When researchers want to uncover cause-and-effect relationships using regression models—especially in social sciences, medicine, or economics—one crucial challenge is ensuring the validity of their identification strategy. Exclusion restrictions play a central role in this process, especially when using instrumental variables (IV) or seeking to address endogeneity. But what happens when data is clustered (for example, students within schools or patients within hospitals)? And how can exclusion restrictions be respected and tested in these clustered linear regression models, where standard assumptions and procedures may not directly apply?
Short answer: Exclusion restrictions in clustered linear regression models must be handled with care to maintain credible inference. This involves ensuring that the chosen instruments (or excluded variables) are valid both within and across clusters, accounting for potential intra-cluster correlation, and often includes robust estimation techniques like cluster-robust standard errors. Additionally, researchers sometimes need to limit their analysis to subsamples where overlap and instrument validity are strongest, as suggested by propensity score trimming approaches, to avoid bias and maintain precision.
Why Exclusion Restrictions Matter in Clustered Settings
At the heart of instrumental variable (IV) estimation is the exclusion restriction: the instrument must influence the outcome only through its effect on the endogenous regressor, not directly. In clustered data settings, such as students nested within schools or workers within firms, this requirement can become even more challenging. If the instrument varies mainly at the cluster level (for example, a policy applied to all schools in a district), then all units within a cluster share the same value of the instrument, raising the risk that unobserved cluster-level factors could violate the exclusion restriction.
NBER's technical working paper by Crump, Hotz, Imbens, and Mitnik (nber.org) highlights another layer of complexity: lack of overlap in the distribution of key covariates (or propensity scores) can make estimation in clustered models fragile. If, for example, some clusters never receive the treatment or instrument, or only receive it under certain conditions, the exclusion restriction can be harder to justify for those clusters. The authors note that “estimation of average treatment effects under unconfoundedness or exogenous treatment assignment is often hampered by lack of overlap in the covariate distributions,” which can lead to imprecise estimates and sensitivity to model specification.
Clustered Standard Errors and Robustness
Another major consideration in clustered models is the correlation of errors within clusters. Standard errors that ignore clustering can be severely underestimated, leading to spurious statistical significance. When dealing with exclusion restrictions, it is therefore standard practice to use cluster-robust standard errors, which account for within-cluster correlation. This adjustment is not just a technicality—it is essential for valid inference about whether the exclusion restriction appears to hold, especially when testing for overidentification or conducting sensitivity analyses.
Sample Trimming and Overlap: Lessons from Propensity Score Methods
The NBER paper further offers a practical guideline derived from propensity score analysis: in cases where overlap is limited, the authors recommend a simple rule—dropping all units with estimated propensity scores outside the range of 0.1 to 0.9. This is because, in these regions, the lack of comparable treated and untreated units (or instrumented and non-instrumented) makes the exclusion restriction especially suspect. The authors found that “for a wide range of distributions a good approximation to the optimal rule is provided by the simple selection rule to drop all units with estimated propensity scores outside the range [0.1,0.9]” (nber.org). In clustered settings, this often translates to dropping entire clusters or subgroups where the instrument or treatment is almost never or almost always assigned, since these are the places where the exclusion restriction is least credible.
Practical Example: Schools and Educational Interventions
Imagine a study evaluating the effect of a new teaching method (the treatment) using the fact that some schools adopted it early (the instrument). If some school districts never adopt the method and others always do, the exclusion restriction—that the instrument affects test scores only through use of the teaching method—becomes hard to defend in districts with no variation. Moreover, unmeasured district-level policies could confound the results. By trimming the sample to focus on districts where both adoption and non-adoption occur, and by applying cluster-robust standard errors, the exclusion restriction becomes more plausible and the inference more reliable.
Testing Exclusion Restrictions in Clustered Data
In clustered models, formal tests of the exclusion restriction (such as overidentification tests using multiple instruments) must also account for clustering. Failing to do so can make these tests invalid. For example, if the instrument is only assigned at the cluster level, the effective number of independent observations is the number of clusters, not the number of individuals. This greatly reduces statistical power and can make violations of the exclusion restriction harder to detect.
Limitations and Ongoing Challenges
As the NBER paper and broader econometric literature emphasize, even with these robust methods, exclusion restrictions can rarely be “proven” in observational data, especially when clustering is present. Researchers must rely on a combination of theory, institutional knowledge, sensitivity checks, and robust estimation techniques. Sometimes, as suggested by the NBER authors, “changing the estimand”—for instance, estimating treatment effects only for the subgroup with sufficient overlap—can be the most credible approach.
Other sources, such as sciencedirect.com and coursera.org, did not provide additional substantive guidance on this topic in the provided excerpts. However, the key themes—about the need for overlap, the use of cluster-robust methods, and careful sample selection—are well supported by both theory and the best-practice recommendations found in the NBER working paper.
Concrete Takeaways and Real-World Application
To summarize, handling exclusion restrictions in clustered linear regression models requires several concrete steps: ensuring that the instrument or excluded variable is valid both within and across clusters, using cluster-robust standard errors to account for intra-cluster correlation, and sometimes trimming the sample to focus on clusters or units where the instrument is plausibly exogenous. As the NBER paper puts it, “lack of overlap can lead to imprecise estimates and can make commonly used estimators sensitive to the choice of specification.” By applying these principles—such as trimming based on propensity scores and using robust inference—researchers can strengthen the credibility of their exclusion restrictions in clustered settings.
In practical terms, suppose a researcher is studying the impact of a healthcare intervention rolled out at the hospital level. If some hospitals always implement the intervention and others never do, the exclusion restriction is weakest in those hospitals. Focusing analysis on hospitals with variation, and applying cluster-robust inference, enhances both the internal validity and the interpretability of the results.
In conclusion, while exclusion restrictions are always a conceptual and practical challenge, clustered data introduces unique complexities that require careful handling. Drawing on best practices from the literature, especially as synthesized in the NBER technical working paper, researchers can navigate these challenges by combining robust statistical techniques with thoughtful sample selection and a clear-eyed assessment of where their identification strategy is most credible. This approach not only protects against bias and invalid inference, but also provides a transparent and defensible framework for causal analysis in real-world, clustered data.
