Adds CR3 estimator to lm_robust and iv_robust

[lukesonnet]

• Version bump for first post-0.12 release bump
• Adds CR3 standard error estimators to both lm_robust and iv_robust (closes issue #261)
• Updates documentation and tests (lots of test refactoring)
• For now gives S-1 as the degrees of freedom, as the next PR will just divide the DoF from the SE

[macartan]

Luke -- has it be possible to check this implementation against others in R or stata?

On Wed, Oct 24, 2018 at 3:37 PM Luke Sonnet [email protected] wrote:

@lukesonnet commented on this pull request.

In src/lm_robust_helper.cpp https://github.com/DeclareDesign/estimatr/pull/263#discussion_r227931837 :

@@ -426,16 +428,30 @@ List lm_variance(Eigen::MapEigen::MatrixXd& X, H2s.block(0, p_pos, r_fe, r_fe) = ME * X.block(start_pos, 0, len, r_fe) * M_U_ct; H3s.block(0, p_pos, r_fe, r_fe) = MEU * Omega_ct; }

•      } else if (se_type == "CR3") {
  

•        Eigen::CompleteOrthogonalDecomposition<Eigen::MatrixXd> At(
  

•            Eigen::MatrixXd::Identity(len, len) -
  

•              Xoriginal.block(start_pos, 0, len, r_fe) *
  

•              meatXtX_inv *
  

•              X.block(start_pos, 0, len, r_fe).transpose()
  

•        );
  

•        if (Xunweighted.isNotNull()) {
  

•          At_WX_inv =  At.solve(Eigen::MatrixXd::Identity(len, len)).transpose() * X.block(start_pos, 0, len, r_fe);
  

It doesn't know the solution is symmetric, so the transpose is not a no-op, right?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/DeclareDesign/estimatr/pull/263#discussion_r227931837, or mute the thread https://github.com/notifications/unsubscribe-auth/AMJO_Shg7vsyo5FFGicEwdjHvTCt-PGFks5uoMF-gaJpZM4X2uhQ .

[lukesonnet]

@macartan, the tests compare it with clubSandwich. Stata does not have CR3 standard errors, but I think base sandwich in R does now and I could add those tests.

[lukesonnet]

I'm stuck right now because the slight (numerical) differences in the variance-covariance matrix from commarobust or from lm_robust are causing different kinds of failures in computing the F-statistic on windows 32bit machines.

This is an absolutely ridiculous error, but I'm trying to eliminate or minimize the numerical deviations that are causing it.

[acoppock]

OK with me if we just round to the 10th decimal place or something to get around this windows issue

[acoppock]

Thanks so much for adding this.