Inference in clustered IV models with many and weak instruments

arXiv:2306.08559 · econ.EM · Submitted 2023-06-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Inference in clustered IV models with many and weak instruments".

Warren: Here's the gist of the paper. Inference in clustered IV models with many and weak instruments shows that clustered data makes it more likely that the instruments are weak or many, which necessitates new robust tests for instrumental variable models The paper proposes extensions to the jackknife Anderson–Rubin and jackknife score tests to handle clustered data by removing clusters rather than individual observations Pruning redundant reasoning tokens The paper demonstrates that data clustering reduces the effective sample size, making instruments more likely to be weak or many, which increases the likelihood of problems for estimators and tests in instrumental variable models Specifically, it shows that previously developed robust tests cannot be applied to clustered data because they require independent observations The authors adapt the jackknife Anderson–Rubin and jackknife score tests to clustered data by removing clusters instead of individual observations This adaptation is crucial because the dependence between observations within the same cluster reduces information, making weak and many instrument problems more likely How the training signal is built The core of the proposed tests involves modifying existing robust tests by removing terms with non-zero expectation using a "cluster jackknife" instead of a standard jackknife The cluster jackknife AR statistic is derived by removing all products of errors within the same cluster to center the statistic, which requires Assumption A2 regarding rank and leverage conditions Furthermore, four extensions are proposed for these tests, including combining the AR and score statistics into a conditional linear combination test Where the gains come from The proposed improvements aim to enhance power and robustness through several mechanisms: 1.

Ruth: Combining AR and score statistics allows for a more powerful test via a conditional linear combination test.

Ruth: The symmetric cluster jackknife retains part of the information in removed observations, leading to a larger signal component than the standard cluster jackknife.

Paper summary: Ruth: So we're looking at this paper, "Inference in clustered IV models with many and weak instruments". The core idea here is that when you have data clustered together, like observations within a specific group, it actually makes those instruments look weaker or too numerous.

Warren: That sounds like a problem for any model trying to figure out how things work based on those instruments. So what does the paper claim they did to fix this issue?

Ruth: They found that data clustering reduces the effective sample size, and this makes the instruments more likely to be weak or many, which in turn increases the chance of problems for estimators and tests in instrumental variable models. They point out that previous robust tests just can't be used on clustered data because they demand independent observations.

Iris: I see. So the authors are adapting existing robust tests, specifically the jackknife Anderson–Rubin and jackknife score tests, to handle this clustering by removing entire clusters instead of looking at individual observations.

Warren: That makes sense from a quantitative standpoint. If you remove the cluster whole, you're not relying on those dependent observations directly for your test statistic. But how does that actually work in practice?

Ruth: The paper explains that the core of the proposed tests involves modifying existing robust tests by removing terms with non-zero expectation using what they call a "cluster jackknife" instead of a standard jackknife. They derive the cluster jackknife Anderson–Rubin statistic by removing all products of errors within the same cluster to center that statistic.

Iris: That sounds like they need an assumption about rank and leverage conditions to make that work, because of Assumption A2 mentioned in the paper. And they propose four extensions for these tests, including combining the Anderson–Rubin and score statistics into a conditional linear combination test.

Warren: So it's not just one thing; they are building out these extensions to make them more flexible for different setups. What's the payoff for all this modification?

Ruth: The gains come from several mechanisms. Combining the Anderson–Rubin and score statistics allows for a more powerful test via that conditional linear combination test. They also have a symmetric cluster jackknife that keeps some of the information from those removed observations, which gives it a bigger signal component than the standard cluster jackknife they started with.

Iris: And they introduced something called a "many controls jackknife" to handle cases where you have many exogenous control variables by correctly projecting them out. Plus, they suggested alternative variance estimators like cross-fitting residuals, which helps power by removing bias in the variance estimator when the hypothesized parameter value is far from its true value.

Warren: That sounds like they're tackling the power issue head-on, which is always a concern when you have many instruments or weak ones. So how did they test all this out?

Ruth: They revisit a study on queenly reign on war using a dataset clustered in reigns, where data were grouped into one hundred seventy-six clusters of varying sizes. The simulations confirmed that the cluster jackknife tests achieve close to nominal size, and they show superior power compared to adaptations of the Anderson–Rubin and score tests for independent data when there are many instruments.

Iris: That empirical application is interesting because it shows real-world relevance. It also confirms that queenly reign has a significant positive effect on war likelihood in the full model when using these new robust tests.

