Doing Economics

2.4.2 Identification Strategy
After showing and discussing what equations are estimated, there needs to
be a discussion of how the coefficient pertaining to the causal relationship
of interest is identified.
The term “identification” has gone through several meanings over time
(Lewbel 2019). For better or for worse, the term more often than not refers
to causal identification nowadays in applied papers. What is causal
identification? Briefly, it refers to situations where a coefficient is more
than just a (partial) correlation between the dependent variable y and some
variable of interest D, and where the estimated coefficient instead reflects a
relationship from cause D to effect y.
Although an unbiased coefficient estimate implies an identified—that is,
causally identified—coefficient estimate, the converse is not true. There are
situations where one knows a coefficient to be biased, but where a
statistically significant coefficient estimate can still be used to denote a
causal relationship.
If you are fortunate enough (i) to have experimental variation in your
treatment variable, and (ii) balance tests suggest the experimental
assignment of observations to treatment and control groups was truly
random, your identification strategy section can be kept short, as your
results are causally identified by virtue of experimental assignment. In other
words, you can estimate what Pearl (2009) denotes E(y|do(x)); that is, the
(causal) effect of treatment x on outcome y.
If you have (i) experimental variation in your treatment variable but (ii)
balance tests suggest the experimental assignment of observations to
treatment and control groups was not truly random, your identification
strategy section can also be short, as you only need to explain how you will

---

add in control x on the right-hand side of your equation of interest to help
rectify the situation, but only somewhat, as unobservables are also likely to
be unbalanced when the observables are unbalanced.
If you do not have experimental variation in your treatment variable,
there is yet more work to be done. This chapter cannot dive into causal
identification with observational data, but there are nevertheless certain
things that can be discussed as being necessary in any good identification
strategy section:
• Explain intuitively why your results have a shot at causal
identification. Practically speaking, this means that you have to
tell your reader why your results bring us closer than ever before
to making a causal statement about the relationship of interest. In
the best-case scenario, this will be because you have a research
design (e.g., a strictly exogenous instrumental variable such as a
lottery) which clearly allows thinking of treatment as if it were
randomly assigned. In less-than-ideal scenarios (e.g., an
instrumental variable that is only plausibly exogenous; cf. Conley
et al. 2012), you need to explain why, even though your research
design does not yield clean and clear causal identification, your
results are the best in the literature.
6
• Discuss in turn the three following sources of statistical
endogeneity:
7
(i) reverse causality, (ii) unobserved heterogeneity,
and (iii) measurement error, explaining whether each of those
sources of statistical endogeneity is a concern in your application,
and how it is dealt with in your application. Here, if there are
issues, admit to them, and explain how they might bias your
estimate of the coefficient of interest. Be honest about what your
paper can and cannot do.
• Once that is done, there is one more threat to internal validity to be
considered, namely violations of the stable unit treatment value
assumption (SUTVA). What SUTVA means is specific to each
application, but in short, if you observe the effect of a treatment
D
it
on outcome y
it
, where i denotes an individual unit of
observation and t denotes a time period, it has to be the case that
the value of D
it
does not affect the value y
−it
, y
i,−t
, or y
−i,−t
. In

---

other words, there cannot be any spillovers from one unit being
treated to another unit’s outcome, and there cannot be any
spillovers from one unit being treated at a given point in time to
that same unit’s outcome in the future, nor can there be any
spillovers from one unit being treated at a given point in time to
other unit’s outcome in the present or in the future. The SUTVA
can be extremely difficult to satisfy. That said, one can often test
for SUTVA violations; see Burke et al. (2019) for an example of a
paper where the authors deal with SUTVA violations very well.
• Again, because this is important: if your results are not causally
identified, be honest about what they can and cannot do. And
generally, do not make claims that are not backed up by your
research designs of your results, no matter how much you wish
those claims to be true. Editors and reviewers would much rather
deal with manuscripts wherein the author candidly admits to the
limitations of their findings than with manuscripts wherein the
author tries to deceive the reader. In plain English: the former
kind of manuscript has a much better chance of not being rejected
than the latter.