[Aashima Singhal]

FernandoRios

Respected Sir,

I recently go through your article "Recentered influence functions (RIFs) in Stata: RIF regression and RIF decomposition". Sir, I have confusion regarding if my outcome variable is binary, then which command should I go for. Is these commands are suitable for only continuous variable or can determine inequality for binary health outcome as well.
I will highly grateful to you, if you could please suggest.

Thanks and Regards

[FernandoRios]

Hi Asshima,
Unfortunately ALL of the RIF's i programmed in the command are valid only for continuous variables. None for discrete. Except, perhaps some of the health related ones.
Overall, if you are looking at binary variables, I don't think you need to worry about unconditional quantiles, (which is what most ppl use this for)
F

[Aashima Singhal]

Sir, my outcome variable is related to health only. For instance: hospitalization occurred or not (1,0), for this I want to check the inequality across various socio-demographics variables. I am not able to understand in such a scenario can I go for RIF, if it produces the results for binary outcome variable as well? If yes, then which command ?
Looking forward to hear from you.

Thanks and Regards

[FENG WEI]

@FernandoRios
Dear Sir,

Recently, I am reading Havnes and Mogstad (2015), they use the RIF-DID for the main results, while two alternative methods (QDID and CIC) are also used for robustness check. I am wondering which command can be used for QDID method and what assumptions should be checked by using the data, e.g., whether the command trend assumption should be checked as well for QDID?

Thanks

[FernandoRios]

Hi Feng
I have read the paper before, but I'm not sure about a particular command for their application of QDID. It seems to me that is an application of machado mata decomposition, which uses multiple quantile regression simulations, but i m not certain about it.
I should also say that i partially disagree with how they are defining RIF-DID components, but that could simply be a dissagreement in implementation.
HTH
F

[FENG WEI]

@FernandoRios
Dear Sir,

Currently, I have a policy which is implemented in 2014, my dataset includes years from 2008 to 2016. I am wondering whether I can just set post >= 2014, and run the RIF-DID regression, it seems to me that we may need some pre-assumption, can you please explain how we should do it?

Thanks again for your help.

[FENG WEI]

Hi Feng
I have read the paper before, but I'm not sure about a particular command for their application of QDID. It seems to me that is an application of machado mata decomposition, which uses multiple quantile regression simulations, but i m not certain about it.
I should also say that i partially disagree with how they are defining RIF-DID components, but that could simply be a dissagreement in implementation.
HTH
F

@FernandoRios
Dear Sir,

I have a quick question regarding how to change the level of confidence interval using rifhdreg, my command is follows:

rifhdreg `v' treated post interaction, abs(firm_id) over(DD) rif(q(`i')) cluster(industry) level(90)

It still reports 95% CI instead of 90%, I am wondering why this is the case. Also a quick question for combining DID with QTE, is it necessary to add the over(DD) that you suggest from another post?

Thank you very much in advance.

[FernandoRios]

That was an oversight of mine. i ll try to add something to correct that next
for now, you could do the following

_coef_table, level(90)

right after the rifhdreg outcome

[Shivani Yadav]

hi
I need a help in understanding the Oaxaca_rif command. I was successfully able to implement the command in caae of linear decomposition. However, I am facing difficulties in using the same command when the dependent variable is non-linear. What shall be the correct usage to get the quantile decomposition results in case of probit model, say?

[Shivani Yadav]

hi
I need a help in understanding the Oaxaca_rif command. I was successfully able to implement the command in caae of linear decomposition. However, I am facing difficulties in using the same command when the dependent variable is non-linear. What shall be the correct usage to get the quantile decomposition results in case of probit model, say?
The following is the error that I am getting:

dropped coefficients or zero variances encountered
specify -noisily- to view model estimation output
specify -relax- to ingnore

[FernandoRios]

The error you see is not because of your dep variable, but model specification
In others words, you may have variables that are zero for one of your groups. But for that you need to explore the model specifications

[Shivani Yadav]

I wish to say that I have used the same model specifications (i.e. exactly the same model) as I did while computing the non-linear decomposition using nldecompose command (i.e for differences in predicting average probabilities). I have also rechecked the variables and as per what you are saying, there will be zeroes for a variable when it is a dummy variable (but none of the variable is zero for any of the group). I am sharing the command here for reference:
oaxaca_rif fanc sector1 wealth1-wealth4 mage1-mage6 meduc1-meduc3 birthorder1-birthorder5 mm1 [aw=weight], by(caste) rif(q(25)) noisily

[FernandoRios]

Perhaps something else to consider
rifs for quantiles are not well defined with non-continuous data
or data with excess zeroes, binary data etc
simply bc the rif cannot be correctly estimated

[Shivani Yadav]

FernandoRios Thanks for the reply. Is there any other command that you could suggest in that case? to do quantile decomposition with a set of categorical variables?

[FernandoRios]

Perhaps if you find a model that estimates quantile regression for categorical variables