Causal inference and intuitive representations of the world

Just a short follow up on the previous piece. I ended up mentioning that while in structural estimation the story of the causal mechanism is told inside the formal-empirical model, the Rubin-Holland framework only deals with “statistical” causality, that is, the causality between two observed variables, leaving to the reader to figure out the causal mechanism. Interestingly, there seems to be a consensus about how much more intuitive the experimentalist framework is.

In a sense, this is understandable: at least at the technical level, the art of reduced-form econometrics relies consistently more on clever experimental designs and or on finding how natural contexts could replicate experimental conditions, and less on the technicalities proper to DCDM. The set of statistical and computing tools is typically more accessible, if only because explicit formal modeling is often omitted. Moreover, the causal model is coupled with a narrative that emphasizes the analogy with the natural and experimental sciences which seems to be easy to understand and explain to the public, comforting the lack of an explicit causal mechanism. While in formal models the mechanism is explicit, they are typically designed to deal with complex interactions and extract counterintuitive hypotheses; by contrast, empiricists claim, and are to some extent, agnostics or even skeptical about (formal) theoretical constructs.

Yet, the causal mechanism is what seems to matter. At least if you are in the business of knowing not just what caused the effect but also the how it was caused. This is often thought to be important so that knowledge accumulates as a body of theory and note just new estimates and, to raise awareness about how it can be extrapolated to different contexts. To quote  one of my favorite books on the methodology of social science:

To explain an event is to give an account of why it happened. Usually… this takes the form of citing an earlier event as the cause of the event we want to explain…. [But] to cite the cause is not enough: the causal mechanism must also be provided, or at least suggested.

Typically, reduced form methods identify the explanans -and Rubin identification is precisely about that. However, the mechanism is left under the carpet. It may seem puzzling that the essence of the identification exercise, which is a purely technical/statistical exercise, is seen as more intuitive. Arguably, this has to do with the fact that what makes a piece of research successful are not its technical details. Actually, to put some flesh in the bones of the analysis, the discussion of the mechanism is typically performed in an informal fashion, relying on some historial evidence, some speculative story of what might be going on, an enormous amount of handwaving and back of the envelope calculation and the deep statistical analysis of the causal relation of the model is offered as a clue of the plausibility of the story. The pseudo experimental designs are typically imaginative and sexy. Indeed one sometimes feels that these informal speculative discussion and the creativity of the IV or experimental design has at least as much to do with the success of a piece of research as it analytical, technical, inferential issues are.

Don’t get me wrong. I am indeed sympathetic about the role of speculative thinking and informal discussion in scientific discovery.  However, I can hardly see how depriving oneself from the tools of formal models to perform signifies progress, since these are precisely a sophisticated form of doing the same task of refining intuition.