A standard assumption in structural equation models with interaction terms is the normality of all the random constituents of the model. In applications, however, there may be predictors, disturbance terms of equations, or measurement errors, deviating from normality. The present paper investigates how deviation from normality affects the consistency of two alternative estimators; namely, the maximum likelihood (ML) and the method of moments (MM) estimators. The ML approach requires full specification of the distribution of observable variables while this is not required in the MM approach. It will be seen that while the MM estimator is insensitive to departures from normality of all the random constituents of the model not involved in the interaction terms, such deviation from normality distorts considerably the consistency property of the ML estimator. The paper provides analytical results showing the consistency of MM when using a proper selection of moments up to order three, and presents a Monte Carlo illustration showing how the consistency of the ML estimator breaks down when there is deviation from normality. It is concluded that for a variety of distributions of the data, the MM method gives consistent estimates while ML does not.