Instrumental variables with treatment clusters
summary
The gist
Instrumental variables with treatment clusters shows that IV methods with treatment clusters identify a weighted average of local average treatment effects (LATEs) under a set of assumptions at the…
In short
Instrumental variables applied to treatment clusters identify a weighted average of local average treatment effects (LATEs) under assumptions made at the unclustered level. The method handles multiple unordered treatments by clustering them and provides testable conditions for causal interpretation, allowing researchers to understand which specific LATEs contribute to the overall effect.
Key concepts
- IV estimation with treatment clusters
- This technique uses instrumental variables when treatments are grouped into clusters. It aims to estimate the causal effect while accounting for these groups. The core finding is that this method yields a weighted average of local average treatment effects (LATEs) based on assumptions made at the unclustered level.
- Local Average Treatment Effects (LATEs)
- LATEs represent the causal effect of a treatment for a specific subgroup of individuals who actually receive that treatment. In this context, the method identifies a weighted average of these LATEs when multiple clusters are involved, providing a more granular view than an unclustered estimate.
- Unclustered Treatment Level Assumptions
- The paper relies on standard assumptions like independence and exclusion at the unclustered treatment level. These assumptions are crucial for the IV method to be valid, even when treatments are clustered. The authors derive necessary conditions for these assumptions and propose tests to verify their validity using data from the unclustered level.
Terminology used across episodes
This episode discusses
- Instrumental variables with treatment clusters · Paper Radio
- Sharp Testable Implications of Encouragement Designs · Paper Radio
- Causal Effects in Matching Mechanisms with Strategically Reported Preferences
- Effect or Treatment Heterogeneity? Policy Evaluation with Aggregated and Disaggregated Treatments
The paper
Instrumental variables with treatment clusters · Read on arXiv
Didier Nibbering, Matthijs Oosterveen, Pedro Luis Silva
Department of Econometrics and Business Statistics, Monash University · Department of Economics, Lisbon School of Economics and Management, and Advance/ISEG Research, University of Lisbon · Department of Applied Economics, Universitat Autonoma de Barcelona
Many policy evaluations include multiple unordered treatments. A common approach is to aggregate these treatments into clusters and apply instrumental variable (IV) methods at the treatment cluster level: The clusters are treated as binary treatments and the IV assumptions are invoked at the cluster level. This paper shows that IV with treatment clusters identifies a weighted average of local average treatment effects (LATEs) under a set of assumptions at the unclustered treatment level. When data on unclustered treatment choice is available, we propose testable necessary conditions for the causal interpretation of IV with treatment clusters. We illustrate our methods with empirical applications to the causal effects of fields of study aggregated into STEM and non-STEM clusters or eight broad study directions.
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: "Instrumental variables with treatment clusters".
Warren: Here's the gist of the paper. Instrumental variables with treatment clusters shows that IV methods with treatment clusters identify a weighted average of local average treatment effects (LATEs) under a set of assumptions at the unclustered treatment level Core Contribution The paper shows that IV estimation with clusters identifies a weighted average of LATEs under assumptions imposed at the unclustered treatment level The authors define mutually exclusive binary treatments and instruments based on multi-valued unordered variables They invoke standard IV assumptions at the unclustered level, including independence and exclusion Identification of Causal Interpretation The IV estimand with two treatment clusters is shown to be a weighted average of LATEs under Assumptions 1-4 The decomposition in Lemma 1 shows that the IV estimand in (1) is not necessarily a positively weighted average of the treatment effects corresponding to the response types (D(j) ∈ S1, D(k) ∈ S2) with j ∈ S1 and k ∈ S2 Empirical Relevance and Testing The results are illustrated through an empirical application studying the causal effect of fields of study on study credit points in the Portuguese public higher education system The necessary conditions for causal interpretation are testable using data at the unclustered treatment level via conditional moment inequalities Robustness and Extensions The authors derive necessary conditions for the identifying assumptions and propose falsification tests for settings where unclustered data is available Furthermore, they extend the TSLS estimand to fuzzy regression discontinuity designs The TSLS estimand in the fuzzy RDD has a causal interpretation under the local continuity assumption The paper concludes that IV methods with treatment clusters identify weighted averages of LATEs and provides testable conditions to justify the required assumptions in the presence of treatment clusters The results allow researchers to precisely consider the validity of required assumptions and clarify which LATEs contribute and their weights References The paper references literature on field choice, instrumental variables with binarised treatments, and regression discontinuity designs It also cites work on testing local average treatment effect assumptions using the clrtest package The empirical application draws upon data from the Portuguese public higher education system The authors acknowledge financial support from various funding bodies The paper also cites work on strategic misreporting in matching mechanisms The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing continuity assumptions in RDDs The data sources are DGES and DGEEC The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching The paper also cites work on testing local average treatment effect assumptions using the clrtest package The references include foundational works such as Imbens and Angrist on IV estimation and the Gale-Shapley algorithm for matching <ref:2611.1331348,Imbens, G.
