Revisiting instrumental variables and the classic control function approach, with implications for parametric and non-parametric regressions

"We show that the well-known numerical equivalence between two-stage least squares (2SLS) and the classic control function (CF) estimator raises an interesting and unrecognized puzzle. The classic CF approach maintains that the regression error is mean independent of the instruments conditional on the CF control, which is not required by 2SLS, and could easily be violated. We show that the classic CF estimator can be modified to allow the mean of the error to depend in a general way on the instruments and control by adding the unconditional moment restrictions maintained by 2SLS. In this case 2SLS and our generalized CF estimator are no longer numerically equivalent, although asymptotically both converge to the true value. We then show that our generalized CF estimator is consistent in parametric or non-parametric settings with endogenous regressors and additive errors. For example, our estimator is consistent when the conditional mean of the error depends on the instruments while the nonparametric estimator of Newey, Powell, and Vella (1999) based on the classic CF restriction is not. Our new approach is also not subject to the ill-posed inverse problem that affects the non-parametric estimator of Newey and Powell (2003). Our estimator is easy to implement in standard programming packages - it is a multi-step least squares estimator - and our monte carlos show that our new estimator performs well while the classical CF estimator and the non-parametric analog of Newey, Powell, and Vella (1999) can be biased in non-linear settings when the conditional mean of the error depends on the instruments"--National Bureau of Economic Research web site.