The paper discusses a stubborn problem of theory construction: deciding between recursive and non- recursive variants of a causal model by testing them against empirical data. After pointing out the consequences of a correlation between an exogeneous variable and the error term of an endogeneous variable as well as certain aspects of the identification problem we show for the asymptotic case (n → ∞) that a test is possible if the correlation between the exogeneous and the error term is in fact zero. Following this we present the results of a Monte-Carlo simulation investigating the robustness of the proposed test when applied to small samples. Finally we suggest conclusions of this testing procedure for empirical research.