Ruth: W.
Warren: and J.
Title and authors: Ruth: So, we're looking at this paper called "Instrumental variables with treatment clusters." It’s dealing with situations where you have a bunch of choices, like different fields of study, and researchers often group those choices into bigger buckets or clusters.
Warren: Right. And the authors are tackling how to use instrumental variables when you apply that approach at the cluster level instead of the individual level. It sounds like they're trying to figure out what exactly they are estimating with that method, especially when you have these grouped treatments.
Iris: The core idea seems to be that if we treat these clusters as binary treatments and apply the instrumental variable assumptions at that cluster level, we can actually identify a weighted average of local average treatment effects, or LATEs. This is something they're doing under certain assumptions made at the unclustered treatment level.
Ruth: That’s exactly what the abstract says. They are showing that this IV estimation with clusters identifies a weighted average of LATEs under those specific assumptions at the unclustered treatment level, which is pretty significant because it gives us a clearer picture of what we are actually measuring.
Warren: It moves beyond just getting some aggregated effect and tries to quantify the contribution of different underlying treatments within those clusters. But the paper immediately points out that this identification relies on a set of assumptions that have to hold true at the unclustered level, like independence and exclusion.
Iris: And they go further by showing how this decomposition in Lemma one reveals that when you have two treatment clusters, say STEM versus non-STEM, the IV estimand isn't necessarily a positive weighted average of the treatment effects corresponding to those two specific groups <ref:2611.13313#pg2>.
Ruth: That’s a crucial detail because it means we can’t just assume everything lines up perfectly with what we expect from a simple combination of effects. It shows that the weights themselves matter and they aren't automatically positive in the way intuition might suggest.
Warren: The empirical application they use to show this involves studying the causal effect of fields of study on study credit points in the Portuguese public higher education system, using administrative data to create instruments in a fuzzy regression discontinuity design.
Iris: And what’s interesting about that application is that they explicitly test the necessary conditions for this causal interpretation using data available at the unclustered treatment level through conditional moment inequalities. This makes it testable.
Ruth: So, before we get into how they improve things, let’s just summarize what this paper is doing: they are taking multiple unordered treatments, grouping them into clusters, and showing that IV methods at the cluster level give you a weighted average of LATEs under unclustered assumptions.
Title and authors: Warren: It sets up the framework for how we interpret those cluster-level estimates, but it leaves the heavy lifting of verifying those assumptions to be done at a finer level.
Iris: And looking ahead, they propose testable necessary conditions for that causal interpretation, which is a big step because it gives us something concrete to check against our real data.
Ruth: They also extend this idea to fuzzy regression discontinuity designs, where the TSLS estimand with an arbitrary number of clusters still has a causal interpretation provided we assume local continuity.
Warren: And they even derive some bounds on the treatment effect weights using Corollary one which helps us understand how policy encouragement actually affects how much of a treatment gets taken up <ref:2611.13313#pg2>.
Iris: There's also something interesting in Corollary three where they identify individual LATEs under Assumption one and Assumptions two through five at the unclustered level, showing that beta ab equals lambda j k delta t j k for specific response types.
Ruth: That’s a lot of detail packed into the results. It clarifies exactly which LATEs are contributing to the final cluster-level estimate, which is hard to do otherwise when you have so many variables mixed together.
Warren: So, what does this mean for us practically? It means if you're working with aggregated data where you can't treat every single individual treatment separately, this paper gives you a rigorous way to interpret the results of your cluster-level IV analysis.
Iris: It tells us that we need to be careful about our assumptions at the unclustered level, and it provides the tools—the testable conditions—to justify those assumptions when we have data available at that more detailed level.
Ruth: And for anyone working on policy evaluation or market design where choices are naturally clustered, this paper offers a path to correctly attributing effects to the underlying individual treatment choices.
Warren: The real cost here, as you always ask, is the complexity of checking those unclustered assumptions; it’s not just plugging in numbers and running a regression.
Iris: And they address that by providing those moment inequalities for testing, which connects the cluster-level estimation back to observable data at the finer level.
Ruth: So, to wrap up this segment on "Instrumental variables with treatment clusters," we have seen how this paper provides a method to interpret cluster-level IV estimates as weighted averages of LATEs under unclustered assumptions, and they give us testable conditions for those assumptions.
