Dear all,

I'm actually performing an Oaxaca-Blinder decomposition to test wether the returns of individual observable characteristics on log-income differ by origin class i.e. working class vs service class.
For the standard OLS i.e. the decomposition at the mean, I use the following code, run without any problem:

Code:
 oaxaca log_income normalize(edur_*) normalize(emps_*) normalize(sex_1 sex_2) ability_scaled jbhrs labour_experience [pw=c_indscub_xw], by(service_origin) relax swap pooled
In this case, the coefficient/unexplained effect is significant and in the direction I was expecting.


However, I'm also interested in verifying wether the unobserved effects changes along the income distribution. Therefore, following the contributions by Firpo et al. (2009, 2018) I'm using the oaxaca_rif command introduced by Rios-Avila (2019). In this case, it is possible to perform the decomposition at different percentiles and the authors suggest to implement a reweight approach in this kind of decomposition. However, I have some doubts.
Indeed, when I run the oaxaca_rif as follows:
Code:
 xi : oaxaca_rif log_income normalize(edur_*) normalize(emps_*) normalize(sex_1 sex_2) ability_scaled jbhrs labour_experience [pw=c_indscub_xw], ///
by(service_origin) rif(mean) swap rwlogit(i.edu_reduced i.emp_status ability_scaled jbhrs labour_experience i.gender)  relax
Using the mean as a comparison with the OLS, I observe very close estimates for the decomposition, however using the reweighting approach the unexplained effect is no more significant, while it is if I do not use reweighting approach. I was wondering, why is this the case? what is the best approach in this "contradicting" case?

Thanks for your support.