Instrumental Variable Estimation (IV)

Instrumental variable estimation identifies the causal effect of an endogenous treatment by using an instrument that shifts the treatment but affects the outcome only through that treatment, typically via two-stage least squares.

도구변수 추정은 처치에는 영향을 주지만 결과에는 처치를 통해서만 영향을 주는 도구변수를 이용해 내생적 처치의 인과효과를 식별하는 방법으로, 주로 2단계 최소제곱법으로 추정합니다.

How do I estimate a causal effect when the treatment is correlated with unmeasured confounders?

처치가 측정되지 않은 교란요인과 상관되어 있을 때 인과효과는 어떻게 추정하나요?

Instrumental variable estimation recovers a causal effect by using a variable that influences treatment but has no direct path to the outcome, so only the instrument-driven variation in treatment is used. It fits when unmeasured confounding or reverse causality is likely and a credible, strong instrument is available, and with heterogeneous effects it estimates a local average treatment effect for compliers.

When to use Instrumental Variable Estimation (IV)

• Treatment likely affected by unmeasured confounders or reverse causality
• Natural experiments, such as lotteries, distance, or policy rules, that shift treatment
• Randomized encouragement designs with imperfect compliance

When to use something else

• No credible instrument but confounders are measured: use propensity score methods such as IPW
• Treatment assigned by a cutoff on a running variable: use regression discontinuity design

Assumptions and requirements

• Relevance: the instrument is strongly related to treatment, with a first-stage F statistic conventionally above 10
• Exclusion: the instrument affects the outcome only through treatment
• Independence: the instrument is as good as randomly assigned
• Monotonicity for a local average treatment effect interpretation

What the analysis reports

• First-stage coefficients and F statistic
• Two-stage least squares estimate with standard error and confidence interval
• Endogeneity test (Durbin-Wu-Hausman)
• Overidentification test (Sargan or Hansen) when there are multiple instruments

How to report it

Two-stage least squares estimates indicated that [treatment] increased [outcome] by [value] (SE = [value], p = [value]), and the first-stage F statistic was [value].

Related methods

• Regression Discontinuity Design
• Difference-in-Differences
• Inverse Probability Weighting (IPW)

Run Instrumental Variable Estimation (IV) on Quorant