Revisions to Official Data on U.S. GNP: A Multivariate Assessment of Different Vintages
K. D. Patterson and S. M. Heravi
Although there is a substantial literature on revisions to data published by official agencies, relatively little work has been undertaken on multivariate aspects of the data measurement process (DMP) producing different vintages of the GNP variable. This is particularly so for nonstationary time series. With a focus on U.S. real Gross National Product (GNP), we show that a number of interesting questions can be answered within a multivariate framework. Defining a well-behaved DMP as one generating a single stochastic trend in a multiple vintage data set, we can then assess whether this is the case for GNP. We also consider whether the short-run properties of the different vintages share the same dynamic structure. Further, given multiple vintages on the same generic variable, is it the case that one vintage, for example the final vintage, in a well defined sense, dominates the others, and that alone can be used? We show that the idea of single (final) vintage representation is related to the idea that data revisions arise through measurement errors, and contrast this with the interpretation of revisions as forecast errors. Also, the existence of multiple vintages of GNP enables a different approach to the much-researched question of whether GNP has a unit root. This can be formulated as the null hypothesis of trend stationarity in the multivariate Johansen framework. Inter alia we show the importance of the concept of weak exogeneity, and how tests for stationarity of revisions and homogeneity of vintages can be formulated and tested.
Data measurement process, cointegration, common trends, common features, weak exogeneity