[Samira Choudhury]

Hi Fernando,
I have a quick question regarding the stata code to do Oaxaca RIF counterfactual decompositions.
For example, if we want to do this for the lowest quantile-
oaxaca_rif lnwage educ exper tenure, by(female) wgt(0) rif(q(10)) rwlogit(educ exper tenure) Is it correct if I specify wgt as 0? Thanks for your help!!
Samira.

[FernandoRios]

Yes that is correct
look into the headings after you do that so you can see what the counter factual is
it would say something like
x1 ~>x2 which means characteristics of group 1 were freweighted to look like those in x2
hth
fernando

[Samira Choudhury]

The reason I'm asking is because I'm getting completely different set of results when I specify wgt as 0 instead of wgt(1). It would be good to know what is the difference.

Thanks for your help,
Samira.

[FernandoRios]

I See
there are couple of options
a) Oaxaca with reweighting produces something equivalent to treatment effect on treated and untreated, depending on the w(#). And it might just be that the treatment has different effects on different groups.
b) you have problems of overlapping.
For this, a simple thing you could do is to manually run the probit model, predict the propensity scores and plot them for both groups.
HTH
Fernando

[Samira Choudhury]

Hi Fernando,
Thank you for this idea. My graph looks fine. For the decomposition results similar to Table 4 in the paper by Srinivasan et al. 2013 (if I am correct they computed the reweighted OB decomposition and not counterfactual for the two fold decomposition), I found different results when using wgt(0) vs wgt(1). For instance, when I run oaxaca_rif using wgt(0), the covariate and coefficient effects are larger than the raw gap between the 2 groups and I find that the contribution attached to covariate and coefficient effects for each of the quantiles (10-90) is more than 100%. The total contribution of covariate and coefficient effect does not add up to 100% like in Table 4.

Instead, when I run oaxaca_rif using wtg(1), the contribution of covariate and coefficients adds up to 100%. In terms of the contribution of individual characteristics, some of them are above 1000%. You have previously mentioned maybe it's better not to show those contributions.

I'm hoping you can advise me as to whether I should use wtg(1) or wtg(0).

Thank you so much for your help,
Samira.

[FernandoRios]

Can you send a log file of the output ?
i need to see what is going on
thank you

[Samira Choudhury]

Thank you so much for your help Fernando. I'm attaching tables 1)with wtg(0) 2) with wtg(1)

[FernandoRios]

Hi Samira
I not really sure what you meant with ". The total contribution of covariate and coefficient effect does not add up to 100% like in Table 4. " I saw the tables you sent, and the contributions add to the totals as they should.
What I did observe is that you have a problem of counterfactual identification. Basically, your reweighted error is coming large and significant, and that is something you need to fix.
As i suggested before, you need to do an overlapping test.
You cant do that directly with oaxaca_rif, but you can use the test from teffects to do that.
Do something like:

webuse cattaneo2
* This is the outcome for the oaxaca
oaxaca_rif bweight mmarried, by(mbsmoke) rwprobit(mmarried c.mage##c.mage fbaby medu) rif(mean) w(1)
* This is what you would get with teffects
 teffects ipwra (bweight ) (mbsmoke mmarried c.mage##c.mage fbaby medu, probit), atet
* With this you check balance. Regardless of the RIF
tebalance summarize

Im suspecting that you have serious balance problems even after reweighting which is causing some of the weird results that you have.
HTH
Fernando

[Samira Choudhury]

Hi Fernando,
I think I understand what you mean. So I have done this overlapping test. Do you think the problem is now fixed?

Thank you so much for your help.
Samira.

[Samira Choudhury]

Hi Fernando,
My results are still the same. The reweighted errors are still large and significant. How can I fix this?

Thank you for your help,
Samira.

[FernandoRios]

Hi Samira
Sorry i wasn't clear. The overlapping test simply shows you that you have an overlapping/balance problem even after reweighting, but it does not fix it.
You can read on general advice regarding balance problems in the treatment estimation literature, or for propensity score matching.
In any case, there are couple of things you can do.
1. Add squared terms and interactions in the reweighting option. Say, add logmpce^2 c0_5^2 or interactions. The point here is to play with the probit/logit specification until ALL standardize differences are small (and possible the reweighted error is non significant)
2. Before doing that, also check on the density of the distributions. Im suspecting that things like logmpce will have very different distributions across groups. one of the teffects options that you can do is:
teffect density logmpce , and it will show you how that looks before and after reweighthing. Pay attention to the tails, which are usually the main problems for overlapping.
HTH
Fernando

[Samira Choudhury]

Hi Fernando,
Thanks for the clarification. For the following command:
oaxaca_rif bweight mmarried, by(mbsmoke) rwprobit(mmarried c.mage##c.mage fbaby medu) rif(mean) w(1) Where is the reweighting option where I can include squared terms and interactions? Sorry to ask this but I am really confused. Also can you give an example as to what is meant by "all standardize differences are small"?

Thanks for your help,
Samira.

[Samira Choudhury]

Do you mean oaxaca_rif bweight mmarried, by(mbsmoke) rwprobit(mmarried mmaried ^2 c.mage##c.mage fbaby^2 fbaby medu medu^2) rif(mean) w(1) ?

Thanks for your help,
Samira.

[FernandoRios]

Yes that is a good example.
But also check on the densities of across groups of all the variables you use, using teffect density command.
Think of it as working with the model specification.
Fernando

[Samira Choudhury]

Thank you so much for your help! Really appreciate it.
Samira.