Probit - Please please please help!

[kian]

I have been running a probit model as part of a HECKMAN TWO STEP method on Eviews. I'm using a panel dataset. When I try and estimate the probit model over all the observations, I get a sign come up saying: -

DEPENDENT VARIABLE HAS NO VARIANCE.

But my dependent variable (a binary choice) clearly does have variance in that there are a large amount of zeros and a large amount of ones.
Why might this sign come up?
What does it mean ordinarily??? Please help, I don't know what else to do I'm so stuck!
Thank you!

[startz]

I have been running a probit model as part of a HECKMAN TWO STEP method on Eviews. I'm using a panel dataset. When I try and estimate the probit model over all the observations, I get a sign come up saying: -

DEPENDENT VARIABLE HAS NO VARIANCE.

But my dependent variable (a binary choice) clearly does have variance in that there are a large amount of zeros and a large amount of ones.
Why might this sign come up?
What does it mean ordinarily??? Please help, I don't know what else to do I'm so stuck!
Thank you!

You should check whether the smpl ended up excluding some values. This can happen "behind your back" if some of the other variables in the probit have NAs.

[kian]

Thanks for the quick reply Startz,

Given that it has removed some observations due to NA, how do I check what observations remain? (i.e. how do I look at the dependent variable series that is left over in order to check that this is the problem I'm experiencing?)

In order for this warning sign to appear, does it require that there is actually no variance in the dependent variable (i.e. that there are only ones in my dependent variable series, or does it only require that there is a negligible amount of zeros?)

I'd appreciate another speedy reply so very much...

[EViews Gareth]

One way to do it is to click on Proc->Make Regressor Group from your equation, then name the group that appears (by clicking on the Name button), then type this into the command pane:

series test = @robs(group01)=@max(@robs(group01))
show test

where Group01 is the name of the group you created. The resultant series will tell you (with a 1 or a 0) whether an observation would be included or not.

[kian]

Thanks Gareth,
That sounds useful but how can I do this when Eviews won't let me run the equation in the first place?

[EViews Gareth]

Estimate it by Least Squares first (just to get a working equation).

[kian]

I workerd through your advice on Eviews and found that for some of my datasets there were indeed no zero values (of my binary dependent variable) making it through to the estimated equation.

HOWEVER... for some I did find a proportion of zero values making it through. For instance one dataset returned 189 zeros and 1277 ones, but Eviews will not run the Probit regression, saying "dependent variable has no variance"...

Is is possible for Eviews to return the message "dependent variable has no variance" when there are a non zero amount of zeros (as in the case above)?
I.E. Will Eviews refuse to run a probit (even when there is some degree of variance) if the amount of zeros is a low proportion of the total observations included?

Please help, you've been so much use so far guys...

[EViews Gareth]

It shouldn't.

If you post your workfile, we'll take a look at what's going on.

[kian]

Never mind a little more number jumbling and I've realised that only the ones with no zero values can't be estimated.
Thanks soooo much for your help guys, now I've just got to work out why no zero values make it through!

P.S. You can't control for random or fixed effects in limitied dependent variable models (i.e. Probit) in Eviews can you?

[EViews Gareth]

Unfortunately, not.

Although you might be able to add dummy variables to your equation for each cross-section which would act as fixed effects.

In a pool setup this could be done simply by adding "C" as a cross-section specific regressor.

In a panel setup this could be done by adding @expand(@crossid, @dropfirst) as a regressor.

[EViews Glenn]

Caveats about the asymptotics required for these latter results should be kept in mind. You need the number of observations for each cross-section unit to go to infinity.

[kian]

And if, say, I had large value of observations (say 200) and a small value of time periods (say15)...?

[EViews Gareth]

You mean 200 cross-sections, and 15 time periods? That should be ok...

[startz]

And if, say, I had large value of observations (say 200) and a small value of time periods (say15)...?

This might be a little dicey, for the same reasons that a probit on a single sample of 15 observations might not be great. Unfortunately, this is an econometric issue more than an EViews issue.

[startz]

You mean 200 cross-sections, and 15 time periods? That should be ok...

Maybe, the issue is that there are only 15 observations to identify each time cross-section constant and that's not a lot in a nonlinear estimator.