( which I am running as a pooled OLS generally, since i do not apply any cross section/time effects as explained by QM), and I'm trying to do a TSLS on it.
How do i also include control variables in the equation I want to estimate without running into the problem of overidentification? I only have three instruments but a whole load of other controls

eg I want to estimate

education outcome on childlabour(instrumented) and a host of other control variables

I've got pretty ghastly results so far, R squared is negative! I didn't even know that was possible! Is it because it is still a PANEL TSLS? I am getting rather panicked looking at these results.

2) Also, is there a Sargan test? the only one I can find in HELP is under dynamic panels..

Many thanks!

[startz]

Hi again, I've got a couple more questions please

1) So I have a panel ( which I am running as a pooled OLS generally, since i do not apply any cross section/time effects as explained by QM), and I'm trying to do a TSLS on it.
How do i also include control variables in the equation I want to estimate without running into the problem of overidentification? I only have three instruments but a whole load of other controls

eg I want to estimate

education outcome on childlabour(instrumented) and a host of other control variables

I've got pretty ghastly results so far, R squared is negative! I didn't even know that was possible! Is it because it is still a PANEL TSLS? I am getting rather panicked looking at these results.

2) Also, is there a Sargan test? the only one I can find in HELP is under dynamic panels..

Many thanks!

Overidentification is not a problem. It's kind of why TSLS was invented. A negative R-square is not unusual in TSLS-it has nothing particular to do with panels.

[cocoabutter]

well it IS a problem, because EVIEWs doesn't allow me to run it with my controls...

[EViews Gareth]

You'll need to be more specific.

[cocoabutter]

Hello Gareth,

I am working on a panel of 4 years, however not applying an cross section/period effects. Basically, I'm just running a pooled OLS ( as in previous question)
I need to do a TSLS in regressing
grade for age = c child labour Z Y
because child labour is endogenous.

Where Z is host of controls for individual level and Y is controls for household level.
I have 3 instruments on which I have already done a first stage regression just to check the correlation between variables. ( Also looking for the Sargan test here,does Eviews provide? )
But when I try to estimate the TSLS, I can't actually put down my controls, because Eviews says " overidentified"
Meaning that I can only have as many regressors as instruments (3, which is not enough). How do I get around this?

As mentioned when I run the TSLS on just child labour, I get a POSITIVE correlation ( wrong) and a negative R squared. According to Startz, the latter doesn't matter?

Thanks very much.

[startz]

You might want to post your workfile, including the equation you're trying to estimate.

[cocoabutter]

Hi startz,

I just tried but my workfile is way too big (1.5 MB) , is there somewhere else I can send it to instead?

I really appreciate it..

[EViews Gareth]

I just increased the maximum file upload limit to 2mb, so try attaching it again.

[cocoabutter]

I want to run TSLS

SAGE on c totaltime age age^2 female eduh exphh numberhh d2 d3 d4 d5 d6

where totaltime is to the endogenous variable. and then totaltime to be replaced with sumchores and sumfarm

instruments are crop crop*age crop*female crop*age*female

I also want to test if my instruments satisfy the usual criteria

Any help will be appreciated. Thank you very much

[EViews Gareth]

Not quite sure what you're saying makes sense, but it seems to me that you want to run the following equation:

Code: Select all

equation eq1.tsls sage c totaltime age age^2 female eduh exphh numberhh d2 d3 d4 d5 d6 @ crop crop*age crop*female crop*age*female age age^2 female eduh exphh numberhh d2 d3 d4 d5 d6

Which appears to work without any problems at all....

[cocoabutter]

Thanks for the quick reply Gareth, but I don't want all the other control variables to be used as instruments...only the ones I listed.
How do I go about doing this?

Or does Eviews just automatically add all the right hand regressors as instruments? And if so, how come?

[EViews Gareth]

I added those as instruments, since you appeared to say you wanted that.


I think you need to be a bit more clear as to what you're trying to do.

If you wish to regress SAGE on c totaltime age age^2 female eduh exphh numberhh d2 d3 d4 d5 d6, with only crop crop*age crop*female crop*age*female as instruments, then you're in trouble since you don't have enough instruments.

But you said you only wanted totaltime to be endogenous. Thus you can add the other regressors as instruments (which would make those regressors exogenous).

[startz]

Thanks for the quick reply Gareth, but I don't want all the other control variables to be used as instruments...only the ones I listed.
How do I go about doing this?

Or does Eviews just automatically add all the right hand regressors as instruments? And if so, how come?

EViews does TSLS, which means that whatever instruments you list are used for all the RHS variables. In principle, you can run the first stage for totaltime yourself and use the fitted values in the second stage. But while this gets the right coefficients, it gets the wrong standard errors.

Out of curiousity, why not do TSLS and use all the instruments?

[cocoabutter]

ah yes. OK noted. I didn't know that you could just add on the other regressors as instruments as well. Thanks for clearing that up!

1)So do I need to worry about a negative R squared or am I right in assuming is no longer holds its typical (OLS) meaning?
2) What is the equivalent of the sargan test in eviews?

[startz]

ah yes. OK noted. I didn't know that you could just add on the other regressors as instruments as well. Thanks for clearing that up!

1)So do I need to worry about a negative R squared or am I right in assuming is no longer holds its typical (OLS) meaning?
2) What is the equivalent of the sargan test in eviews?

1) R squared has the same meaning, but isn't restricted to be positive. Think of it this way, if you were estimating both supply and demand by TSLS, at least one of them probably wouldn't go through the data points very well.
2) I'm pretty sure that the test isn't built in.