Understanding boottest

[Samuel Makikalli]

Hello,

I am running a simple OLS but have only 12 clusters. Following the "Fast and Wild: Bootstrap Inference in Stata Using boottest" paper I ran a boottest following the simple regression. I use the following commands:

ivreghdfe sharedk inflation_yearly, a(country date) cl(country date) bw(12)

boottest inflation_yearly, boot(wild) bootcl(country date)

The results show a negative coefficient of -0.129 and the boottest result shows a significant value (t-stat=3.91), yet the 95% confidence interval is [0.2991, 1.053]. I am confused regarding the interpretation of the result. Does boottest imply that the simple regression result is extremely biased? Any guidance into how to read these results would be much appreciated.

[David Roodman]

Hi Samuel,
There are a few issues here:
* boottest doesn't support ivreghdfe.
* boottest doesn't support the Newey/Driscoll-Kraay type serial correlation adjustment triggered by the bw() option. This is not conceptually compatible with the wild cluster bootstrap because it doesn't break the data into non-overlapping clusters in the time dimension.
* boottest does support reghdfe but only with one set of fixed effects.

I could easily fix the first, but not the second or third.

Best,
--David

[David Roodman]

I should that if it is practical to explicitly include i.country or i.date in your regressor list (and I'd think it is), then that will solve the third issue. And at least here you are not doing an IV regression, so you could just use reghdfe, which boottest supports. The bw() problem though has no straightforward solution. You would need to abandon that to use boottest.

[Manoel Calvo]

Hi David,
Does boottest support command xtpmg?
Thanks!

[David Roodman]

Sorry, no.

[Manoel Calvo]

Thanks for your answer David, I was obtaining some confusing results as Samuel using boottest after xtpmg.

[Gabriel Facchini]

Dear david, and community,

Continuing with better grasping what boottest does, ...should the boottest and cgmwildboot command give the same results? (using the same seeds and regression). I find that the results are different (although they are closer between each other than with the non-bootstrapped errors).

Below the example using the nlsw88 data like in the cgmwildboot help file.


set seed 293867483
webuse nlsw88, clear
drop if industry==. | tenure==.

regress wage tenure , cluster(industry)
boottest {tenure} , cluster(industry) rep(1000) seed(12345) nograph boottype(wild) bootcluster(industry)

cgmwildboot wage tenure, cluster(industry) seed(12345) reps(1000) bootcluster(industry)


In the case of the cgmwildboot the CI for tenure is [.13334021 .23370585], while when using boottest the CI for tenure is [.1250953 .2410478].

thansk for any help!