The paper focus on the VAR approach to modelling nonstationary panel data. Two statistical frameworks are compared, the large-scale VAR model of IP-dimensional process, as proposed by Larsson and Lyhagen (2007), and the global VAR model, as proposed by Pesaran, Schuerman and Weiner (2004). The study investigates the small-sample performance of the ML estimator of the long-run parameters in case of cross-sectional adjustments to the cointegrating vectors.