JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.. American Statistical Association is collaborating with JSTOR to digitize, preserve and extend access to Journal of the American Statistical Association.For an m-variate ("partially") nonstationary vector autoregressive process {Y}, we consider the autoregressive model ID(L)Y, = ,, where P(L) = I -,L --IPLP and det{D(z)} = 0 has d < m roots equal to unity and all other roots are outside the unit circle. It is also assumed that ranok{?(1)} = r, r = md, so that each component of the first differences W, = Y, -Y,1 is stationary. The relation of the model to error correction models and co-integration (Engle and Granger 1987) is discussed. The process {Y,} has the error correction representation I*(L)(1 -L)Y, = -P2(Ir -Ar)Q2Y,-+ c, where Q(I,, -ID(1))P = J = diag(Id, Ar) is in Jordan canonical form and Q' = [Q, Q2]. It follows that the transformation Z, = QY, = [Z Z2,]' is such that the d x 1 process Z,, is nonstationary with Z, -Z,, stationary while Z2, is stationary. Asymptotic distribution theory for least squares parameter estimators of the model is first considered. A Gaussian partial reduced rank estimation procedure that explicitly incorporates the unit root structure in the model is then presented, and an asymptotically equivalent two-step reduced rank estimation procedure is also considered. A numerical example is presented to illustrate the methods and concepts. The finite sample properties of the estimators are also briefly examined through a small Monte Carlo sampling experiment.