Doing Economics

2.5.4 Mechanisms
As a result of the Credibility Revolution (Angrist and Pischke 2010),
applied microeconomists have been answering questions of the form “Does
D cause y?” or “What is the effect of D on y?” first and foremost.
In recent years, however, much has been written in the quantitative social
science literature about how to test for whether a given variable m is a
mechanism whereby some other variable D causes some outcome y—what
is called mediation analysis—and this remains a very active area of
research.
9
A good section on mechanisms does its best to investigate
potential mechanisms. In the best-case scenario, this involves a proper
mediation analysis. In many cases, this means doing what one can do with
the data at hand, such as presenting descriptive (i.e., not causally identified)
regressions or correlations. In other cases, this means simply admitting that
there are some mechanisms one cannot test for, not even with imperfect
proxies. When anything but the ideal is feasible, you should clearly explain
why you cannot test for specific mechanisms to leave no doubt in your
readers’ minds that you have thought about the question “How does D
cause y?”
2.5.5 Limitations
A good Empirical Results section should be honest about what it can and
cannot do. Though this is often discussed quickly in the Conclusion, it
should be discussed more fully in a separate subsection of the estimation
results section.
What limits one’s results? Typically, limitations come in three varieties.
First and foremost, internal validity may be limited. In other words, one
might not be able to make a causal statement but instead only get close to
doing so relative to the literature. For instance, your instrumental variable
might only be plausibly exogenous, but not strictly so. This would be a
good time to remind the reader that this is so.
Second, external validity may be limited as well. This is often the case
with lab or a lab-in-the-field experiments,
10
 or with RCTs. Or you may have
a strictly exogenous instrumental variable, but it is not entirely clear who

---

the compliers and defiers are, and so who the local average treatment effect
applies to is a nebulous subset of the sample.
Finally, the variables you use as your treatment or your outcome variable
might only be proxies for what you are truly interested in. For instance,
though you may be interested in looking at whether economic shocks push
people to commit suicide, data on suicides may not be available (or suicides
may be significantly under-reported), and so you might have to resort to
using mortality rates instead.