IVREGHDFE does not allow for score option using SUEST

[james xu]

Dear colleague,
I figured this issue is exclusive to ivreghdfe so I post here. Basically, I'm trying to comparing if the two coefficients are statistically indifferent from two 2SLS second-stage regressions.

I am following the help guide of suest and conduct the following commands:

qui ivreghdfe DISP  log_ME  (h_diffs =test lngdp epu_state) if large, a(cusip st year)
estimates store large
qui ivreghdfe DISP  log_ME  (h_diffs =test lngdp epu_state) if !large, a(cusip st year)
estimates store small
suest large small

Basically I divide my sample into two groups, large=1 and large=0. I got the error message when I try to run the last command:

unable to generate scores for model large
suest requires that predict allow the score option

Does anyone know how I could enable score option with ivreghdfe or what should I do to get around this error?

Thank you!

[George Ford]

Best option is to put the two models together with dummy/interactions for large, and test them directly in a single model.

It would be nice if suest could be modified to accommodate more models.

[Sebastian Kripfganz]

It would be nice if suest could be modified to accommodate more models.

It is not a problem with suest. To correctly calculate the variance-covariance matrix, it relies on the previous estimation command to provide scores. If the previous command, here ivreghdfe, does not provide those scores, then there is nothing that suest can do about it.

From a programmer's perspective, providing scores is quite cumbersome and most user won't need them. I thus understand why many commands (even official Stata commands) do not provide scores. With IV estimators, the additional complication is that equation-level scores do not exist, only parameter-level scores. This would require additional adjustments to make it work with suest.

[james xu]

I do appreciate all your responses! I will think of another way to get around this.

[George Ford]

for what you're doing, you only need one model.