The papers of both Spady and Carlos represent sound and useful discussions of some of the problems that we all face in analysing a complex body of data. Each author has written clearly, has taken care to illustrate the points made, and has warned us of the pitfalls involved in indiscriminate use of the analyses discussed. For this reason, a detailed critical commentary may be less fruitful than a more extended discussion of one or two general points that arise. These can be grouped under two general headings: the preparation of a theoretical model, and the nature of parsimony. The Preparation of a Theoretical Model Carlos places the ball firmly in the theoretician's court, as he stresses that the key decisions are theoretical and not statistical. The researcher who wishes to use causal analysis effectively must ask specific questions of the data. It is not enough to wonder, "Which model fits my data best?" Causal analysis cannot answer that question; as a rule, many different complex models will seem consistent with our data. The traditional justification for a theory, that it works, may not be enough; other models can be found that work equally well. Path analysis puts the burden of inventiveness on the user, at the research design stage when the model is being constructed. One must prepare the theoretical model with the same care and attention to detail that is exercised in designing a questionnaire. One must review potential causal variables and alternative timesequences, and be able to present an effective justification for the model that is finally chosen. If we assume in advance that there will be causal relationships here and here but not there, path analysis will tell us that the strengths of these relationships will be this and this. What the analysis tells us may be so different from our expectations that our model has to be rejected; but if the model is complex rejection is unlikely. The decision to use or avoid a causal model cannot sensibly be made until the theory has been set out. If the researcher cannot confidently make the necessary decisions about a causal model, he should not use one. These choices imply not only the inclusion of certain causal paths and the elimination of others; but also a description of the shape of the relationship wherever there is a causal link. The researcher may lind this very difficult; and may be baffled by discussions of nonlinear and non-additive relationships. In the first section of his paper, Carlos considers the links between theory and procedures of analysis, and