Instrumental Variables (IV) Method for Causal Inference
Also known as: IV, two-stage least squares, TSLS, causal estimation
Instrumental variables (IV) is an econometric method to estimate causal effects when treatment or exposure is not randomly assigned and confounding is severe or unmeasured. IV relies on a third variable (instrument) that influences treatment but does not directly affect the outcome, allowing researchers to isolate the causal effect from the noise of confounding. Developed extensively in econometrics (Angrist & Pischke, 1990s–2000s), IV methods are increasingly used in health economics and health services research to leverage natural experiments and policy changes.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+60 more
When to use it
When randomization is infeasible or unethical, treatment is endogenous (determined by patient/provider factors correlated with outcome), and a valid instrument exists. Common scenarios in health: (1) Health services research: causality of hospital/physician practices on outcomes. (2) Health policy evaluation: effects of insurance coverage expansion, price changes, or regulatory policies. (3) Medication/treatment effects: when observational data has severe selection bias but exogenous variation (policy, geography) exists. (4) Behavioral interventions: when encouragement (randomized) is feasible but compliance imperfect (use encouragement as IV). Not appropriate when: no valid instrument available (no exogenous variation in treatment); outcome is rare (IV method needs sufficient events); treatment effect is small (requires large sample for precision); heterogeneous effects dominant (LATE not informative).
Strengths & limitations
- Isolates causal effects in observational data: leverages natural experiments (exogenous variation in treatment) to estimate causality without RCT.
- Addresses selection bias: when unmeasured confounding or reverse causality severe, IV can provide unbiased estimates (unlike standard regression).
- Efficient use of data: observational data often abundant; IV methods extract causal insight without requiring new data collection (unlike RCT).
- Generates population-relevant estimates: when instruments target specific populations (e.g., geographic, policy), LATE reflects real-world variation in practice.
- Weak instrument bias: if instrument (Z) is weakly correlated with treatment (X), IV estimator is biased and unreliable; larger bias than OLS. Weak IV (F-stat <10) is common problem.
- LATE ≠ ATE: IV estimates the effect for 'compliers' (those influenced by instrument), not population average. If compliance varies by subgroup, generalizability uncertain.
- Exogeneity assumption untestable: IV assumes Z does not directly affect Y (only through X). This assumption cannot be tested directly from data; relies on reasoning and sensitivity analysis.
- Instrument scarcity: finding valid instruments is hard. Geographic variation, policy changes, and randomized encouragement designs are limited in scope; many research questions lack valid instruments.
- Interaction with confounding: if both confounding and instrument endogeneity present, IV may not fully resolve bias.
- Sample size: IV requires larger sample than OLS for equivalent precision, especially if instrument is weak or compliance low.
Frequently asked
What is the difference between an instrument and a confounder?
Confounder: variable affecting both treatment (X) and outcome (Y), biasing X→Y association. Instrument (Z): variable affecting only X (predicting X), with no direct path to Y. Confounders must be controlled (statistical adjustment). Instruments are leveraged to estimate causal effects despite unmeasured confounding.
How is the 'exogeneity' of an instrument tested?
Exogeneity (Z independent of Y except through X) is NOT directly testable from data. Validation relies on: (1) Domain reasoning: 'Is it plausible that Z affects Y only through X?' (2) Balance test: does Z correlate with pre-treatment baseline covariates? If yes, Z may not be exogenous. (3) Falsification test: do 'bad' outcomes unaffected by treatment predict differently by Z? If yes, Z is endogenous. (4) Literature review: do published studies validate similar instruments? Sensitivity analysis to alternative instruments strengthens claims.
What does 'weak instrument' mean and why is it a problem?
Weak instrument (IV): Z weakly predicts X (low correlation, F-stat <10). When IV is weak, TSLS amplifies any violation of exogeneity assumption, producing estimates more biased than OLS. Example: if Z is only weakly correlated with X but somewhat correlated with omitted confounder U, the bias in IV can exceed OLS bias. Staiger & Stock rule: F-stat <10 suggests weak IV, avoid relying on those estimates.
Can you have multiple instruments for one treatment?
Yes. Multiple valid instruments (Z1, Z2, Z3 all predict X, all exogenous) can be used in generalized method of moments (GMM) or system IV, potentially increasing precision. Over-identification test (Sargan or Hansen J-test) examines whether all instruments satisfy exogeneity. Rejection suggests at least one instrument is endogenous; investigate which.
How do you interpret LATE (Local Average Treatment Effect) in practice?
LATE is the causal effect of treatment on outcome for the 'complier' subpopulation—individuals whose treatment status is influenced by the instrument. If 30-year-old patients living near specialists receive treatment due to access (influenced by distance Z), LATE describes treatment effect for them. If 70-year-old patients have treatment access regardless of distance (uninfluenced by Z), LATE may not apply to them. When reporting IV results, always clarify which population (compliers) the LATE represents.
Sources
- Angrist, J. D., & Pischke, J. S. (2009). Mostly Harmless Econometrics: An Empiricist's Companion. Princeton: Princeton University Press. link ↗
- Bound, J., Jaeger, D. A., & Baker, R. M. (1995). Problems with Instrumental Variables Estimation When the Correlation Between the Instruments and the Endogenous Explanatory Variable is Weak. Journal of the American Statistical Association, 90(430), 443-450. DOI: 10.1080/01621459.1995.10476536 ↗
- Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.). Cambridge, MA: MIT Press. link ↗
How to cite this page
ScholarGate. (2026, June 4). Instrumental Variables (IV) Method for Causal Inference. ScholarGate. https://scholargate.app/en/health-economics/instrumental-variables
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cost-Effectiveness AnalysisHealth Economics↔ compare
- Decision Analytic ModelingHealth Economics↔ compare
- Markov Model in Health EconomicsHealth Economics↔ compare