Instrumental Variables
A causal inference technique that uses a variable affecting treatment but not outcome directly to identify causal effects when treatment is endogenous.
What Is Instrumental Variables?
Instrumental variables (IV) is the Nobel-honored causal inference technique that rescues you when treatment is confounded with outcome. The idea: find a variable Z (the instrument) that affects treatment T but has no direct effect on outcome Y and is not confounded with Y. Variation in Z induces exogenous variation in T, which identifies the causal effect T→Y even in the presence of unobserved confounders.
Also Known As
- Data science: IV estimation, 2SLS (two-stage least squares), Wald estimator
- Growth: natural experiments, quasi-experimental leverage
- Marketing: exogenous variation in treatment
- Engineering: identification via induced variation
How It Works
Classic marketing IV: you cannot randomize who downloads an app, but you can randomize who receives an install prompt (the instrument Z). The prompt affects installs (T) but shouldn't directly affect downstream revenue (Y) except through installation. You compute IV estimate: (ΔY in prompted - ΔY in unprompted) / (ΔT in prompted - ΔT in unprompted). If the prompt raised installs by 20pp and raised revenue by $1.40, the local average treatment effect is $7 per induced install.
This only identifies the effect on "compliers" — users who install because of the prompt. It does not identify effects on always-takers or never-takers.
Best Practices
- Check the first stage F-statistic. Below 10 indicates weak instruments, and IV estimates become wildly biased.
- Defend the exclusion restriction explicitly. Why does the instrument not affect Y except through T? This is not testable — it must be argued from substance.
- Interpret as LATE, not ATE. Your estimate applies to the subpopulation moved by the instrument.
- Report sensitivity to violation of exclusion via Anderson-Rubin confidence intervals or sensitivity bounds.
- Triangulate with other estimators — IV, DR, and differences should tell coherent stories.
Common Mistakes
- Using an instrument that likely violates exclusion. Most "clever" instruments in applied work turn out to have direct effects on the outcome.
- Ignoring weak-instrument bias. A 2SLS point estimate with an F of 3 is essentially noise.
- Claiming ATE when your instrument only identifies LATE for a small compliance subgroup.
Industry Context
In SaaS/B2B, IV can leverage randomized outreach, price experiments, or feature gating as instruments for downstream usage effects. In ecommerce, it is the right tool when randomizing discount offers (instrument) to estimate the effect of actually using the discount (treatment). In lead gen, randomized invitations to demo requests serve as instruments for demo-attended-to-close-rate analysis.
The Behavioral Science Connection
IV disciplines the human urge to treat every correlation as a candidate cause. It forces explicit argument about which variable is exogenous and why — a level of rigor that, honestly, most experimentation programs never reach. The exclusion restriction is a high bar and teams often fool themselves into thinking they clear it.
Key Takeaway
When randomization of treatment is infeasible but a randomized nudge exists, IV is the right bridge from correlation to causation. The catch: the validity of IV rests on assumptions that cannot be tested, only argued.