Forum: help

RE: Functionality beyond tobit-II? [ Reply ] By: Arne Henningsen on 2016-07-04 06:04	[forum:43336]
I guess that you mean by 'exclusion restrictions' that the dependent variables have different sets of explanatory variables, right? In a Seemingly Unrelated Regression (SUR) model, you can have different explanatory variables for the different dependent variables. Hence, 'exclusion restrictions' do not rule out SUR models.

RE: Functionality beyond tobit-II? [ Reply ] By: B R on 2016-06-13 08:03	[forum:43277]
Hi Arne, thanks so much for your hint -- I have not been aware of this paper and will certainly have a look. I certainly can rule out SURE models, because I do have an exclusion restriction. All best, Bernhard

RE: Functionality beyond tobit-II? [ Reply ] By: Arne Henningsen on 2016-06-13 07:05	[forum:43276]
You may take a look at the following paper (although your model is simpler, because AFAIK it does not have endogenous explanatory variables, i.e. your model is not 'simultaneous'): DEVELOPING CONSISTENT ESTIMATES OF MARGINAL EFFECTS IN A SIMULTANEOUS EQUATION MODEL WITH LIMITED DEPENDENT VARIABLES Atwood, Joseph & Joglekar, Alison http://purl.umn.edu/235554

RE: Functionality beyond tobit-II? [ Reply ]
By: Arne Henningsen on 2016-06-12 05:58

[forum:43271]

Dear Bernhard

Do I correctly understand that you want to estimate both the first equation and the second equation with all observations in your sample? If this is the case, there is no 'sample selection' and thus, you do not need to correct for this. You can consistently (but not efficiently) estimate your model by two separate (independent) estimations (e.g. one probit estimation and one OLS estimation).

Given that the error terms of the two equations are potentially correlated, it seems to me that you have a system of equations that could be efficiently estimated, e.g., by 'seemingly unrelated regression' (SUR). If you estimate the first equation as 'linear probability model,' you can use the 'systemfit' package for estimating your model.

I am not sure if there exists a ready-to-use software for simultaneously estimating a probit equation and an (Gaussian) linear equation, but it should be not too difficult to derive the (log) likelihood function for this model based on the bivariate normal distribution and to maximize it using the 'maxLik' package.

Your model specification suggests that the coefficients 'b' are the same in the first and in the second equation. If this is the case, you should impose these cross-equation parameter restrictions, because this gives more efficient estimates.

Best regards,
Arne

Functionality beyond tobit-II? [ Reply ]
By: B R on 2016-06-11 20:52

[forum:43270]

Dear users,

I want to estimate a sampleSelection model in which the second stage is not confined to positive cases.

y1={0, if y1*<0
1, if y1*>=0

y1*=z'j +x'b + u1

y2=x'b+u2

The idea is that Cor(u1,u2)<>0, so I need the sample selection, but again, do not want E(y2|x) not E(y2|x, y1==1), which is implemented in sampleSelection (with either method).

Alternatively, is there a routine to estimate correct standard errors (as proposed by Heckman) for my problem?

Thanks for any help.

Bernhard