IV(2SLS) regression with fixed effects and two-way clustering

[SeongMyeong Kang]

Hello,

I tired IV(2SLS) regression with STATA, but I do not know whether what I do is right or not.

Please let me know better solutions.

Fixed effects: year fixed effect and id-code fixed effect
Clustering: standard errors clustered at the region and year levels.


egen id1= group(id)
egen code1 = group(code)
egen region1 = group(region)
egen double_fixed = group(id1 code1)
egen double_cluster = group(region1 year)

1)

ivreghdfe Y X1 X2 X3 (endogenous variable=instrument variable), absorb(i.year id1 code1) cluster(region1 year)

2)

ivreghdfe Y X1 X2 X3 (endogenous variable=instrument variable), absorb(i.year double_fixed) cluster(double_cluster)

3)

xtset double_fixed
xtreg endogeneous_variable = Instrument_variable X1 X2 X3 i.year, fe
predict endo_var_hat, xb
xtreg Y endo_var_hat X1 X2 X3 i.year, fe

4)

xtreg endogeneous_variable = Instrument_variable X1 X2 X3 i.year, fe vce(cluster double_cluster): error - panels are not nested within clusters

Which one is correct? any better solution?

[SeongMyeong Kang]

typo: vreghdfe -> ivreghdfe (revised)

[Tilman Deissinger]

I have the same problem and do not know which way is correct. Can somebody please explain this?

[Andrew Musau]

ivreghdfe is from https://github.com/sergiocorreia/ivreghdfe (FAQ Advice #12).

For two-way clusters, you need two variables, so

ivreghdfe Y X1 X2 X3 (endogenous variable=instrument variable), absorb(i.year id1 code1) cluster(region1 year)

is correct. All the rest are one-way clusters at different levels. Two-way clustering appears to be applied in an ad-hoc manner since statistical software packages made it accessible. However, there needs to be a theoretical basis for this, so it's important not to employ it simply because it's available. In the case of panel data, clustering using the panel identifier is usually sufficient.