Meaning of the Oaxaca_rif result?

[Jang JinHee]

Thank you for your answer.
The following results were obtained using oaxaca_rif.

---------------------------------------------------
overall
group_1 | .2900839
group_2 | .2913806
difference | -.0005414
explained | -.0574823
unexplained | .0569409
----------------------------------------------------

I understand what the following example means.
- for example difference 0.05, explained 0.02, unexplained 0.03 (unexplained ratio = 0.03/0.05*100)
But I do not know what it means to be a negative number.
How can I get the unexplained ratio? (0.569/-0.0005414 *100 ?)

I'm sorry for starters...
I hope to help you

[Liu Qiang]

Dear Jang JinHee, to be frank, I had the same puzzling question days ago and consulted one of the professional experts in this field.
The result is negative maybe partly due to the model specification error.
hope this thread could help you understand it
https://www.statalist.org/forums/for...-decomposition

[FernandoRios]

HI Jang,
In addition to Liu's advice i think you have to rephrase your problem/interpretation of the results.
You are correct in saying that often, for OB type of decomposition, people interpret the components as shares or proportion of the raw gap. This is not the only way, specially for a case like yours.
Think about the following example. You run a OB decomposition, using wages as depend variables, and education as the ONLY explanatory variable. YOu get the following results
Group 1. Wage=0+5*educ avg education =3
group 2 Wage=0+3*Educ avg education =5

If you compare their wages, you will see, like in your example, that the RAW gap is zero.This is hiding the fact that there is discrimination (or at least differences in wage structures). Group 2 returns to education is much lower than groups 1. Group 2 however has more education, which counteracts the coefficients impact.

In your case, the raw gap (mean gap?) but the decomposition actually reveals differences in characteristics and returns to those characteristics.
HTH
Fernando

[Liu Qiang]

HI Jang,
In addition to Liu's advice i think you have to rephrase your problem/interpretation of the results.
You are correct in saying that often, for OB type of decomposition, people interpret the components as shares or proportion of the raw gap. This is not the only way, specially for a case like yours.
Think about the following example. You run a OB decomposition, using wages as depend variables, and education as the ONLY explanatory variable. YOu get the following results
Group 1. Wage=0+5*educ avg education =3
group 2 Wage=0+3*Educ avg education =5

If you compare their wages, you will see, like in your example, that the RAW gap is zero.This is hiding the fact that there is discrimination (or at least differences in wage structures). Group 2 returns to education is much lower than groups 1. Group 2 however has more education, which counteracts the coefficients impact.

In your case, the raw gap (mean gap?) but the decomposition actually reveals differences in characteristics and returns to those characteristics.
HTH
Fernando

thank you for your nice answer. I once again learn a lot

clear all
set obs 20000
gen u=invnorm(uniform())
gen group=.
replace group=1 in 1/10000
replace group=2 in 10001/l
gen edu=.
replace edu=invnorm(uniform())+3 in 1/10000
replace edu=invnorm(uniform())+5 in 10001/l
gen wage=.
replace wage=5*edu+u in 1/10000
replace wage=3*edu+u in 10001/l
reg wage edu if group==1
reg wage edu if group==2
oaxaca wage edu, by(group)

This is a simulation based on Fernando's advice.
If I understand it correctly, there may still exist discrimination although the wage difference between two different groups is zero. In this way, you cannot interpret the results using the "share" perspective any longer.

[Jang JinHee]

FernandoRios and Liu Qiang, Thank you for your advice.
But I still did not understand it properly. T_T)
I want to find the share of discrimination in the total wage gap.
Studies show that the unexplained part is negative, or the total wage gap is negative.
So, should I consider other calculations regardless of the oaxaca_rif overall result?
Or should I do additional work after estimating oaxaca_rif?
I understand what you told me, but it's hard to apply to me.
I'm sorry I do not know.
Please let me know a little more detail.

