[Paper Review] Optimal Shrinkage Estimation of Fixed Effects in Linear Panel Data Models

Soonwoo Kwon|arXiv (Cornell University)|Aug 24, 2023

Monetary Policy and Economic Impact9 citations

TL;DR

This paper develops a data-driven, distribution-free shrinkage estimator for fixed effects in linear panel data, achieving minimum MSE within a broad class of estimators and handling time-varying, serially correlated fixed effects.

ABSTRACT

Shrinkage methods are frequently used to improve the precision of least squares estimators of fixed effects. However, widely used shrinkage estimators guarantee improved precision only under strong distributional assumptions. I develop an estimator for the fixed effects that obtains the best possible mean squared error within a class of shrinkage estimators. This class includes conventional shrinkage estimators and the optimality does not require distributional assumptions. The estimator has an intuitive form and is easy to implement. Moreover, the fixed effects are allowed to vary with time and to be serially correlated, in which case the shrinkage optimally incorporates the underlying correlation structure. I also provide a method to forecast fixed effects one period ahead in this setting.

Motivation & Objective

• Motivate the need to estimate many fixed effects with limited per-effect data in linear panel models.
• Develop a shrinkage estimator that achieves optimal mean squared error within a flexible class of estimators.
• Allow fixed effects to vary over time and be serially correlated, and incorporate their correlation structure.
• Provide a practical method to forecast fixed effects one period ahead under time-varying settings.

Proposed method

• Start from a multivariate normal means model linking least squares fixed effect estimates to true fixed effects.
• Define a class of shrinkage estimators parameterized by a mean vector mu and a positive semidefinite matrix Lambda.
• Use an unbiased risk estimate (URE) to select hyperparameters by minimizing estimated risk.
• Show that the URE-based estimator dominates or matches common EB shrinkage under mild conditions and without distributional assumptions.
• Introduce three URE estimators (grand mean, general localization, and covariate-assisted) that shrink toward different targets.
• Provide a specialized forecast method for one-period-ahead prediction of fixed effects.

Experimental results

Research questions

• RQ1 Can we construct a shrinkage estimator for fixed effects that minimizes mean squared error without relying on strong distributional assumptions?
• RQ2 How should the shrinkage target and covariance structure be chosen when fixed effects are time-varying and potentially serially correlated?
• RQ3 Does the proposed URE-based estimator achieve asymptotic optimality and perform well in finite samples compared to EB methods?
• RQ4 How can fixed effects be forecast one period ahead in the presence of time variation and serial correlation?
• RQ5 What is the impact of incorporating covariates or broader target locations on estimator performance in practice?

Key findings

• The URE-based estimator minimizes the MSE within a flexible class of shrinkage estimators and is robust to violations of EB distributional assumptions.
• The method accommodates time-varying and serially correlated fixed effects and leverages the correlation structure to improve precision.
• Simulations show URE estimators achieve MSE within 10% of the best possible MSE for sample sizes ≥ 600 and can outperform EB methods when distributional assumptions fail.
• When applied to New York City teacher value-added data, the proposed method materially changes policy-relevant selections, such as bottom 5% teacher releases, compared with conventional methods.
• A forecast variant of the estimator yields about 20% lower average value-added for released teachers in out-of-sample exercises, relative to conventional estimators.