When interdependent models are regarded as approximations to recursive models, a specification error is made. What is the importance of that specification error for the maximum likelihood estimation procedure? This question is analyzed for the case where lagged endogenous variables are treated as current on the grounds that the lag is small relative to the observation period. It is argued that the appropriate estimation procedure is qualitatively different from that which ignores the specification error, however small the omitted lag may be. It is contended that as the lag approaches zero the limit form of the likelihood function is not the likelihood of the limit form of the model. The article is, however, exploratory and the analysis lacks complete rigor.