Basic Meta Aspects

  • Economics is not philosophy. Economics looks at data, develops models, looks at more data, offers different models, etc. There is no clean distinction between theory and evidence. The idea that a theory is polluted because the researcher has looked at data first, or vice-versa, is common but ludicrous.

  • Economic theories are like lenses. Pick too narrow or tinted a lens, and you are not really seeing what is going on or at best a distorted view. Pick no lens and you cannot see or focus on anything.

  • Everything researchers do has or needs a model — even if it is as simple as an econometric model. The use of a proxy measure is based on / implies a model. The question is how much to lean on the model and/or how rigorous it is supposed to be.

  • If there is no clear and good alternative to the tested theory, it is not a reasonable test. Think about the lens analogy. Under the model, of course the model is true. silly. what happens if we don’t know for sure whether it is the right model? what alternatives and strawmen can our tests capture? (calibration is not testing.)

  • Theories that are always true or where evidence in the paper’s situation always has external validity everywhere else are not theories but tautologies. (Cases where there is no external validity anywhere are interesting only if the context itself is interesting.) The point of a theory is to offer hypotheses about how local phenomena exists when there are/were alternative explanations.

  • All theories reject if the power is strong enough. Models are, after all, just models, and try to strip reality down to the most important aspects. You can assess how badly a theory is rejected. For example, it is pretty clear that stock prices are not perfect random walks, but it is also clear that the deviations are trivially small.

  • If a theory predicts A, B, and C; and A and C work but B is rejected, the theory is false. (This is more a confusion in the context of asset-pricing theory.)

  • In my mind, we use techniques to answer economic questions. This means questions first, techniques later. In many other minds, the point is about improving methods first (and perhaps showing off ability, especially on the rookie market).


  • This means distinguishing whether X is merely associated with Y, or whether X causally contributes to determining Y.

  • If Y = X*Z, X’s causal influence depends on Z, too. For firms where Z is zero, there is no causal influence. The point is that causal influence and identification is not a zero or one aspect, either. X can also be partly exogenous and partly endogenous.

  • If X -> Y, then dX -> dY. Otherwise, correlation between X and Y is spurious. (If X !-> Y, then dX -> dY is not.)

    • this assumes that there is enough statistical power
    • power is the probability of rejecting the null hypothesis if the alternative hypothesis is true.
  • The 2021 Nobel Prize was for identification. You should probably have read Angrist-Pischke, already, too. In order of believability (strength):

    1. experiments (assignment of X, observation of Y)
    2. regression discontinuities
    3. IV designs from truly random sources of variation
    4. diff-in-diff designs (the source of variation is both time and space)
    5. diff design (the source of variation is either time or space)
    6. level association design (i.e., just level correlations between variables)
  • Causation causes correlation. Correlation does not need causation. If no correlation is observed, then presumably there is no causation.

    • the point is that correlation evidence is not uninformative.
    • correlation evidence is just not very strong. It suffers from more Type-1 error.
  • Spurious correlation = other factor responsible, such as selection.

  • All models make some identification assumptions. Even defining a proxy in a reduced-form test is a method of theoretical identification. Structural models just lean more strongly on theoretical identification (such as particular functional associations) than reduced-form model tests.

  • Identification has become central to empirical economics. In some cases, researchers have hammers (quasi-experiments) and are looking for nails (y variables).

/teaching/phd-empcf/meta-aspects.html Last modified: