Brunello et al (2009) show that extended compulsory schooling leads to increased wages respectivly to the individual gender. Their empirical model first uses quantile regression to show the impact of compulsory schooling years (ycomp, defined as the instrument variable z) on actual years of education (s). Afterwards they subtract those fitted values of the regression from s to get the ability of a person of a specific quantile.[2]

They claim that their model is exactly identified to do so.[3]

In the end they come up with a quantile regression aproach which is augmented by the control variate computed in [2]. But if I understand them correctly they compute the inverse of the $\tau$ - quantiles of distribution $a$ and $u$. [4]

If I got that right, could somenone help me to show how this is done? I suspect some kind of Monte-Carlo Method, e.g. importance sampling, but I'm unsure. A solution with R-code is appreciated but not necessary.

EDIT : Simplfying the question : How does one calculate $G_{a}^{-1}\left(\tau_{a}\right)$ and $G_{u}^{-1}\left(\tau_{u}\right)$?

[2]:First, we estimate the conditional quantile functions of schooling $s$ and compute the control variate $$ a\left(\tau_{a}\right)=s-\bar{Q}\left( \tau_{a} \mid X, z \right) $$

[3]: Omitting subscripts for simplicity, the earnings-cum-education model presented above can be written in the format of an exactly identified triangular model, as in Chesher's approach $$ \begin{array}{c} \ln(w)=\beta s+s(\lambda a+\phi u)+\gamma_{w} X+a+u &(6)\\ s=\gamma_{s} X+\pi z+\xi a &(7)\end{array} $$

[4]: Define $\tau_{a}=G_{a}\left(a_{\tau_{a}}\right) \text { and } \tau_{u}=G_{u}\left(u_{\tau_{u}}\right)$, where $a_{\tau_{a}}$ and $u_{\tau_{u}}$ are the $\tau-$ quantiles of the distributions of $a$ and $u,$ respectively. Furthermore define $Q_{w}\left(\tau_{u} \mid s, X, z\right)$ and $Q_{s}\left(\tau_{a} \mid X, z\right)$ as the conditional quantile functions corresponding to log wages and years of education. Ma and Koenker (2006) show that recursive conditioning yields the following model $$ \begin{array}{c} Q_{w}\left[\tau_{u} \mid Q_s\left(\tau_{a} \mid X, z\right), X, z\right]=Q_s\left(\tau_{a} \mid X, z\right) \Pi\left(\tau_{a}, \tau_{u}\right)+\gamma_{w} X+G_{a}^{-1}\left(\tau_{a}\right)+G_{u}^{-1}\left(\tau_{u}\right)& (8) \\ Q_{s}\left(\tau_{a} \mid X, z\right)=\gamma_{s} X+\pi z+\xi G_{a}^{-1}\left(\tau_{a}\right) & (9)\end{array} $$ Given the restrictions imposed by (6) and $(7),$ the key parameter of interest $\Pi\left(\tau_{a} \tau_{u}\right)$ is a matrix with the following structure $$ \begin{array}{c} \Pi\left(\tau_{a}, \tau_{u}\right)=\beta+\lambda G_{a}^{-1}\left(\tau_{a}\right)+\phi G_{u}^{-1}\left(\tau_{u}\right) \quad \end{array} $$

EDIT : Added some clarification to the equation numbering, fixed the equation in the first qoute and added more information on variable ycomp/z. Fixed the main quantile regression equation (8) from subscript v to w (wages). Clarified the importance-sampling tag. Edited the subscripts of the distribution functions $\tau_a$ and $\tau_u$

  • $\begingroup$ Is the importance-sampling tag relevant here? $\endgroup$ Sep 21, 2020 at 9:26
  • $\begingroup$ That article takes some time to dig through, and the question is not easy to understand. Already the first line $\left(\tau_{a}\right)=s-\bar{Q}\left( \tau_{a} \mid X, z \right)$ is confusing to me without context. So the $\tau$ are quantiles but where does the subscript $a$ suddenly come from? Why is $\tau_a$ in both sides of the equation? What are the variables $X$ and $z$? .... Could you make the question more concise/simpler with a self-contained example? $\endgroup$ Sep 21, 2020 at 9:38
  • $\begingroup$ @RichardHardy I thought that for the inverse of the quantile distribution one need some kind of Monte-Carlo integration. If I'm wrong here, I will delete the tag. $\endgroup$
    – mugdi
    Sep 21, 2020 at 9:43
  • $\begingroup$ @SextusEmpiricus I'm sorry that I made a mistage in the first equation. Of course It should be $a(\tau_a)$ . Fixed that! $\endgroup$
    – mugdi
    Sep 21, 2020 at 9:44
  • 1
    $\begingroup$ For the sake of a question on cross validated, whenever you can distill the essence of the question from noisy/confusing field specific context, then it will always be better. I am not sure whether there is an economics stackechange, the question might do well there. Or otherwise the quants stackechange. $\endgroup$ Sep 22, 2020 at 9:33

1 Answer 1


I think I found an answear to this on my own. Following Ma and Koenker one simply does the following :

For tau = 0.1 :

$G_{a}^{-1}\left(\tau_{a}\right)$ =


$G_{u}^{-1}\left(\tau_{u}\right)$ =


Maybe someone can verify this theoretically?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.