# Difference between an “Empirical Strategy” and an “Identification Strategy” in econometrics?

What are the substantive differences between these terms, specifically for economists?

Are instrumental variables, difference-in-differences, and regression discontinuity designs considered ‘identification strategies?’

For example, if I am using difference-in-differences to obtain a causal effect, and someone asked me ‘what is your identification strategy?’ would ‘I used a difference-in-differences design’ be the appropriate response?

• That is economics-speak for methods of using observational data to identify effects when a proper experiment is not possible. Here are some lecture notes on identification strategies which might help you understand some of the ways the phrase is used. May 18, 2021 at 22:51

Both terms are fairly loaded terms with meanings that will depend on who is using it.

Broadly speaking, "empirical strategy" is an umbrella term used by researchers to indicate their overall "process" in approaching a question and delivering an answer. Indeed, Angrist and Krueger write in Empirical Strategies in Labor Economics (my own highlighting):

We use the term empirical strategy broadly, beginning with the statement of a causal question, and extending to identification strategies and econometric methods, selection of data sources, measurement issues, and sensitivity tests.

This re-affirms the idea that empirical strategy is a catch-all term to indicate your overall method.

In contrast, I would argue "identification strategy" means something very specific. I wrote a bit about identification in this CV question. Based on my definition of identification presented there, I would define identification strategy as the process of

1. defining a parameter you are interested in (such as the causal effect of treatment on outcome), and
2. proving that your observed data (a DGP) and imposed assumptions (such as parallel trends in diff in diffs) identify this parameter

In applied work, you'll probably notice people are not so rigorous about this, and instead will say something hand-wavy such as "our identification strategy leverages a difference in difference design..." This is fine, because in many commonly used designs, previous work has already done the tedious work of showing the formalities, so applied research can simply state that they are doing a diff in diff without having to explain every detail. But in works where the identification strategy is novel, then they have to actually prove the identification strategy "works".

• Thank you for this clarification! Taking the identification strategy process definition, then is it fair to say for instance difference in differences is an identification strategy because it is a way to make comparisons with observed data that uncover (or identify) the causal parameter of interest, if the underlying assumptions are valid? is that at least a slightly less hand waivy way of saying it May 18, 2021 at 23:51
• @Steve yeah, that's right (in the hand wavy sense :-) ). When you say your identification strategy is a difference in difference design, under the hood you're making all the standard assumptions that are required. For example, if you're also identifying the effect using a linear regression design, then you're probably also assuming standard gauss markov assumptions. May 19, 2021 at 1:01