Logit/Probit Estimation and Panel Data

[michelleli]

Hello everybody,

I am using EViews 7 and my workfile comprises binary panel data. I know, that Logit/Probit models also work on Panel data, but I have some questions about how I can account for fixed effects in the panel data when estimating a Logit/Probit model.

From previous posts, I understand the only way to do it is by creating dummy variables for the fixed effects. I have created firm dummies for the fixed effects. However, when I include these dummies (I have 152 firm dummie variables) in the estimation, Eviews show that "Parameter estimates for limit points are non-ascending 0". How can correct for this? Also is this precedure the same as estimating standard errors with firm-level clustering?

Kind Regards
Michelle

[EViews Glenn]

That error message sounds like it's masking a different error. Are you including both an intercept C and the dummy variables? If so, you're falling into the standard dummy variable trap. In that case, removing the C (or one of your firm dummies should handle the problem).

You can post the workfile here and we can take a look for sure. If it's a bad error message, I'll see about getting it corrected.

Note that adding the dummies is not the same as estimating with clustering. The former estimates with firm-specific means but does not do anything about the covariances between errors. We currently don't have probit/logit with firm-level clustering, but it is possible with some work to roll your own. But again, it's not built-in.

[michelleli]

Given the answer you provided below about firm level clustering, what can I do to deal with autocorrelation in the dependent variable within firm clustering and possibly autocorrelation in the residuals.

Cheers

That error message sounds like it's masking a different error. Are you including both an intercept C and the dummy variables? If so, you're falling into the standard dummy variable trap. In that case, removing the C (or one of your firm dummies should handle the problem).

You can post the workfile here and we can take a look for sure. If it's a bad error message, I'll see about getting it corrected.

Note that adding the dummies is not the same as estimating with clustering. The former estimates with firm-specific means but does not do anything about the covariances between errors. We currently don't have probit/logit with firm-level clustering, but it is possible with some work to roll your own. But again, it's not built-in.

[EViews Glenn]

http://forums.eviews.com/viewtopic.php?f=23&t=3029 does this for non-panels.

In the panel setting, you should look at the literature on Generalized Estimating Equations.

[michelleli]

Hi Glenn,
I don't think it is the problem of dummy variable trap. I have uploaded my workfile, please have a look.

I am trying to regress ordered on lrsale_w segbus segloca agelist agelist^2 qratio_w leverage block_shareholding_w board ceoyears ceo_ss ceogender founderinvolvement _indpt chdirlist chdiridpt_w dirshrexceobene_w ind2 ind3 ind5 ind6 ind8 ind9 ind10 ind11 ind12 ind13 ind15 ind16 yr98 yr99 yr00 yr01 yr02 yr03 yr04 yr05 yr06 yr07 yr08 and firm dummy variables firm1-firm 152 with ordered probit model.

Cheers
Michelle

That error message sounds like it's masking a different error. Are you including both an intercept C and the dummy variables? If so, you're falling into the standard dummy variable trap. In that case, removing the C (or one of your firm dummies should handle the problem).

You can post the workfile here and we can take a look for sure. If it's a bad error message, I'll see about getting it corrected.

Note that adding the dummies is not the same as estimating with clustering. The former estimates with firm-specific means but does not do anything about the covariances between errors. We currently don't have probit/logit with firm-level clustering, but it is possible with some work to roll your own. But again, it's not built-in.

[EViews Glenn]

Unless your firms change industries during the course of your panels, you can't have both firm dummies and industry dummies. You'll have to remove the industry dummies to estimate with firm dummies. And beyond that, there appear to be other numerical issues.

I think the error message is correct. We're unable to get starting values for the limit points that work. But I think this is because the regressor matrix is singular in general. Try running least squares on the same specification with only firm and time dummies and you'll see that it's singular.

[michelleli]

Hi Glenn,
Thank you very much for your help.
One more question with regard to the firm level clustering standard error. Since it is not built in for ordered probit model with panel data, what's the steps I need to take to estimate robust standradr erros incorporating firm level clustering?

Regards
Michelle

That error message sounds like it's masking a different error. Are you including both an intercept C and the dummy variables? If so, you're falling into the standard dummy variable trap. In that case, removing the C (or one of your firm dummies should handle the problem).

You can post the workfile here and we can take a look for sure. If it's a bad error message, I'll see about getting it corrected.

Note that adding the dummies is not the same as estimating with clustering. The former estimates with firm-specific means but does not do anything about the covariances between errors. We currently don't have probit/logit with firm-level clustering, but it is possible with some work to roll your own. But again, it's not built-in.

[EViews Glenn]

It's not a trivial calculation, but can be done. It's going to involve some programming.

http://pages.stern.nyu.edu/~wgreene/Dis ... Survey.pdf

has a discussion. Basically, you'd have to use the estimates that EViews provides to obtain H^-1, and then compute the middle term using the saved gradients.

[rjoreilly]

Hi Glenn,

I am undertaking something very similar and I was hoping you could elaborate on the process involved in your last post on this thread?

Thanks