Best way to do quantile regression for instrumental variable model with fixed effect

[Jeffery Xu]

Dear Statalist,

As a robustness check for my main model, I attempt to run a set of quantile regressions for an instrumental variable model with the fixed effect. The instrumental variable model includes both birth year and birth state fixed effect. The standard errors are clustered at the birth state level.

In my main analysis, the model is defined by the following code:

ivreg2 cognition (infl_pneumonia_rate1000=inter) female agesurvey black postsulfa `postAtheen' i.birth_year i.birth_state if  eduqt == 5, cluster(birth_state) partial(i.birth_year i.birth_state)

Currently, I am trying to use IVQREG2 to do the IV quantile regression with the fixed effect. Clustered SE was done using bootstrap. The code I used to estimate the IV quantile regression with FE is:

bootstrap, cluster(birth_state) idcluster(birth_state2) seed(22): ivqreg2 cognition infl_pneumonia_rate1000 female agesurvey black postsulfa `postAtheen' by* bs*, inst(inter female agesurvey black postsulfa `postAtheen' by* bs*) q(0.1(0.1).9)

The problem here is that running IVQREG2 under bootstrap is very time-consuming. For example, my STATA has run two hours to execute the code above but is still processing. Are there any better ways to estimate an IV quantile regression with FE but are more time-efficient?

Thanks in advance!

[Joao Santos Silva]

Dear Jeffery Xu,

The estimator implemented in ivqreg2 is not valid if the model has fixed effects. Actually, as far as I know there is no estimator that is valid in those conditions and you are better of using the correlated random effects approach; the estimator has recently been discussed in this forum.

Best wishes,

Joao

[Jeffery Xu]

Dear Jeffery Xu,

The estimator implemented in ivqreg2 is not valid if the model has fixed effects. Actually, as far as I know there is no estimator that is valid in those conditions and you are better of using the correlated random effects approach; the estimator has recently been discussed in this forum.

Best wishes,

Joao

Hi Joao,

Thanks for your response, and I really appreciate it! Even though I googled the correlated random effects approach and searched within the forum, I still have no idea about how to handle this problem using the CRE approach.

Could you give me some specific suggestions on using the correlated random effects approach to address this problem? For example, do you mean I can integrate the correlated random effects approach within the IVQREG2 estimator (e.g., include a vector of time averages as additional covariates)? Or should I estimate the first stage and second stage of the 2SLS separately to estimate quantile regression and include CRE in all stages of 2SLS?

Thanks again for your help!

Best,
Jeffery

[Joao Santos Silva]

Dear Jeffery Xu,

Please see details here

https://www.statalist.org/forums/for...se-model-tobit

Best wishes,

Joao