← Feed
@ctesta@fediscience.org
Post #2594107
2026-05-06 14:50 UTC
I’ve been trying to read more carefully about instrumental variables and make up my mind about when IV arguments are scientifically convincing.
Here's a tension I keep running into:
Should the scientific question alone determine the causal parameter of interest?
Or is it legitimate for the target parameter to reflect an interplay between scientific interest and the identifying assumptions we actually find tenable?
IVs can be difficult to interpret when instruments are weak, who “compliers” are is opaque, exclusion restrictions are debatable, or linear models are used in settings where the true data-generating process may be nonlinear.
On the other hand, when an entire body of (aspirationally causal) literature rests on methods that try to close backdoor paths, IVs offer a genuinely different identification strategy. That seems valuable for evidence triangulation, even if IV analyses have their criticisms.
What do you think? Are you a big IV proponent? Are you an IV critic?
When do you find IV evidence persuasive?
Some literature I've been reading & re-reading:
https://pubmed.ncbi.nlm.nih.gov/16755261/
https://academic.oup.com/ije/article/47/4/1289/3095892
https://pmc.ncbi.nlm.nih.gov/articles/PMC4285626/
https://arxiv.org/abs/2402.09332
https://arxiv.org/abs/2402.05639
#CausalInference #InstrumentalVariables #Econometrics #Statistics #DataScience #HealthPolicy
Replies (1)
-
Some choice quotes on the topic of instrumental variables:
From Swanson and Hernán 2014
> Imbens states we are “limited in the questions we can answer credibly and precisely.” We agree, but there are differences between the questions we can answer and the questions we want answered. Choosing only answerable questions (e.g., identifying the LATE in some settings) can mislead decision-making efforts: our estimates may be misinterpreted as directly relevant to a decision when in fact they are only tangentially related.
From Levis, Kennedy, and Keele 2024
> Often in the literature, it is noted that the choice of estimand should be based
on a scientific judgement and should reflect the specific research question. However, it is
often the case that there is some interplay between the estimand and the plausibility of the identification assumptions. Nowhere is this more true than in an IV analysis.
What do you think? Do you have a favorite instrumental variable?
Tell me it's not rainfall!
https://www.jonathanmellon.com/publication/rain/
Open ##3040182