The limits of extrapolation in economics

from Lars Syll

There are two basic challenges that confront any account of extrapolation that seeks to resolve the shortcomings of simple induction. One challenge, which I call extrapolator’s circle, arises from the fact that extrapolation is worthwhile only when there are important limitations on what one can learn about the target by studying it directly. The challenge, then, is to explain how the suitability of the model as a basis for extrapolation can be established given only limited, partial information about the target … The second challenge is a direct consequence of the heterogeneity of populations studied in biology and social sciences. Because of this heterogeneity, it is inevitable there will be causally relevant differences between the model and the target population.

In economics — as a rule — we can’t experiment on the real-world target directly.  To experiment, economists therefore standardly construct ‘surrogate’ models and perform ‘experiments’ on them. To be of interest to us, these surrogate models have to be shown to be relevantly ‘similar’ to the real-world target, so that knowledge from the model can be exported to the real-world target. The fundamental problem highlighted by Steel is that this ‘bridging’ is deeply problematic​ — to show that what is true of the model is also true of the real-world target, we have to know what is true of the target, but to know what is true of the target we have to know that we have a good model  …

Most models in science are representations of something else. Models “stand for” or “depict” specific parts of a “target system” (usually the real world). A model that has neither surface nor deep resemblance to important characteristics of real economies ought to be treated with prima facie suspicion. How could we possibly learn about the real world if there are no parts or aspects of the model that have relevant and important counterparts in the real world target system? The burden of proof lays on the theoretical economists thinking they have contributed anything of scientific relevance without even hinting at any bridge enabling us to traverse from model to reality. All theories and models have to use sign vehicles to convey some kind of content that may be used for saying something of the target system. But purpose-built tractability assumptions — like, e. g., invariance, additivity, faithfulness, modularity, common knowledge, etc., etc. — made solely to secure a way of reaching deductively validated results in mathematical models, are of little value if they cannot be validated outside of the model.

All empirical sciences use simplifying or unrealistic assumptions in their modeling activities. That is (no longer) the issue – as long as the assumptions made are not unrealistic in the wrong way or for the wrong reasons.

Theories are difficult to directly confront with reality. Economists therefore build models of their theories. Those models are representations that are directly examined and manipulated to indirectly say something about the target systems.

There are economic methodologists and philosophers that argue for a less demanding view on modeling and theorizing in economics. And to some theoretical economists it is deemed quite enough to consider economics as a mere “conceptual activity” where the model is not so much seen as an abstraction from reality, but rather a kind of “parallel reality”. By considering models as such constructions, the economist distances the model from the intended target, only demanding the models to be credible, thereby enabling him to make inductive inferences to the target systems.

But what gives license to this leap of faith, this “inductive inference”? Within-model inferences in formal-axiomatic models are usually deductive, but that does not come with a warrant of reliability for inferring conclusions about specific target systems. Since all models in a strict sense are false (necessarily building in part on false assumptions) deductive validity cannot guarantee epistemic truth about the target system. To argue otherwise would surely be an untenable overestimation of the epistemic reach of surrogate models.

Models do not only face theory. They also have to look to the world. But being able to model a credible world, a world that somehow could be considered real or similar to the real world, is not the same as investigating the real world. Even though all theories are false, since they simplify, they may still possibly serve our pursuit of truth. But then they cannot be unrealistic or false in any way. The falsehood or unrealisticness has to be qualified (in terms of resemblance, relevance etc). At the very least, the minimalist demand on models in terms of credibility has to give away to a stronger epistemic demand of appropriate similarity and plausibility. One could of course also ask for a sensitivity or robustness analysis, but the credible world, even after having tested it for sensitivity and robustness, can still be a far way from reality – and unfortunately often in ways we know are important. Robustness of claims in a model does not per se give a warrant for exporting the claims to real world target systems.

Questions of external validity — the claims the extrapolation inference is supposed to deliver — are important. It can never be enough that models somehow are regarded as internally consistent. One always also has to pose questions of consistency with the data. Internal consistency without external validity is worth nothing.