ABSTRACT

This paper presents the results of a Monte Carlo study that compares the small sample performance of various unit root tests in short panels using simulated data that mimic the time series and cross sectional properties of commonly used firm level variables. Our conclusion is that in the presence of firm-level heteroskedasticity two methods are preferred, depending on the nature of the preferred alternative: the simplest method based on the ordinary least squares regression of the variable under consideration on its own lag and a version with a more complex alternative hypothesis suggested by Im, Pesaran, and Shin. The paper also reports the results of using these tests for sales, employment, investment, R&D, and cash-flow in three panels of large French, Japanese and US manufacturing firms. In most cases our data reject the presence of a unit root in favor of a first order autoregressive model with a very high autoregressive coefficient, so high that fixed effects are of negligible additional importance in the model.

INTRODUCTION

In this paper, we investigate the properties of several unit root tests in short panel data models using simulated data that look like the data typically encountered in studies on firm behavior. This investigation arose from a previous exploration of a simple question – could we find a simple parsimonious model that accounts for the time series properties of key observable variables characterizing the behavior of individual firms: sales, employment, investment, R&D, and cash flow or profits in France, Japan and the United States.