Difference between manual 2sls and ivregress

[Alessandro Fusari]

I have a regression with an endogenous variable for which I have an instrument.
By computing
ivreg income (immigrants=enclave) YearDummy*
I obtain a given coefficient.

If instead i first compute the first stage and then regress the dependent variable on the exogenous fitted value of the endogenous variable I obtain a completely different result.
reg immigrants enclave YearDummy*
predict immhat, xb
reg income immhat YearDummy*

If I'm not mistaken the result shoud be the same, as a 2sls regression is computed that way. Is there a reason why this is actually not the case?

[Andrew Musau]

It can happen if you are using different samples. See, e.g., https://www.statalist.org/forums/for...ares-procedure

[Fei Wang]

Other than the issue of inconsistent samples pointed out by Andrew, using predicted regressor requires adjustment of standard errors. So if you fix the problem of inconsistent samples, you'll find coefficients are identical in the two ways, but SEs differ.

[Øyvind Snilsberg]

Other than the issue of inconsistent samples pointed out by Andrew, using predicted regressor requires adjustment of standard errors. So if you fix the problem of inconsistent samples, you'll find coefficients are identical in the two ways, but SEs differ.

which is why one should use ivregress

[Craig Stocky]

Other than the issue of inconsistent samples pointed out by Andrew, using predicted regressor requires adjustment of standard errors. So if you fix the problem of inconsistent samples, you'll find coefficients are identical in the two ways, but SEs differ.

Hi both - I've seen this issue come up a few times. I'm running a manual IV currently - and I've seen elsewhere that I need to adjust the standard errors, but I don't know the reason why. Could you please help to explain / point me in the direction of the relevant literature?

[Andrew Musau]

The correct search term is the generated regressor problem. See https://en.wikipedia.org/wiki/Generated_regressor and the references therein.

[Nitin Jain]

Hi both - I've seen this issue come up a few times. I'm running a manual IV currently - and I've seen elsewhere that I need to adjust the standard errors, but I don't know the reason why. Could you please help to explain / point me in the direction of the relevant literature?

Hi, Craig, Were you able to identify how to correct the SE in the first stage manually? Thanks.