Unconditional quantile regression (rifreg) and svy

[JoCT]

Hi everyone

I am trying to run an unconditional quantile regression using the user-written 'rifreg' command by Firpo, Fortin and Lemieux (2009) (available for download from the link attached). It is not compatible with svy, and while weights are allowed how do I account for the complex survey design (clustering and strata) when running the regression?

Thanks in advance

Jo

PS I am a novice STATA user so I hope this isn't a stupid question

[FernandoRios]

Hi Jo,
I Dont think there is a straight forward way to solve this. I think, however, you should consider using bootstraps for survey design.
HTH

[JoCT]

Thank you. I will look into the bootstrapping.
There are a number of papers using rifreg and complex survey data, but unfortunately they never discuss how they accounted for the survey design.

[JoCT]

For those who are interested the best solution is to calculate the RIFs yourself for each quantile of interest, and then run separate regressions for each using 'reg' and then you can use 'svy'

[Daniela Vaz]

For those who are interested the best solution is to calculate the RIFs yourself for each quantile of interest, and then run separate regressions for each using 'reg' and then you can use 'svy'

Hi, JoCT,

Did you suceed in doing so?

Could you send the code you used?

Thanks in advance!

[FernandoRios]

Hi Daniela
WHat he suggest is doing the following:
1. Create RIF of interest using RIFREG
2. Use OLS to estimate the RIF regression with the SVY option

svyset psuid [pweight=finalwgt], strata(stratid)
rifreg  zinc [aweight=finalwgt], retain(rif_10) q(10)
svy: reg rif_10 age age2 weight female black orace rural

I may add, however, that this doesnt estimate the "corrected" standard errors for unconditional quantile regressions. The reason being that for you to estimate that, you also estimate two additional parameters. The density at the qth point of the distribution, and the qth quantile. This is why its recommended to use bootstrap procedure.

HTH
Fernando

[Jana Belzer]

Dear Fernando,

thank you very much for sharing your expertise with this community!
I am a STATA beginner and maybe you could support me in the following:

I have survey data, that uses syv option for extrapolating survey data to overall population representation.
I try to reproduce an oaxaca_rif step for step. First creating a rif_variable for income at the 15th percentile (for two points in time with the by(time) option). Here it is not possible to use the survey option, instead I used weights. Is this a problem for the RIF variable?

Than I used oaxaca with the created RIF variable as dependent variable and my list of independent variables with the option by(time). I have read (also here in the forum) that the survey option does not correct standard errors for unconditional quantile regression, instead bootstrap is supposed to be used. However, when only using bootstrap, my coefficients and also the income at the 15th percentile for the two years are incorrect in the oaxaca output, as it has not been accounted for the survey design or weights. I also implemented bootstrap, stata(strpsu) but this didn't improve the results.
In the end it seems only one of the options can be used svy or bootstrap. Unfortunately I don't have the expertise to know how I should best continue to receive correct coefficients but also corrected standard errors. Especially because oaxaca_rif also only works with bootstrap or svy or weights.

Could you help me in this matter?

Thank you very much in advance!

Jana