Thoughts on Rational Inference or Is Information Theoretic Inference Rational?
Abstract: Reasoning, modeling, and inference given imperfect, limited or incomplete information remains a major challenge, especially in the social, behavioral, and other complex sciences. We need a coherent, decision-theoretic, ‘rational’ framework on both the theoretical and empirical fronts. In my talk, I will discuss some thoughts on rational inference. I will touch on several questions including, how to define and implement ‘rational’ approaches to inference (and is it possible to do so)? How can ‘rational’ methods to inference handle possible misspecifications? This includes issues of model uncertainty (ambiguity), robustness, and sensitivity analysis, as well as falsification.
Some background: we are all trying to add to knowledge (find the ‘truth’) and solve problems from observed information. But the available information is (in most cases) insufficient to pin down a unique representation or solution. In fact, we are generally faced with a continuum of possible solutions – theories or models – that are consistent with our observed information (usually, in terms of circumstantial evidence). This problem is universal across all disciplines and decision makers. An inference that is rational may provide a way to circumvent that problem. Though, the dialogue on ‘rational’ inference has its roots in the work of Neyman, Pearson and Fisher, and in the debate (and controversy) between the first two and Fisher, there is much more to do. In my talk I will try to go beyond that initial discussion earlier in the previous century.