JamesL.PowellDepartmentofEconomicsUniversityof.pdfVIP

Models, Testing, and Correction of Serial Correlation James L. Powell Department of Economics University of California, Berkeley First-Order Serial Correlation The general problem of serially correlated disturbances in the linear regression model 0 yt = x β+ ε ,t t t= 1,...,T, is to find methods to accommodate lack of independence in the disturbances across observations, i.e., C(ε ,ε ) =6 0 if t =6 s. Unlike the case of heteroskedasticity, we usually assume the error terms are t s stationary, so that Var(y 2 t) = σ is a constant across observations. The usual model for serial correlation y assumes that the errors are first-order autoregressive (or AR(1)), ε = ρε + u , t t−1 t where the “primitive” error terms ut are assumed to be mutually uncorrelated (or, stronger still, i.i.d.) with E(u 2 ) = 0 and V(u ) = σ . By the usual AR(1) calculations, assuming |ρ| 1, we get E(ε ) = 0, t t u t C(ε 2 |t−s| 2 ,ε ) = σ · ρ /(1 −ρ ), so the covariance matrix of the vector of error terms ε is of the form t s u 2 2 σ · Ω≡σ · [ω ], with u u ts |t−s| 2 wts = ρ /(1 −ρ ).