The epistemic benefits of generalization in economic modelling
时间：2018年3月28日 14:00
地点：哲学院317
主讲人：Aki Lehtinen
(University of Helsinki, Academy of Finland Centre of Excellence in the Philosophy of the Social Sciences) graduated from the Erasmus Institute for Philosophy and Economics in 2007. He has written on various topics such as modelling and robustness, philosophy of game theory, social choice theory, and various topics in the philosophy of economics. He has also constructed models on voting, which explains why he has publications in economic, statistical, and political science journals (European Journal of Political Economy, Theory and Decision, Journal of Theoretical Politics) in addition to philosophical ones (British Journal for the Philosophy of Science, Erkenntnis, Philosophy of Science, Philosophical Studies etc.). His most recent work has been on confirmation in climate models and philosophy of macroeconomics, and he is also working on the philosophy of meta-analysis and computer simulations.
ABSTRACT: This paper spells out why generality is an important desideratum in economic modelling. Generalising models serves similar epistemic functions as robustness analysis: it provides a solution to the epistemic uncertainty that arises from the presence of unrealistic assumptions. We present our arguments by discussing examples from economic modelling: the Dixit-Stiglitz (1977) model of monopolistic competition and Abraham Wald’s proofs for the existence of general equilibria.
Proving the same result with more general assumptions is an important achievement for economists. Generality is typically defined as ‘the property of applying widely’ and is measured either by the number of phenomena a model can explain or predict, or by the number of systems to which a model applies. But why is ‘applying widely’ a desideratum?
We will argue that not all kinds of generalizations provide epistemic benefits. Specifically, generalizations have epistemic benefits only when they involve either increases in expressive power or they entail fewer false assumptions about the target. Increasing a model’s generality via either of these routes helps solving some of the problems that arise from the necessity of making unrealistic assumptions. Tractability considerations often imply that modellers describe their targets with assumptions that are less general than they think can be truly asserted about them. Thus they often are able to prove a result only for a special case. They know the specificities introduce falsehoods. Yet they do not know whether the model’s results crucially depend on those falsities. When the model is generalized but continues to imply the same result, we learn that the particular falsehoods were not responsible for the results. That is, obtaining the same result with less restrictive assumptions increases the modellers’ confidence that the result is not an artefact of specific assumptions that are known to be unrealistic. We will thus argue that the importance of generality derives from the same kind of epistemic considerations that motivate derivational robustness analysis (see esp. Kuorikoski, Lehtinen, and Marchionni 2010), and that herein lies its main epistemic advantage.
There are three kinds of generality and corresponding generalizations. First, a model may be generalized such that it applies to more phenomena. Second, the level of abstraction of the target may be increased. Third, a given target may be described with more general assumptions. If we simply count the number of systems, the three cases are indistinguishable – in all the number of systems becomes larger. Let us call them increasing the number of target phenomena (G1), generalizing the target (G2), andincreasing the number of subsumed systems (G3), respectively. We argue that the three kinds of generalizations are very different, and that only G3 provides genuine epistemic benefits.
A characterization of epistemically beneficial generalizations is as follows:
Model M1 provides an epistemically beneficial generalization of model M2 if the model-implications of M1 and M2 include the same generalized (or actual) targets and M1 describes them in such a way that they subsume a larger number of possible systems than M2.
