An economist wants to estimate the causal effect of years of education on adult earnings, but a direct regression of earnings on education is confounded by omitted variables like family background, innate ability, and motivation, all of which plausibly affect both how much education someone gets and how much they earn independent of education itself. Randomly assigning years of education to people isn't feasible. Instrumental variables analysis offers a way to approximate what a randomized experiment would provide, by finding some other variable — an instrument — that moves education in a way unrelated to those confounding factors.
What a valid instrument actually needs to satisfy
An instrument must satisfy two core conditions: it needs to be genuinely correlated with the variable of interest (here, years of education), and it needs to affect the outcome (earnings) only through its effect on that variable, with no other separate, direct pathway to the outcome. The first condition, called relevance, is directly testable with the data — you can check whether the proposed instrument actually predicts the variable of interest. The second condition, called the exclusion restriction, cannot be directly tested with the data itself and has to be argued for on theoretical and contextual grounds, which is exactly why finding a genuinely convincing instrument is often the hardest and most debated part of this kind of analysis.
A well-known illustration of this logic in practice
Researchers studying the returns to education have used distance from a person's childhood home to the nearest college as an instrument for years of education, reasoning that living closer to a college plausibly increases the likelihood of attending, and plausibly has no direct, separate effect on adult earnings except through its effect on educational attainment — meaning proximity to a college provides a source of variation in education that's plausibly unrelated to family background, ability, or motivation, the confounders that make a direct regression of earnings on education unreliable.
Why the exclusion restriction is the genuinely contestable part
Whether growing up near a college truly has no other plausible pathway to affecting adult earnings — through, say, a systematically different local economy, different job opportunities, or different community characteristics correlated with college proximity — is a substantive, debatable claim rather than something the statistical analysis itself can verify, which is precisely why instrumental variables studies are frequently and legitimately debated on the specific credibility of their chosen instrument's exclusion restriction, not just on their statistical execution.
Why this approach matters despite these genuine limitations
Instrumental variables analysis provides a path to estimating a causal effect in situations where a genuine randomized experiment is infeasible or unethical, using naturally occurring variation instead — a genuinely valuable tool for exactly the many important economic and social questions where randomization simply isn't possible, provided the analysis is transparent about the specific instrument used and honest about the exclusion restriction's underlying, contestable assumptions.
What this means for evaluating research that uses instrumental variables
- Check whether the proposed instrument plausibly satisfies both relevance (directly testable) and the exclusion restriction (a substantive, contestable claim)
- Be specifically skeptical of an instrument whose exclusion restriction rests on an assumption that seems contestable given what else is known about the broader context
- Recognize that instrumental variables analysis is attempting to approximate a randomized experiment using naturally occurring variation, not a substitute that avoids all the underlying causal inference challenges
- Look for the specific justification a study provides for its exclusion restriction, since this argument, not the statistical machinery itself, is usually where the real analytical vulnerability lies
Instrumental variables analysis is a genuinely clever and useful tool precisely because it works around the impossibility of randomizing many variables of real interest — its core vulnerability is always the same: finding an instrument whose only path to the outcome runs through the variable you actually care about, a claim that has to be argued for, not simply assumed.