VECTOR DISTRIBUTED LAG MODELS WITH SMOOTHNESS PRIORS

Distributed lag models with smoothness priors have been found useful in econometrics since they were introduced by Shiller [12] but the estimation of the smoothness parameters is still an open problem. The assumption of smooth distributed lag or regression coefficients requires the estimation of variance (hyper-) parameters and the order of the lag distribution. Akaike [1] suggested for univariate distributed lag models an estimation procedure which maximizes the conditional likelihood function of the hyperparameters. This empirical Bayes method is extended to the multivariate case for models with simple covariance structures. For complicated models a fully Bayesian approach based on the posterior-imputation (PI) algorithm of Tanner and Wong [14] is suggested. © 1990.