Boostrapping with two-stage estimation

[James Henning]

Hi all,

I'm estimating a two-stage model where a coefficient from the first stage is used as a regressor in the second.

The first stage estimates region-specific coefficients (either by running separate regressions for each region or by using region interactions).

The second stage pools observations from all regions and uses the region-level coefficients from the first stage as a regressor (in addition to other individual-level variables).

Samples used for the two stages do not overlap - is it still possible to bootstrap standard errors or otherwise correct them to account for sampling error in the first stage?

There is also the issue of clustering since an aggregate-level variable is used in an individual-level regression.

Would be very grateful for any pointers or ideas.

[James Henning]

Resolved - code posted at http://www.statalist.org/forums/foru...n-boostrapping.

Still looking for ways to adjust standard errors for clustering - would appreciate any suggestions on the new thread!

[Clyde Schechter]

The -bsample- command itself permits you to do bootstrap sampling that respects clusters. See -help bsample- You will need to specify the -cluster()- and -idcluster()- options.

[James Henning]

Ah thank you! Didn't think of looking at the -bsample- command itself.