Convex Combination of Ordinary Least Squares and Two-stage Least Squares Estimators

TL;DR

This paper proposes a convex combination of OLS and TSLS estimators, called CLS, which adaptively balances bias and variance to improve estimation accuracy in the presence of confounders.

Contribution

It introduces the CLS estimator that optimally combines OLS and TSLS, estimates the mixing proportion from data, and demonstrates improved performance over TSLS.

Findings

CLS outperforms TSLS in MSE in simulations

The mixing proportion is uniquely determined and consistently estimable

Application to econometric data illustrates practical benefits

Abstract

In the presence of confounders, the ordinary least squares (OLS) estimator is known to be biased. This problem can be remedied by using the two-stage least squares (TSLS) estimator, based on the availability of valid instrumental variables (IVs). This reduction in bias, however, is offset by an increase in variance. Under standard assumptions, the OLS has indeed a larger bias than the TSLS estimator; and moreover, one can prove that the sample variance of the OLS estimator is no greater than the one of the TSLS. Therefore, it is natural to ask whether one could combine the desirable properties of the OLS and TSLS estimators. Such a trade-off can be achieved through a convex combination of these two estimators, thereby producing our proposed convex least squares (CLS) estimator. The relative contribution of the OLS and TSLS estimators is here chosen to minimize a sample estimate of the mean squared error (MSE) of their convex combination. This proportion parameter is proved to be unique, whenever the OLS and TSLS differ in MSEs. Remarkably, we show that this proportion parameter can be estimated from the data, and that the resulting CLS estimator is consistent. We also show how the CLS framework can incorporate other asymptotically unbiased estimators, such as the jackknife IV estimator (JIVE). The finite-sample properties of the CLS estimator are investigated using Monte Carlo simulations, in which we independently vary the amount of confounding and the strength of the instrument. Overall, the CLS estimator is found to outperform the TSLS estimator in terms of MSE. The method is also applied to a classic data set from econometrics, which models the financial return to education.