suest after bootstrap

[Rijo John]

Hi Statalist,
Trying to run suest after bootstrap using the code below:

cap program drop Expelast
program define Expelast, rclass
args i
qui areg yvar xvar1 xvar1 xvar2 xvar3 xvar4 if dum==`i', absorb(clust)
local a0 =_coef[xvar1]
qui sum yvar if dum==`i'
local vbar =r(mean)
return scalar Expel`i' =1-b1+(`a0'/`vbar')
end

forvalues i=1/3 {
bootstrap Expel`i'=r(Expel`i'), reps(1000) seed(1): Expelast `i'
estimates store exp_el`i'
}
suest exp_el*

Here the variable "dum" takes values 1,2, and 3. The code runs just fine. the command <estimates dir> correctly returns all the three stored results.
However, the last line suest returns the error "exp_el1 was estimated with a nonstandard vce (bootstrap)"
I would like to be able to test the coefficient Expel1=Expel2=Expel3, etc. after suest. Is there any way to pull it off?
I tried the code

suest exp_el*, vce (robust)

but to no avail. But it says "vcetype bootstrap not allowed".

Thank you in advance for any potential leads.

Best,
Rijo.

[Rijo John]

Bumping it up, just in case.

[Sebastian Kripfganz]

The error message is quite clear about it: You cannot use suest after bootstrap. suest requires score functions (internally generated with the postestimation command predict for each estimation result) in order to compute the variance-covariance matrix for the system of estimates, but those scores are not available after your Expelast program.

[Rijo John]

The error message is quite clear about it: You cannot use suest after bootstrap. suest requires score functions (internally generated with the postestimation command predict for each estimation result) in order to compute the variance-covariance matrix for the system of estimates, but those scores are not available after your Expelast program.

Thank you. Can I ask what could be the best way to make a comparison of the bootstrapped coefficients across subgroups (variable dum in my program Expelast)?

[Joro Kolev]

Instead of bootstrapping the coefficients separately, bootstrap some statistic that tests that they are the same.

[Rijo John]

Instead of bootstrapping the coefficients separately, bootstrap some statistic that tests that they are the same.

Thank you. That was one approach I saw elsewhere. However, in this case, because each statistic is estimated in a forvalues loop, one at a time, I couldn't find a straight forward way to bootstrap something like say, Expel1-Expel2=0 since only one of them would be available to the program at a time given the way the program is currently implemented.

[Joro Kolev]

Move your loop -forvalues i=1/3 {}- inside the programme you are bootstrapping.

And then the simplest statistic that you can bootstrap is probably (Expel1-Expel2)^2 + (Expel2-Expel3)^2.