[Sebastian Kripfganz]

The first differences of time-invariant regressors vanish. As a consequence they will be flagged as "collinear" variables by Stata and hence dropped. You can think of a variable that takes on the value zero everywhere as an extreme case of collinearity to any other variable. The option hascons forces Stata not to do collinearity checks for the first differences.

[Jian Zhang]

Hi Sebastian,

I was trying to replicate your listed code:

xtdpdsys z_score age

xtdpd z_score L.z_score age, dgmmiv(z_score) div(age) lgmmiv(z_score)

The two syntax get identical results. But when I add grade dummies and time dummies, the results are supposed to be same again. But the results are different. Here are the syntax:

xtdpdsys z_score age g1 g2 g3 t1 t2 t3 t4 t5

xtdpd z_score L.z_score age g1 g2 g3 t1 t2 t3 t4 t5, dgmmiv(z_score) div(age g1 g2 g3 t1 t2 t3 t4 t5) lgmmiv(z_score)

I checked their outputs. both two syntax use the exactly same instruments, and the sample are the same sample. Then why would not one get the same results? seems wrong with the dummy variables. But why and how?

Jian

[Sebastian Kripfganz]

The reasons seems to be that some of your dummy variables are dropped because of collinearity but the two commands drop different dummies. Can you confirm this observation? If you restrict the number of dummies in the command syntax (and the corresponding instrument options) such that no variables are dropped any more, the results should be the same again.

[Jian Zhang]

Thanks, Sebastian! You are right! As long as I restrict the two commands to have the same dummies and same instruments, I get the same results, for instance:

xtdpdsys z_score age g2 t3 t5

xtdpd z_score L.z_score age g2 t3 t5, dgmmiv(z_score) div(age g2 t3 t5) lgmmiv(z_score).