Jon -- My comments on your very interesting argument under each step:
Jon Williamson wrote: > > 1. Structural Equation Models (involving equations of the form E=f(C,e) > where E is an effect, f is a function, C are the direct causes of E and e is > an error term) are used in econometrics and the social sciences for > forecasting and as decision support tools. Agreed. > 2. These models tend to make very strong assumptions. E.g. all relevant > variables have been identified; the qualitative causal structure has been > correctly identified; variables only depend on those variable that are > perceived to be their direct causes; f is linear; the error terms are > independent and normally distributed. Agreed. > 3. Consequently these models are rarely reliable or robust. I agree that "these models are rarely reliable or robust" However, I don't think this property is necessarily a consequence of the previous two claims. Rather it is a consequence of the domain of application and the relationship between that domain and the model. Robustness and reliability are properties of models relative to some domain. A model may be very robust if the domain it seeks to model is continuously dependent upon its initial conditions and the model's assumptions are aligned with these, but not at all robust if either of these two conditions is not met. It has long been my opinion that economies are not continuously dependent upon their initial conditions -- i.e. a small change in initial conditions can result in major differences downstream. > 4. Consequently it would be better to use causal Bayesian networks instead, > since in that case we only need to estimate the probability distribution of > each variable conditional on its direct causes. Such models should be > checked (one should check that the causal Markov condition holds for the > model and that the model is robust for forcasting and modeling > interventions) and refined as necessary. Bayesian networks using probability distributions will still be making strong assumptions, although weaker than structural equation models, as you say. Other types of causal networks, such as QPNs, would likely be more robust, although the conclusions may also be weaker. > 5. Only when we are confident that a causal Bayesian network has captured > the correct qualitative causal structure would it be fruitful to investigate > functional relationships between cause and effect. > 6. Slogan: "causal net before functional model". Not sure I agree. How will causal nets deal with dynamic domains -- those where the relationships between variable are changing -- or with what one could call reflective domains -- i.e. those where the actions taken by a decision-maker can potentially alter the relationships between variables, as would seem to be the case for monetary policy decisions? Does anyone know if such issues have been considered in the causal net literature? -- Peter