Warren: If we look at the numbers, it seems like these new cluster jackknife tests perform well, actually showing superior power relative to the cluster adaptations of the Anderson–Rubin and score tests for independent data when there are many instruments. So it sounds like they solved a specific problem in a specific setting.

Ruth: The conclusion of this paper is that the problems associated with many and weak instruments are more likely in clustered data than when they are independent, making these robust tests essential. The cluster jackknife tests are size correct, whereas the tests for independent data overreject, and these new robust tests show good power.

Iris: So what does this mean for us if we're working with clustered data in our actual research? It suggests that standard methods might be unreliable when instruments overlap or observations are grouped.

Warren: It means we need to move toward these cluster jackknife tests when dealing with clustered data because they seem more reliable, showing good power even when the instruments are weak or there are many of them. The paper also confirms the effect in their example about queenly reign on war likelihood.

Ruth: Exactly. It's about acknowledging that the structure of the data—the clustering—changes how we have to test our models, and this paper gives us a way to do that robustly. That's what this work is really focused on: providing tools for inference when you have these tricky instrument situations in clustered settings.

Conclusion: Ruth: So we've seen how data clustering makes those instrumental variables look weak or numerous, and this paper tackles that head on with new cluster jackknife tests.

Warren: It sounds like the big takeaway is that standard ways of testing these models just don't work well when your data has that clustering structure.

Iris: The authors are essentially saying you have to change how you do your inference, moving away from methods that assume independent observations when they are actually grouped together.

Ruth: Exactly. They introduce these cluster jackknife statistics that remove entire clusters instead of just looking at individual data points in a way that respects the dependence within those groups.

Warren: That's the mechanism—removing the whole cluster to center the statistic—and it seems to be size correct, which is usually a big win when you're dealing with complex models.

Iris: The paper shows these tests have superior power compared to adapting existing tests for independent data, especially when you have a lot of instruments and they are weak.

Ruth: So the conclusion is that if your data is clustered, you need these specific robust tools because the standard tests just overreject.

Warren: It means if someone's using an IV model with clustered data, they really need to switch to this cluster jackknife approach for their inference to be accurate.

Iris: The authors confirm that even in their specific example about queenly reign and war likelihood, these new tests still give them a significant positive effect result.

Ruth: It shows the practical value of this work because it's not just theory; it's working on real data sets right now to show these tests can actually find effects when they matter.

Warren: So we've seen how the math works and why it matters for clustered data, but next up we should look at what happens when you try to run this kind of model in a massive, high-dimensional setting with even more instruments.

econ.EM

Submitted: 2023-06-14

Updated: 2026-11-06

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: Inference in clustered IV models with many and weak instruments shows that clustered data makes it more likely that the instruments are weak or many…

Key concepts

Clustered Data Effect
When data points are grouped into clusters, the effective sample size decreases. This reduction makes it more probable that instruments in an IV model are either too weak or there are too many of them, which can cause standard tests to fail.
Cluster Jackknife Tests
These are modified versions of existing robust tests. Instead of using a standard jackknife on individual data points, the authors remove entire clusters. This adaptation is necessary because the dependence within a cluster reduces information, making these new tests more reliable for clustered settings.
Many and Weak Instruments
This refers to situations in instrumental variable models where there are either too many instruments relative to the number of observations or when some instruments have very low relevance. Clustered data exacerbates these problems, requiring specialized inference methods.

Terminology

Summary

Inference in clustered IV models with many and weak instruments shows that clustered data makes it more likely that the instruments are weak or many, which necessitates new robust tests for instrumental variable models The paper proposes extensions to the jackknife Anderson–Rubin and jackknife score tests to handle clustered data by removing clusters rather than individual observations

Pruning redundant reasoning tokens

The paper demonstrates that data clustering reduces the effective sample size, making instruments more likely to be weak or many, which increases the likelihood of problems for estimators and tests in instrumental variable models Specifically, it shows that previously developed robust tests cannot be applied to clustered data because they require independent observations The authors adapt the jackknife Anderson–Rubin and jackknife score tests to clustered data by removing clusters instead of individual observations This adaptation is crucial because the dependence between observations within the same cluster reduces information, making weak and many instrument problems more likely

How the training signal is built

The core of the proposed tests involves modifying existing robust tests by removing terms with non-zero expectation using a cluster jackknife instead of a standard jackknife The cluster jackknife AR statistic is derived by removing all products of errors within the same cluster to center the statistic, which requires Assumption A2 regarding rank and leverage conditions Furthermore, four extensions are proposed for these tests, including combining the AR and score statistics into a conditional linear combination test

