RIF regression and RIF decomposition following Heckley's methodology

[atta asif]

I want to make analysis like
https://www.google.com.pk/url?sa=t&r...YZxnFLNF1RYiYe.
however our response variable is child health care such as (no-vaccinated, partially-vaccinated; fully-vaccinated). we want to measure socioeconomic inequality using RIF-CI-OLS methodology and rig decomposition presented by Heckley's et al. 2016.
please any body could guide me how can I adopt the methodology presented in the article provided in above link.

thanks and best wishes
ATTA

[FernandoRios]

Hi Atta
Im almost completely sure you can do that type of analysis using the command rifhdreg (you need to install it from ssc install rif).
If you take a look at the help file (help rifvar and help rifhdreg) you will see how any of the concentration indices can be estimated with rifhdreg directly, or how can you create the RIF's using rifvar (an extension for egen)
rifhdreg works like "regress", so you should be able to easily estimate what you need to.
At the end of the helpfile, I also provide my own interpretation of RIF regressions and RIF decompositions that you may find useful.
Best Regards
Fernando

[atta asif]

Respected FernanadoRios thank you for your guideline.

as I my response variable is of ordered type so we have to estimate ordered probit model first (as in the article I have given the link in above post). our explanatory variables are categorical such as education level, residence type etc. so how we may estimate inequality in this case. please if you have some stata example simple or some other guiding material such as any simple article which have discussed the same casev , please share it if possible.

I shall be very thankful for this kindness.

best
atta

[FernandoRios]

Hi Atta
I see. If I understand the process correctly, it would have to be something as follows:

webuse nhanes2f, clear
oprobit health female black age c.age#c.age
predict xb_pr,xb
sum xb_pr
gen rchealth=(xb_pr-r(min))/(r(max)-r(min))
rifhdreg rchealth female black age c.age#c.age, rif(eindex(rchealth) lb(0) ub(1) )

the only difference would be that within the option "rif" you will need to add the income or wealth variable instead of the RCHEALTH variable).

I also still read carefully the paper they cite:
Van Doorslaer E, Jones AM. Inequalities in self-reported health: validation of a new approach to measurement. J Health Econ. 2003;22:61–87.
As it may explain alternatives to how to implement the initial transformation.

Best

[atta asif]

thank you for help

I need one more guidance.
1. how we may perform the RIF-decomposition process.
2. please share some ideas about interpretation of this decomposition outputs in my case.

thaks

[FernandoRios]

Hi Atta,
for the decomposition, I would look into Heckley's et al. 2016.
My understanding from what they do is to simply report the output from the RIF regression. I often do not call this decomposition, but seems to be what it is done.
Better yet, I would contact the author Gawain Heckley, to ask for clarification. Perhaps he could even share his data or code used for the paper.
In terms of interpretation, i would also relay on Heckley's et al. 2016. or the examples on my paper (see in the help files)
HTH

[atta asif]

thanks respected Farnando for guidance.

I have attached the ADO file of codes which they have given as supplementary file.
one more request is:
I have read the your paper of RIF-regression and RIF-decomposition. we may also want to decompose on the lines you have given. please give some simple example and interpretations in this regards if you fell easy. Moreover, the predicted values as you suggested in earlier post on this page using rifhdreg can be used for the decomposition process? if yes, then please guide me using a simple syntax of command using simple variables.

I shall be very thank full to you

[atta asif]

thanks respected Farnando for guidance.

I have attached the ADO file of codes which they have given as supplementary file.

one more request is:

I have read the your paper of RIF-regression and RIF-decomposition. we may also want to decompose on the lines you have given. please give some simple example and interpretations in this regards if you fell easy. Moreover, the predicted values as you suggested in earlier post on this page using

rifhdreg

can be used for the decomposition process? if yes, then please guide me using a simple syntax of command using simple variables.

I shall be very thank full to you

Heckley's ADO file for data analysis.pdf

[FernandoRios]

Hi Atta,
Sorry I didnt respond to your previous email.
So couple of points.
1. rifhdreg does exactly what the PDF you sent me does. So, if you would like to replicate their tables you simply run the regression, as I previously suggested. Something like:

rifhdreg rchealth female black age c.age#c.age, rif(eindex(rchealth) lb(0) ub(1) )

My understanding from Heckleys et al paper is that the coefficients you get from this regression is what they use and interpret as decomposition. The first paper that you cite does exactly the same thing.
2. if what you are more interested is a Oaxaca-Blinder decomposition, that can be easily done using the command oaxaca_rif (also part of "rif")

gen age_sq=age*age
oaxaca_rif rchealth black age age_sq, rif(eindex(rchealth) lb(0) ub(1) ) by(female) 
** note that oaxaca_rif does not work with factor notation.

or you can use the more advanced strategy (reweighted analysis).

oaxaca_rif rchealth black age age_sq, rif(eindex(rchealth) lb(0) ub(1) ) by(female)  rwlogit(black age age_sq)

3. for the interpretation, you can follow the examples I have on my paper. The working paper has examples for the analysis of inequality in the US using CPS data, and the newer version has an example using the "oaxaca" toy data.

