Doing Economics

2.5 Results and Discussion
The section of an applied economics article that discusses the paper’s
findings is obviously the most important section of the paper. Somewhat
paradoxically, it is perhaps also the least-read section of a paper: after a
reader has read the title, the abstract, the introduction, looked at a few
tables, and maybe looked at the Empirical Framework section to answer
any lingering questions, your reader knows whether she can trust you and
your findings, and she is often only interested in your core finding. Only
reviewers and critical readers (e.g., graduate students reading your paper for
a class, should your article end up on someone’s syllabus, or for their
dissertation) will read the entirety of the results section. Nevertheless,
results sections have their own structure, which is discussed below.

---

2.5.1 Order of Results
There is a certain logical order in which results should be presented.
Typically, results progress from most parsimonious (e.g., a simple, bivariate
regression of y on D) to least parsimonious (i.e., a regression of y on D and
a full set of control variables x). With experimental variation in D, this is
not as useful as with observational variation in D. In the former case,
adding controls on the right-hand side of the equation of interest will in
principle not change the sign and the magnitude of the estimated treatment
effect. Rather, it will only make the estimate of the treatment effect more
precise (i.e., it will reduce the standard error around it).
8
In the latter case, where one cannot assume that E(y|x) = E(y|do(x)), the
most-to-least-parsimonious approach is one first step toward assessing the
robustness of one’s results: if the sign and the magnitude do not change
much or at all as one adds in control variables on the right-hand side, this
suggests that one’s results are already somewhat robust. This is in the spirit
of Altonji et al.’s (2005) approach to robustness (although Oster 2019
critiques Altonji et al. 2005 and suggests a new method aimed at assessing
how important unobserved heterogeneity is in a given application).