Using recentred influence functions (RIFs) for regression analysis

[Ashani Abayasekara]

Hi everyone,

I'm interested in using recentred influence functions (RIFs) to identify socio-demographic factors that are the largest contributors to inequalities in gambling harms, and how these have changed over time. I'm interested in identifying factors that are associated with harm and are themselves unequally distributed. So far, I have used concentration indices (mainly the Erreygers's index) to identify levels of inequality by income, for different gambling-harm related outcomes, using code such as:

conindex pgsi [pweight=scq_wgt], rankvar(income) robust cluster(xwaveid) truezero bounded limits(0 27) erreygers

I have two main questions:

1. Are RIFs the appropriate regressions to use in this case, and do the RIF regressions suffice or do I also need to conduct decomposition analysis?
2. I'm hoping to use the commands rifvar and rifhdreg developed by Fernando Rios-Avila for this analysis, and do something as follows; is this the correct approach?

egen rif_pgsi = rifvar(pgsi), eindex(income) lb(0) ub(27) //outcome is pgsi which ranges from 0 to 27

rifhdreg rif_pgsi age female, rif(eindex(income) lb(0) ub(27)) robust //how much do age and gender contribute to inequalities in the pgsi distribution

Also, can I use absolute income values here or do I need to rank income by quintiles first?

Many thanks,
Ashani

[FernandoRios]

Hi
Couple of notes.
1. May be good if you read the Health Econ paper cited on my RIF-regressions paper. They are the best reference to use for using RIF for this type of indices. Otherwise, yes, its ok to use ut
2. You do not need to create the RIF first (egen ...) if you are going to use rifhdreg. The command rifhdreg Creates the RIF for you on the background.
F

[Ashani Abayasekara]

Thank you very much for your reply. I will look at the health econ paper, that is the one by Heckley et al. published in the Journal of Health Economics in 2016?

Re your second point, when I run only the rifhdreg command, the results I get are substantially different to those I get when I run both commands. If using only rifhdreg, would the outcome variable just be pgsi? This is what I tried but perhaps this is incorrect?

[FernandoRios]

Yes those results would be different because what you were doing right now is wrong.
So whats happening.
rifhdreg does the following
1. get rif using rifvar
2. use regress on the rifvar

what you are doing
1. get rif using rifvar
2. use reghdrif on the rifvar
3. reghdrif estimates the RIF of rifvar...call it rif_rifvar This is now meaningless
4. use regress on rif_rifvar. this is wrong

F

[Ashani Abayasekara]

Thank you for the clarification.

[Ashani Abayasekara]

Hi,

Revisiting this question, I have used your command rifhdreg to look at contributors to inequality in gambling harm. I'm still not fully sure on how to interpret the results. Below is my code and output. For this simple case, none of the coefficients are statistically significant, so would this mean that none of these factors contribute to observed inequalities in the PGSI score?

rifhdreg pgsi $pre1 if wave==15, rif(eindex(hheqinc) lb(0) ub(27)) robust

Also, is the rifhdreg command accomplishing what O'Donnell et al. (2007) show in chapter 13 of the World Bank Report on analysing health equity on decomposing inequalities in one step, that is combining the impact of each regressor on PGSI (elasticities), and each regressor’s degree of unequal distribution across income ranking (measured by the concentration index)? Is there a way to separate these two effects using rifhdreg?

[FernandoRios]

I have no idea about O'Donnell et al (2007) but i would say they do not do RIF for their analysis, since the method came after that.
For interpretation, I would redirect you to the Health Economics article that discussed the use of those indicators to the use of Rif-regressions.
F

[Ashani Abayasekara]

Thanks, I'll refer to the Health economics paper. Just one more clarification - that paper mentions that "decomposition of the index is then performed by a two-step procedure of first computing the RIF of the rank dependent index, and then regressing the RIF on a set of covariates yielding the marginal effects of the covariates on the index." From what I understand, the rifhdreg command executes both of these steps in one go? Is there a way I can use this command to do each step separately?

[FernandoRios]

Yes, RIFhdreg does that
If you look over the help file it will tell you how to best do it separately.
F