IVQE command in Stata (unconditional quantile regression)

[John Mendes]

Dear Statalist members,

I am having a problem with the IVQE command.


The problem is the following. I am trying to obtain unconditional quantile regression results through IVQE (estimator provided by Firpo’s 2007 for unconditional quantile regression estimator).

I want to estimate the partial effect of a dummy variable (say, D1) on Y (the explanatory variable). The control variables are x (a continuous variable) and D2, D3 and D4 (two additional dummy variables).

The Stata command for the 0.1th unconditional quantile is:

IVQTE Y (D1), continuous(x) dummy(D2 D3 D4) quantiles(0.1) variance


After I run this line, I get a point estimate, standard error and confidence interval for the 0.1th quantile of D1.

However, when I add together D2 and D3 into a single dummy say, D23 ( D23 is equal to 1 when D2==1 or D3==1), and run the line:

IVQTE Y (D1), continuous(x) dummy(D23 D4) quantiles(0.1) variance

I obtain exactly the same coefficient estimate, standard error and confidence interval as when I run the first line.

Is this normal to happen? For the same data, when I use OLS, I get very similar results for the conditional mean point estimate, confidence intervals and standard deviation (for the conditional mean point estimate); however, they are not exactly the same.

Has anyone come across this issue or might have an explanation for this? Perhaps this has to do with the method? I am rather new in terms of using this option.

Thanks.

[FernandoRios]

Hi John
what you are finding its not a bug or a problem, but just something that happened by chance.
If you revise Firpo (2007) paper, or the paper that introduces the command ivqte (https://www.stata-journal.com/sjpdf....iclenum=st0203) you will see how the treatment effect its being estimated, and may bring you some closing of what is happening.
In a nutshell, when estimating exogenous treatment effects (as you are trying to do), the command will estimate the quantiles of Y for the sample when D1=0 and D1=1 using a reweigthing strategy based on the probability that an observation was assigned to the treatment group, based on some controls.
In your specification, the treatment D1 is a function of X and D2 D3 and D4. P(D1=1|D2,D3,D4)
ivqte estimates this conditional probability using a nonparametric method. At it may be that because its nonparametric, that combining D2 and D3 has no impact on the predicted conditional probability, which is why your estimated treatment effect does not change in your two specifications.
HTH
Fernando

[John Mendes]

Hi Fernando.

Thanks a lot for your answer! It helped me a lot.

Best regards.