An Examination of the Dynamic Behaviour of Local Governments Using GMM Bootstrapping Methods

SUMMARY
Even though recent Monte Carlo evidence has shown that the use of bootstrap critical values, instead of
asymptotic ones, improves the size of the tests substantially, empirical applications using GMM bootstrap
techniques are largely missing. In this paper, the dynamic relationship between local government revenues
and expenditures is re-investigated using GMM bootstrapping techniques on a panel of 265 Swedish
municipalities over the period 1979-1987. A lag of one year is found in the expenditures equation, while no
dynamics is found in the own-source revenues and grants equations. These results, while contrasting sharply
with those obtained when asymptotic critical values are used, are well in line with the theoretical
explanations given in the literature for dynamic behaviour in the local public sector. Copyright C2000 John
Wiley & Sons, Ltd.

1. INTRODUCTION
Methods for estimating dynamic panel data models have become standard components in the
applied econometrician's toolbox. Unfortunately it has been shown that asymptotic theory may
provide poor approximations to the distributions of test-statistics obtained when Generalized
Method of Moments (GMM) estimation techniques are used. Monte Carlo evidence in e.g.
Brown and Newey (1995), Hall and Horowitz (1996), and Bergstrom, Dahlberg and Johansson
(1999) shows that the Sargan test (i.e. the overidentifying restriction test), which is the most
common specification test used, is oversized in small samples, implying that we tend to reject a
true null too often. A possible remedy, proposed by Brown and Newey and Hall and Horowitz
and further investigated and developed by Bergstrom et ul. is to use bootstrap critical values
instead of asymptotic ones when conducting the tests. This improves the tests' size properties
substantially. Furthermore, Monte Carlo evidence in Bergstrom et al. indicates that GMM
bootstrapping can be successfully used in sequential testing. However, empirical applications
using GMM bootstrap techniques are largely missing.