In recent years, atheoretical econometrics has been in vogue, exploiting natural variations and clever instruments to identify the effects of policies. This is in contrast to structural models, which rather than “letting the data speak for themselves” instead attempt to identify the effects of policies by describing a theoretic model of some process and making clear identifying restrictions in order to identify parameters of interest. Keane makes three particularly interesting points. First, “letting the data speak” is not always what it claims to be: even simple instrumental variable regressions are making assumptions (for instance, exogeneity assumptions in fixed effects estimators) that often, on careful investigation, turn out to be very similar to the type of more clearly stated identifying assumptions in a structural model. Second, data cannot often speak for itself. An estimate of the intertemporal elasticity of money based on wage data alone depends very much on assumptions about how individuals learn on the job; in particular, learning-by-doing models will drastically increase the intertemporal substitution. Third, structural models help with induction, or external validity. An atheoretic model is a model of history. Without some explanation of why an effect was seen – something that comes from theory – is it not clear how to apply the data from one natural experiment to a future policy decision. Keane suggests that arguments for “letting the data speak for themselves” draw on a mistaken understanding of scientific progress where empirical data guides theory; as Kuhn and others have noted, it is almost always the theory that comes first, for without a theory suggesting some relationship, why would one bother to collect and investigate a certain set of data at all?
http://gemini.econ.umd.edu/jrust/research/JE_Keynote_7.pdf (WP – full version in J. of Econometrics 196, with responses by Blundell and others)