Where the gains come from

The proposed improvements aim to enhance power and robustness through several mechanisms:

  1. Combining AR and score statistics allows for a more powerful test via a conditional linear combination test. p. 8

  2. The symmetric cluster jackknife retains part of the information in removed observations, leading to a larger signal component than the standard cluster jackknife. p. 8

  3. A many controls jackknife is introduced to handle many exogenous control variables by correctly projecting them out.

  4. Alternative variance estimators are proposed, such as cross-fitting residuals, which improve power by removing bias in the variance estimator when the hypothesized parameter value is far from its true value

Empirical application

The authors revisit the study on queenly reign on war using a dataset clustered in reigns, where data are clustered in 176 clusters of varying size They find that having a queen significantly increases the likelihood of an armed conflict when using these new cluster robust tests Simulations confirm that the cluster jackknife tests attain close to nominal size, and they show superior power relative to adaptations of the AR and score tests for independent data when there are many instruments The empirical application confirms that queenly reign has a significant positive effect on war likelihood in the full model using robust tests

Conclusion

The paper concludes that the problems associated with many and weak instruments are more likely in clustered data than when they are independent, making robust tests essential The cluster jackknife tests are size correct, whereas tests for independent data overreject, and the new robust tests show good power The application to Dube and Harish (2020) confirms that queenly reign has a significant positive effect on war likelihood

References

Abadie, A., S. Athey, G. W. Imbens, and J. M. Wooldridge (2023). p. 0 When should you adjust standard errors for clustering? p. 0 The Quarterly Journal of Economics 138 (1), 1–35

Ackerberg, D. A. and P. J. Devereux (2009). p. 8 Improved JIVE estimators for overidentified linear models with and without heteroskedasticity. p. 6 The Review of Economics and Statistics 91 (2), 351–362

Anatolyev, S. and N. Gospodinov (2011). p. 1 Specification testing in models with many instruments. p. 0 Econometric Theory 27 (2), 427–441

Anderson, T. W. and H. Rubin (1949). p. 1 Estimation of the parameters of a single equation in a complete system of stochastic equations. p. 9 The Annals of Mathematical Statistics 20 (1), 46 – 63

Andrews, I. (2016). Conditional linear combination tests for weakly identified models. p. 1 Econometrica 84 (6), 2155–2182

Barrios, T., R. Diamond, G. W. Imbens, and M. Koles´ar (2012). p. 1 Clustering, spatial correlations, and randomization inference. p. 0 Journal of the American Statistical Association 107 (498), 578–591

Bekker, P. A. (1994). Alternative approximations to the distributions of instrumental variable estimators. p. 0 Econometrica 62 (3), 657–681

Bekker, P. A. and F. Crudu (2015). p. 1 Jackknife instrumental variable estimation with heteroskedasticity. p. 0 Journal of Econometrics 185 (2), 332–342

Bekker, P. A. and J. van der Ploeg (2005). p. 2 Instrumental variable estimation based on grouped data. p. 0 Statistica Neerlandica 59 (3), 239–267

Billingsley, P. (1995). Probability and Measure (Third ed. p. 1). John Wiley & Sons Ltd.

Boot, T. and J. W. Ligtenberg (2025). p. 1 Identification-and many moment-robust inference via invariant moment conditions. p. 4 arXiv preprint arXiv:2303. p. 0 p. 007822

Chao, J. C., N. R. Swanson, J. A. Hausman, W. p. 8 K. Newey, and T. p. 2 Woutersen (2012). p. 1 Asymptotic distribution of JIVE in a heteroskedastic IV regression with many instruments. p. 5 Econometric Theory 28 (1), 42–86

Chao, J. C., N. R. Swanson, and T. Woutersen (2023). p. 0 Jackknife estimation of a cluster-sample IV regression model with many weak instruments. p. 2 Journal of Econometrics

Chetverikov, D., J. Hahn, Z. Liao, and A. Santos (2023). p. 0 Standard errors when a regressor is randomly assigned. p. 0 arXiv preprint arXiv:2303.10306

Crudu, F., G. Mellace, and Z. S´andor (2021). p. 0 Inference in instrumental variable models with heteroskedasticity and many instruments. p. 0 Econometric Theory 37 (2), 281–310

Djogbenou, A. A., J. G. MacKinnon, and M. p. 5 Ørregaard Nielsen (2019). p. 5 Asymptotic theory and wild bootstrap inference with clustered errors. p. 4 Journal of Econometrics 212 (2), 393–412