4. For further examples on how to use the syntax, please see the examples within each command. Most of them are executable, so you should be able to replicate them.
HTH
Fernando

[Magashi Joseph]

Hi Atta
Im almost completely sure you can do that type of analysis using the command rifhdreg (you need to install it from ssc install rif).
If you take a look at the help file (help rifvar and help rifhdreg) you will see how any of the concentration indices can be estimated with rifhdreg directly, or how can you create the RIF's using rifvar (an extension for egen)
rifhdreg works like "regress", so you should be able to easily estimate what you need to.
At the end of the helpfile, I also provide my own interpretation of RIF regressions and RIF decompositions that you may find useful.
Best Regards
Fernando

Dear FernandoRios
I have a dependent variable which is binary (stunting=0 and no stunting=1). Do I need to transform this variable after fitting a logistic model so that I can have a variable that can be used as dependent variable in the rifdreg model and then oaxaca_rif decomposition? If yes how do I it?

Thanks in advance
Joseph

[FernandoRios]

Hi Magashi,
I think the answer will depend on the question you are trying to answer.
As far as I know, Wagstaff and Erreygers ( and other indices described in Heckley et al 2016 https://doi.org/10.1016/j.jhealeco.2016.03.006.) are concentration based indices that already account for the nature of the variable to measure inequality. These measures are already programmed within rifhdreg (see help rifhdreg or help rifvar )
Now, if you want to analyze the dependent variable directly, I think it will depend on what is it you are trying to analyze. (which distribution function)
Quantiles are basically meaningless when the dependent variable is discreet, or dichotomic as in your case. Variances may be a better measure.
You could get the transformed fitted prediction, but I cannot say if that is correct or not. My answer to Atta was based on a paper he cited himself. You will need to see if that is appropriate for you or not.
HTH
Fernando

[Magashi Joseph]

Hi Magashi,
I think the answer will depend on the question you are trying to answer.
As far as I know, Wagstaff and Erreygers ( and other indices described in Heckley et al 2016 https://doi.org/10.1016/j.jhealeco.2016.03.006.) are concentration based indices that already account for the nature of the variable to measure inequality. These measures are already programmed within rifhdreg (see help rifhdreg or help rifvar )
Now, if you want to analyze the dependent variable directly, I think it will depend on what is it you are trying to analyze. (which distribution function)
Quantiles are basically meaningless when the dependent variable is discreet, or dichotomic as in your case. Variances may be a better measure.
You could get the transformed fitted prediction, but I cannot say if that is correct or not. My answer to Atta was based on a paper he cited himself. You will need to see if that is appropriate for you or not.
HTH
Fernando

Hello FernandoRios

Thank you for clarification. Much appreciated

[atta asif]

Dear Farnando

thanks for your helpful comments.
one think I need to be clear is that if we use the reweighted command for decomposition command we get decomposed structured effect and composition effect. please give some interpretations of these decomposed part by giving a single variable if yo have some time. I have read your suggested paper more than one time but not cleared yet. please define specification error and reweighting errors. what they show and why they are important. more over during reweighting process we used logit or probit command which need binary response variable . so, either grouping variable will be response variable here?

best wishes
Atta

[atta asif]

Dear Farnando

thanks for your helpful comments.

one think I need to be clear is that if we use the reweighted command for decomposition command we get decomposed structured effect and composition effect. please give some interpretations of these decomposed part by giving a single variable if yo have some time. I have read your suggested paper more than one time but not cleared yet. please define specification error and reweighting errors. what they show and why they are important. more over during reweighting process we used logit or probit command which need binary response variable . so, either grouping variable will be response variable here?

best wishes

Atta

[FernandoRios]

Hi Atta
Perhaps you should read Firpo, et al (2018). They may be doing a better job explaining the Reweighted decomposition, and also provide an example on how to interpret the results.
They also define the specification and reweighting errors, and should eco my own description of what they are, but putting more emphasis on why are they important.
For the reweigthing process, the variable defined in "by()" is the dependent variable. So if you say, oaxaca_rif wage age, by(female) rwlogit(age), "female" is the dependent variable for the logit model.
HTH