This article is concerned with the specification error implied by the use of a non-recursive model as a discrete approximation to a system of stochastic differential equations. One of the questions considered is whether, in view of this specification error, it is better, for the purpose of prediction, to ignore the over identifying restrictions and estimate the reduced form parameters directly. Another is the question of how to make the best use of the estimates of the parameters of the approximate model in order to obtain estimates of the parameters of the exact model.