[FernandoRios]

Hi jang
the bottom line is that using shares will give you weird results and I would personally not recommend it.
on the other hand you haven’t said anything about your data, outcome variable or what is what you are trying to analyze
also Liu that was a nice replication
Fernando

[Liu Qiang]

FernandoRios and Liu Qiang, Thank you for your advice.
But I still did not understand it properly. T_T)
I want to find the share of discrimination in the total wage gap.
Studies show that the unexplained part is negative, or the total wage gap is negative.
So, should I consider other calculations regardless of the oaxaca_rif overall result?
Or should I do additional work after estimating oaxaca_rif?
I understand what you told me, but it's hard to apply to me.
I'm sorry I do not know.
Please let me know a little more detail.

hi Jang, I may offer a new insight that may help you understand what I am saying. Let's say there are two different individuals A and B. A has a higher education level with 5 years whereas B only has 3 years. But what we can observe is that they have the same hourly wage，both 15 dollars/hour for example. One may argue that wage is not solely affected by education and you have to include other covariates, a model specification error. But if we assume the model is correctly set, we should have observed A's wage higher than B's but in fact we don't. This means there's still discrimination since A has endowment advantage that is not rewarded in his salary. But you cannot say how much of the difference is caused by discrimination because there's no total difference at all.
Hope I put it clearly
Best Regards

[Jang JinHee]

Thank you for your advice to me that I do not know.
I understand your comments.
Nevertheless, I will ask you one more question.

I did the following :

1) The wage gap decomposition of women with children and women without children.

2) use 'oaxaca_rif'
------------------------------------------------------------------
local i = 10
while `i' <= 90 {
oaxaca_rif lnhwage age age2 dum_pt2 dum_union2-dum_union3 dum_stat2 dum_fsize2-dum_fsize4 dum_ind2-dum_ind17 dum_edu2-dum_edu3 dum_job2-dum_job9 if gender==1, by(fm) wgt(1) rif(q(`i'))
local i = `i'+10
}
----------------------------------------------------------------------
fm : women with children = 1, women without children = 0

3) As a result, I obtained these results in several quantiles.
-----------------------------------------------------------------------
example 1)

RIF : q(70)
Group 1: fm = 0 N of obs 1 = 8473
Group c: x2*b1 N of obs C = .
Group 2: fm = 1 N of obs 2 = 37546


lnhwage Coef. Std. Err. z P>z [95% Conf. Interval]

overall
group_1 .2908392 .0064496 45.09 0.000 .2781982 .3034802
group_2 .2913806 .0046469 62.70 0.000 .2822729 .3004884
difference -.0005414 .0079493 -0.07 0.946 -.0161217 .0150389
explained -.0574823 .0165785 -3.47 0.001 -.0899756 -.024989
unexplained .0569409 .0173246 3.29 0.001 .0229853 .0908964

---------------------------------------------------------------------------
example 2)

RIF : q(40)
Group 1: fm = 0 N of obs 1 = 8473
Group c: x2*b1 N of obs C = .
Group 2: fm = 1 N of obs 2 = 37546


lnhwage Coef. Std. Err. z P>z [95% Conf. Interval]

overall
group_1 -.0557652 .0062168 -8.97 0.000 -.0679499 -.0435806
group_2 -.121906 .0028722 -42.44 0.000 -.1275354 -.1162767
difference .0661408 .0068482 9.66 0.000 .0527186 .079563
explained .0641534 .0160752 3.99 0.000 .0326466 .0956602
unexplained .0019874 .0163487 0.12 0.903 -.0300554 .0340302
---------------------------------------------------------------------------------

4) In the example 2, it can be easily seen that pay discrimination is 0.3%. (unexplained/difference*100 = 0.0661408/0.0019874)
But the example 1 is somewhat difficult.
Because the difference part is negative(-0.005414).
In this case, the same calculation as the example 1 is (0.0569409/-0.0005414 = ?)
Is not calculation impossible?

5) What is the interpretation of the unexplained part in this case(q70)?
In this quantile(q70), the wages of women without children are higher than those with children?

I am so confused now.
Please help me.

p.s Do you need my data for an answer?
I will link the data if necessary.