Dov´ı, M.-S., A. B. Kock, and S. Mavroeidis (2023). p. 0 A ridge-regularized jackknifed AndersonRubin test. p. 1 Journal of Business & Economic Statistics 0 (0), 1–12

Dube, O. and S. P. Harish (2020). p. 2 Queens. Journal of Political Economy 128 (7), 2579–2652

Feng, Q., S. Jaidee, and W. Wang (2025). p. 0 Robust inference with high-dimensional instruments. p. 2 arXiv preprint arXiv:2506. p. 023834

Frandsen, B., E. Leslie, and S. McIntyre (2025). p. 0 Cluster jackknife instrumental variables estimation. p. 1 Review of Economics and Statistics, 1–19

Hansen, B. E. and S. Lee (2019). p. 5 Asymptotic theory for clustered samples. p. 4 Journal of Econometrics 210 (2), 268–290

Hausman, J. A., W. K. Newey, T. Woutersen, J. C. Chao, and N. p. 1 R. Swanson (2012). p. 1 Instrumental variable estimation with heteroskedasticity and many instruments. p. 0 Quantitative Economics 3 (2), 211–255

Kleibergen, F. (2002). Pivotal statistics for testing structural parameters in instrumental variables regression. p. 1 Econometrica 70 (5), 1781–1803

Kleibergen, F. (2005). Testing parameters in GMM without assuming that they are identified. p. 2 Econometrica 73 (4), 1103–1123

Kline, P., R. Saggio, and M. Sølvsten (2020). p. 1 Leave-out estimation of variance components. p. 4 Econometrica 88 (5), 1859–1898

Lim, D., W. Wang, and Y. Zhang (2024a). p. 1 A conditional linear combination test with many weak instruments. p. 1 Journal of Econometrics 238 (2), 105602

Lim, D., W. Wang, and Y. Zhang (2024b). A valid Anderson-Rubin test under both fixed and diverging number of weak instruments. p. 6 Unpublished

MacKinnon, J. G., M. Ø. Nielsen, and M. D. Webb (2022). p. 0 Leverage, influence, and the jackknife in clustered regression models: Reliable inference using summclust

Matsushita, Y. and T. Otsu (2022). p. 0 Jackknife Lagrange multiplier test with many weak instruments. p. 4 Econometric Theory, 1–24

Mikusheva, A. and L. Sun (2022). p. 0 Inference with many weak instruments. p. 0 Review of Economic Studies 89 (5), 2663–2686

Mikusheva, A., M. Sølvsten, and B. p. 1 Jing (2025). p. 0 Estimation in linear models with clustered data. p. 2 arXiv preprint arXiv:2508. p. 012860

Moreira, M. J. (2003). A conditional likelihood ratio test for structural models. p. 2 Econometrica 71 (4), 1027–1048

Newey, W. K. and J. R. Robins (2018). p. 2 Cross-fitting and fast remainder rates for semiparametric estimation. p. 2 arXiv preprint arXiv:1801. p. 009138

Phillips, G. D., A., and C. Hale (1977). The bias of instrumental variable estimators of simultaneous equation systems. p. 0 International Economic Review 18 (1), 219–228

Sølvsten, M. (2020). Robust estimation with many instruments. p. 4 Journal of Econometrics 214 (2), 495–512

Staiger, D. and J. H. Stock (1997). Instrumental variables regression with weak instruments. p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Robust Inference in Clustered IV Models: The improved system can apply many and weak instrument robust jackknife Anderson–Rubin and jackknife score tests to clustered data by removing clusters rather than individual observations from the statistics, as described in the paper.

  2. Cluster Jackknife Extensions: The AI system can implement four extensions of the cluster jackknife AR and score tests, including combining them into a conditional linear combination test for enhanced power and incorporating symmetric jackknife procedures to retain more signal component information.

  3. Many Controls Handling: The system can utilize a many controls jackknife to correctly project out control variables while centering the statistic, even in complex clustered models where partialling out controls introduces dependence between clusters.

  4. Cross-Fit Variance Estimation: The AI system can employ cross-fit variance estimators for clustered data, which improve power by removing bias in the variance estimator when the hypothesized value of the parameter of interest is far from its true value.

  5. Size Correctness and Power Assessment: The improved system can perform Monte Carlo simulations to show that cluster jackknife tests attain close to nominal size and exhibit superior power relative to their counterparts for independent data, while also identifying when tests become oversized or conservative in specific scenarios like dominant clusters.

Sources

Related papers