Bayesian Instrumental Variables (Bayesian IV)
Bayesian Instrumental Variables Estimation · Also known as: Bayesian IV, Bayesian 2SLS, Bayesian LIML, BayesIV
Bayesian Instrumental Variables combines the instrumental variable strategy for addressing endogeneity with Bayesian posterior inference. Instead of relying on asymptotic sampling distributions, it places prior distributions over all structural parameters and recovers a full posterior distribution for the causal effect, providing probability statements about the parameter rather than p-values — especially valuable when instruments are weak or the sample is small.
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.
When to use it
Use Bayesian IV when you face endogeneity — the treatment variable is correlated with the error term — and you have a valid instrument: one that is relevant (strongly associated with treatment) and excludes itself from the outcome equation directly. Prefer the Bayesian formulation over classical 2SLS when (a) the instrument is weak (first-stage F-statistic below 10), (b) the sample is small and asymptotic approximations are unreliable, (c) you wish to incorporate genuine prior information about parameters, or (d) you need full posterior uncertainty rather than point estimates. Do not use it when no credible instrument exists — a weak or invalid instrument produces a wide and potentially misleading posterior regardless of the estimation framework. Also avoid it when OLS is consistent (no endogeneity) because IV is inefficient relative to OLS in that case.
Strengths & limitations
- Provides full posterior distributions for causal parameters, enabling direct probability statements rather than hypothesis tests.
- Handles weak instruments more honestly than classical 2SLS: posterior uncertainty is wide when the instrument is weak, rather than producing falsely precise but biased estimates.
- Finite-sample validity — does not rely on asymptotic approximations, which makes it better suited to small or moderate samples.
- Allows incorporation of prior knowledge about instrument strength or the range of the causal effect from theory or earlier studies.
- Joint estimation of both stages naturally propagates endogeneity and instrument uncertainty into the posterior for the causal parameter.
- Requires a valid instrument; no amount of Bayesian machinery compensates for an invalid or irrelevant instrument.
- MCMC sampling can be computationally intensive, especially in large datasets or models with many endogenous variables.
- Prior sensitivity may be non-trivial near weak-instrument regions of the parameter space; results should be checked under alternative priors.
- Exclusion restriction — the instrument affects the outcome only through the treatment — is untestable and must be defended on subject-matter grounds.
Frequently asked
How does Bayesian IV differ from classical 2SLS?
Classical 2SLS estimates both stages sequentially using OLS, then derives asymptotic standard errors. Bayesian IV estimates both stages jointly via a full-information likelihood paired with priors, yielding a posterior distribution. The key practical difference is in finite samples and weak-instrument settings: 2SLS can be severely biased and misleadingly precise, while the Bayesian posterior is wide and honest about the extra uncertainty.
What makes an instrument valid?
Two conditions: relevance (the instrument must be correlated with the endogenous treatment, verifiable in the first stage) and exclusion (the instrument must affect the outcome only through the treatment, not directly). The first is testable; the second is not and must be justified on theoretical or institutional grounds.
How do I choose priors?
Common defaults are weakly informative normal priors on regression coefficients and inverse-Wishart priors on the error covariance. If you have theoretical bounds on the causal effect or evidence on instrument strength from prior studies, you can encode these in tighter priors. Always check sensitivity to prior choice, especially when the instrument is weak.
Can I use Bayesian IV with binary or count outcomes?
Yes, but the model becomes a structural probit or Poisson system rather than a linear one. Bayesian estimation via MCMC accommodates these extensions naturally, though the exclusion restriction argument still requires careful justification.
When should I stick with classical 2SLS instead?
When you have a large sample, a demonstrably strong instrument, and no useful prior information, classical 2SLS is simpler and gives nearly equivalent answers. The Bayesian approach adds the most value in small samples, weak-instrument scenarios, or when prior knowledge is genuinely informative.
Sources
- Kleibergen, F., & Zivot, E. (2003). Bayesian and classical approaches to instrumental variable regression. Journal of Econometrics, 114(1), 29-72. DOI: 10.1016/s0304-4076(02)00219-1 ↗
- Lancaster, T. (2004). An Introduction to Modern Bayesian Econometrics. Blackwell Publishing. ISBN: 978-1405117203
How to cite this page
ScholarGate. (2026, June 3). Bayesian Instrumental Variables Estimation. ScholarGate. https://scholargate.app/en/causal-inference/bayesian-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.
- 2SLS RegressionEconometrics↔ compare
- Bayesian Difference-in-DifferencesCausal inference↔ compare
- Bayesian RegressionBayesian↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Instrumental Variables in Health ResearchHealth Economics↔ compare
- Propensity Score MatchingResearch Statistics↔ compare