Backward mean transformation in unit root panel data models

Artūras Juodis,Rutger W Poldermans

doi:10.1016/j.econlet.2021.109780

Artūras Juodis, Rutger W Poldermans

Open Access

https://doi.org/10.1016/j.econlet.2021.109780

Copy DOI

Abstract

The effectiveness of an orthogonal to backward mean transformation is investigated in the context of a non-stationary panel data model. It is shown that the corresponding estimator is as efficient as Transformed Maximum Likelihood when the autoregressive parameter is equal to unity. Furthermore, a recently introduced bias-corrected version is almost as efficient as the Pooled Least Squares estimator.

Highlights

Dynamic panel data models play a prominent role among empirical tools used by applied researchers
As is well-known from Nickell (1981), the conventional Fixed Effects (FE) estimator suffers from a sizeable finite sample bias for small values of T
We show that in a model with the autoregressive parameter equal to unity the estimator of Everaert (2013) is as efficient as the Transformed Maximum Likelihood (TML) estimator studied by Kruiniger (2008)

Summary

Introduction

Dynamic panel data models play a prominent role among empirical tools used by applied researchers. We investigate the Least Squares (LS) estimators of Choi et al (2010) and Everaert (2013). Both papers, among other things, establish that the corresponding estimators are nearly (asymptotically) unbiased under stationarity, while they are asymptotically unbiased in case of a unit root. Neither of the two studies investigate the asymptotic variance in a non-stationary unit-root setup. We show that in a model with the autoregressive parameter equal to unity the estimator of Everaert (2013) is as efficient as the Transformed Maximum Likelihood (TML) estimator studied by Kruiniger (2008). The backward-means based estimators outperform the conventional (bias-corrected) FE estimator when data are persistent

The model

Main results

Concluding remarks