Anderson-Rubin test in IVprobit using weakiv

[Sylvie Rod]

I am running an IVprobit with one endogenous regressor and only one instrument. Since I am calculating the marginal effects, I am not using the twostep option.

However I am aware that in order to obtain the test for weak instrument (Anderson-Rubin is the only one available for this type of model according to Finlay and Magnusson), I must use twostep.

So my questions are:

1. If the final specification that I am using is the IVprobit without the twostep option (cause I need to calculate the marginal effects), is it valid to use the results from the Anderson-Rubin weak IVprobit estimation using twostep to make conclusions about my instrument? I ask this because I noticed that the coefficients with and without the twostep option are not identical.

2. Is there any other option to test the strength of the instrument in an IVprobit for those cases when the twostep option is not used?

3. If the Anderson-Rubin test using twostep in IVprobit is significant (p-value<0.05), then I could straightforward conclude that my instrument is not weak? Or this is just to reject the null hyphotesis that the B of the endogenous regressor is not zero?

4. When using the weakIV with IVprobit twostep I also get the Wald test. Is the information provided for this test also helpful?

Thanks in advance.

[Irene Selwaness]

Having similar questions.
Any help here?
Thanks

[Jeff Wooldridge]

The weak instruments problem has to do with the reduced form for the endogenous explanatory variable. Call that variable y2. In the model underlying ivprobit, y2 follows a linear reduced form with additive, normal errors. Therefore, all diagnostics for weak instruments are done just as in the usual linear model case. I would estimate the reduced form for y2 directly and look at the F statistic, as in Staiger and Stock, and see if it is sufficiently large (at least 10). After you determine your IV or IVs are strong enough using the usual weak IV diagnostics, you may use the full MLE or the two-step estimator. That choice should have little to do with weak versus strong IVs.

[Irene Selwaness]

Dear Prof. Wooldridge,
Thank you for your prompt response.
If I understand correctly, you advise to run the usual weak IV diagnostics (i.e. after 2SLS estimation - ivregress for instance) to determine whether instruments are strong or not. Then, we can correct for endogeneity estimating an ivprobit or a biavariate probit.. etc.

Irene

[Jeff Wooldridge]

Dear Prof. Wooldridge,
Thank you for your prompt response.
If I understand correctly, you advise to run the usual weak IV diagnostics (i.e. after 2SLS estimation - ivregress for instance) to determine whether instruments are strong or not. Then, we can correct for endogeneity estimating an ivprobit or a biavariate probit.. etc.

Irene

Yes that will do it. You can use the

estat first

command after using -ivregress-.

[Irene Selwaness]

Yes that will do it. You can use the

estat first

command after using -ivregress-.

Dear Professor Wooldridge,
Thank you so much for your reply.
Do you think that we can do the endogeneity tests even if the endogenous regressor y2 is binary (and the main outcome y1 as well); or better to rely on weakiv ?

Best,

Irene

[Irene Selwaness]

Yes I did that.

Do you think that I can do the usual diagnostics for weak instruments in case of a binary outcome y1 and a binary endogenous regressor y2 is binary; or better to rely on weakiv ?

Also, to make sure, the Anderson-Rubin (in the weakiv command) tests the joint hypothesis that the instruments are (1) weak and (2) exogenous... Correct?

Best,

Irene

[Irene Selwaness]

Sorry I double posted this message by mistake! --

Yes that will do it. You can use the

estat first

command after using -ivregress-.

Yes I did that.

Do you think that I can do the usual diagnostics for weak instruments in case of a binary outcome y1 and a binary endogenous regressor y2 is binary; or better to rely on weakiv ?

Also, to make sure, the Anderson-Rubin (in the weakiv command) tests the joint hypothesis that the instruments are (1) weak and (2) exogenous... Correct?

Best,

Irene

[Dorgyles Kouakou]

The weak instruments problem has to do with the reduced form for the endogenous explanatory variable. Call that variable y2. In the model underlying ivprobit, y2 follows a linear reduced form with additive, normal errors. Therefore, all diagnostics for weak instruments are done just as in the usual linear model case. I would estimate the reduced form for y2 directly and look at the F statistic, as in Staiger and Stock, and see if it is sufficiently large (at least 10). After you determine your IV or IVs are strong enough using the usual weak IV diagnostics, you may use the full MLE or the two-step estimator. That choice should have little to do with weak versus strong IVs.

Dear Professor,

Can we do what you said when we have two endogenous explanatory variables ????,

Please advise!!!!

Regards.

[Sumaira Ashraf]

I am running an IVprobit with one endogenous regressor and only one instrument. Since I am calculating the marginal effects, I am not using the twostep option.

However I am aware that in order to obtain the test for weak instrument (Anderson-Rubin is the only one available for this type of model according to Finlay and Magnusson), I must use twostep.

So my questions are:

1. If the final specification that I am using is the IVprobit without the twostep option (cause I need to calculate the marginal effects), is it valid to use the results from the Anderson-Rubin weak IVprobit estimation using twostep to make conclusions about my instrument? I ask this because I noticed that the coefficients with and without the twostep option are not identical.

2. Is there any other option to test the strength of the instrument in an IVprobit for those cases when the twostep option is not used?

3. If the Anderson-Rubin test using twostep in IVprobit is significant (p-value<0.05), then I could straightforward conclude that my instrument is not weak? Or this is just to reject the null hyphotesis that the B of the endogenous regressor is not zero?

4. When using the weakIV with IVprobit twostep I also get the Wald test. Is the information provided for this test also helpful?

Thanks in advance.

Hi. Were you able to solve this problem?

I am also running an IVprobit with one endogenous regressor and only one instrument, and with a binary dependent variable.

The Wald test of exogeneity: chi2(1) = 0.02 Prob > chi2 = 0.8972

Does that mean that there is no endogeneity in my independent variables?

How did you check the strength of the instrumental variable? I am unable to perform the weakIV test in Stata 14.2 "command WeakIV is unrecognized". Were you able to do that?

Regards,
Sumaira

[Kazi Aiman Udoy]

Sylvie Rod Would you please share the code you have run for ivprobit and endogeneity test. I am actually running the same thing for my regression. But I am unable to run it appropriately. Your help would be a great assistance for me.

[Zeynep Yilmaz]

Can we do the Anderson-Rubin tests (weakiv) when we have two endogenous variables and two instruments?

Thank you.