Estimation and inference in short panel vector autoregressions with unit roots and cointegration

This paper considers estimation and inference in panel vector autoregressions (PVARs) with fixed effects when the time dimension of the panel is finite, and the cross-sectional dimension is large. A Maximum Likelihood (ML) estimator based on a transformed likelihood function is proposed and shown to be consistent and asymptotically normally distributed irrespective of the unit root and cointegrating properties of the underlying PVAR model. The transformed likelihood framework is also used to derive unit root and cointegration tests in panels with short time dimension; these tests have the attractive feature that they are based on standard chi-square and normal distributed statistics. Examining Generalized Method of Moments (GMM) estimation as an alternative to our proposed ML estimator, it is shown that conventional GMM estimators based on standard orthogonality conditons break down if the underlying time series contain unit roots. Also, the implementation of extended GMM estimators making use of variants of homoskedasticity and stationarity restrictions as suggested in the literature in a univariate context is subject to difficulties. Monte Carlo evidence is adduced suggesting that the ML estimator and parameter hypothesis and cointegration tests based on it perform well in small sample; this is in marked contrast to the small sample performance of the GMM estimators.