Judging Contending Estimators by Simulation: Tournaments in Dynamic Panel Data Models

Abstract

An attempt is made to set rules for a fair and fruitful competition between alternative inference methods based on their performance in simulation experiments. This leads to a list of eight methodologic aspirations. Against their background we criticize aspects of many simulation studies that have been used in the past to compare competing estimators for dynamic panel data models. To illustrate particular pitfalls some further Monte Carlo results are produced, obtained from a simulation design inspired by an analysis of the (non-)invariance properties of estimators and occasionally by available higher-order asymptotic results. We focus on the very specific case of alternative implementations of one and two step generalized method of moments (GMM) estimators in homoskedastic stable zero-mean panel AR(1) models with random individual specific effects. We compare a few implementations, including GMM sytem estimators with alternative weight matrices, and illustrate that an impartial evaluation of the outcome of a Monte Carlo based contest requires evidence - both analytical and empirical - on the completeness, orthogonality and relevance of the simulation design.

Keywords: finite sample behavior, generalized method of moments, initial conditions, Monte Carlo methodology, orthogonal parametrizations

JEL Classification: C13, C15, C23