Replicating results of Fuzzy Regression Discontinuity with covariates between rdrobust and 2SLS

[Brayden Paur]

Hello,

I am currently struggling to get the same results when I run a fuzzy RD with covariates using rdrobust compared with ivregress 2SLS. I know there have been posts concerning how to get the same results between the commands for fuzzy RD and sharp RD with covariates but I have not found any information on fuzzy RD with covariates.

I apologize, I struggled to find a relevant dataset to use here but I am able to get the same results when I run the fuzzy RD with no covariates but as soon as I add a few dummy variables for fixed effects, I am unable to get the same estimates. For the fixed effects covariates, let's say I have three years and four regions. I will omit the first year and region in the following regressions.



What follows results in similar estimates.

*no cov
*assume the running var is centered at zero
gen above=running_var>0

rdrobust outcome running_var, fuzzy(fuzzyvar) kernel(uni)
ivregress 2sls outcome (fuzzy=above) running_var c.running_var#i.above if running_var< `e(h_r)' & running_var>-`(e(h_l)'

I am not able to do the same as soon as I use the following covariates.

rdrobust outcome running_var, fuzzy(fuzzyvar) kernel(uni) covs(year2 year3 region2 region3 region4)
ivregress 2sls outcome year2 year3 region2 region3 region4 (fuzzy=above) running_var c.running_var#i.above if running_var< `e(h_r)' & running_var>-`(e(h_l)'

Any guidance or suggestions would be greatly appreciated.

Thank you!

[Jackson Monroe]

I assume it is because the coefficients on

year2 year3 region2 region3 region4

are allowed to vary according to the side of the threshold they are on in the rdrobust specification and not in yours.

[Brayden Paur]

I assume it is because the coefficients on

year2 year3 region2 region3 region4

are allowed to vary according to the side of the threshold they are on in the rdrobust specification and not in yours.

Hi Jackson,

Thank you for your response. I want to ensure that I am understanding your point correctly: what I believe you are suggesting is to include an interaction term for each of the variables you mentioned (year2, year3, region2, region3, region4) with the variable 'above.' I attempted this and it still resulted in an estimate that did not closely resemble the estimate from rdrobust.

If this is not what you meant, please feel free to correct me. Any further thoughts or ideas would be fantastic! Thank you for your feedback thus far.

[Lucas Smith]

Hi Jackson,

Thank you for your response. I want to ensure that I am understanding your point correctly: what I believe you are suggesting is to include an interaction term for each of the variables you mentioned (year2, year3, region2, region3, region4) with the variable 'above.' I attempted this and it still resulted in an estimate that did not closely resemble the estimate from rdrobust.

If this is not what you meant, please feel free to correct me. Any further thoughts or ideas would be fantastic! Thank you for your feedback thus far.

Hi Brayden,
Did you receive any updates regarding these questions? I'm also facing the same problem.
Thank you.

[Marianna Portincasa]

Hi Brayden,
Did you receive any updates regarding these questions? I'm also facing the same problem.
Thank you.

Hi Lucas, Have you found any solution? I have the same problem as you

[Lucas Smith]

Hi Lucas, Have you found any solution? I have the same problem as you

Sadly, I haven't.