On the dynamic shape of aggregated error correction models (Q1106605)

The paper investigates the conditions under which ARMAX microequations can be aggregated into macroequations (defined in terms of aggregates which are simple sums or weighted averages of the microvariables) having the same form. Three cases of increasing difficulty are treated. The simplest one occurs whenever the microparameters, or the right-hand variables, are equal across agents. A case of intermediate difficulty arises when the microparameters or microvariables are unequal, but the macrovariables are stationary. In these two cases the existence and uniqueness of the macrorepresentation can be shown. The main part of the paper is devoted to the investigation of the more difficult case where the macrovariables and microvariables are nonstationay. In this case the macrovariables can either be cointegrated [in the sense of \textit{R. F. Engle} and \textit{C. W. J. Granger}, Econometrica 55, 251-276 (1987; Zbl 0613.42140)] or not cointegrated. In both cases an ARMAX macrorelation can be written in the first differences of the macrovariables. But when the macrovariables are not cointegrated, an ARMAX macrorepresentation in levels does not exist. When the macrovariables are cointegrated, an ARMAX macrorepresentation in levels can be written under plausible assumptions on the microparameters. It is then shown that in the cointegrated case, even very simple (e.g. static) micromodels can have quite complicated implications on the dynamics of the corresponding macromodels. The reason is that the structure of the macromodel cannot, in general, be simplified into a rational distributed lag between levels with a white-noise disturbance, or into a finite distributed lag between levels with an ARMA disturbance term. The implications for tests of the rational expectations hypothesis on aggregate data are briefly discussed at the end of the paper.