Warren: It’s a lot of machinery, but the goal is to provide clarity on what those cluster-level numbers actually represent in terms of the underlying individual choices.
Iris: I just want to emphasize that if you can't check the unclustered assumptions, this interpretation doesn't fully hold up for causal inference.
Ruth: Exactly. We’ll take a quick break, and when we come back, we’re going to look at some other papers on identification-aware MCMC methods.
The paper's summary: Ruth: So, to recap, this paper is about using instrumental variables when you have these clusters of treatments instead of just looking at every single thing separately.
Warren: Right. They’re showing that if you treat those groups as binary treatments and run the instrumental variable math at the cluster level, you end up with a weighted average of local average treatment effects, or LATEs.
Iris: That’s the main identification they're making: you get this weighted average of LATEs under these assumptions that are set at the unclustered treatment level.
Ruth: It means we can interpret those cluster-level numbers as a blend of individual treatments, but only if those underlying assumptions hold true at the finer level.
Warren: It’s not a simple average, though they point out in Lemma one that it’s not automatically a positive weighted average of the effects you expect from those two clusters.
Iris: That’s important because it means we can't just assume everything lines up neatly based on our intuition about how these groups interact.
Ruth: And they do offer ways to test those necessary assumptions, which is key for causal inference, and they even extend this idea to fuzzy regression discontinuity designs.
Warren: They show that the TSLS estimand with an arbitrary number of clusters still has a causal interpretation if you assume local continuity.
Iris: They also provide bounds on the treatment effect weights using Corollary one, which helps us see how policy encouragement actually affects how much of a treatment gets taken up.
Ruth: And there’s this idea in Corollary three where they identify individual LATEs under Assumption one and Assumptions two through five at the unclustered level.
Warren: So what this means for us practically is that if you're working with aggregated data where you can't treat every single thing separately, you have a rigorous way to interpret those cluster-level estimates.
Iris: It tells us we need to be careful about our assumptions at the unclustered level and gives us the testable conditions to justify those assumptions when we have that more detailed data available.
Ruth: For anyone doing policy evaluation or market design where choices are naturally grouped, this paper offers a path to correctly attributing effects to the individual treatment choices underneath.
Warren: The real hurdle is checking those unclustered assumptions, though they give you tools for that using moment inequalities.
Iris: And if you can't check those unclustered assumptions, then this interpretation doesn't fully hold up for causal inference, which is the big caveat there.
Ruth: So it boils down to having a clear way to see how the cluster-level numbers relate back to what’s actually happening with the individual treatments.
Warren: And that brings us right back to the testing conditions they propose, which is where the real work for anyone trying to use this method comes in.
The paper's improvements: Tom: So, to wrap up the main points, this paper sets up an identification scheme for cluster-level instrumental variables, showing that you get a weighted average of local average treatment effects under specific assumptions at the unclustered level.
Ruth: That’s right. Now they spend a lot of time looking at how to make these results more robust and useful in real life by suggesting improvements to the original framework.
Warren: They propose testable necessary conditions for those identification assumptions, which is a big deal because it gives us something concrete to check against our actual data using methods like conditional moment inequalities.
Iris: And they extend the TSLS estimand with an arbitrary number of clusters, showing that it still has a causal interpretation provided you make the local continuity assumption.
Ruth: They also derive bounds on the treatment effect weights using corollary one, which helps us understand exactly how policy encouragement affects how much a treatment gets taken up.
Warren: That gives us some real insight into the trade-off between encouraging something and actually getting people to use it in the first place.
Iris: And they also show this works even in fuzzy regression discontinuity designs, where they prove the TSLS estimand is still a weighted average of LATEs under local continuity.
Ruth: The biggest thing I see is how Corollary three lets them identify individual LATEs when response type restrictions are imposed, which really helps clarify which specific treatment effects are contributing to the final estimate.
Warren: It moves beyond just knowing the aggregate effect and lets you pinpoint what’s driving that aggregate effect in a way that’s mathematically grounded.
Iris: But they also flag limitations, pointing out that this whole framework relies on those unclustered assumptions holding true at a finer level, which is where the real econometric work has to happen.
Ruth: So the paper isn't just giving us a new estimation trick; it’s giving us a roadmap for how to actually verify if that trick works in practice.
Warren: It’s about making sure that when we use these cluster methods, we aren't just getting some noise because the underlying structure doesn't match the assumptions.
Iris: And they also propose falsification tests for settings where unclustered data is available, which gives us a way to prove when this identification strategy breaks down.
Ruth: It’s about adding a layer of verification so that researchers aren't just running models and hoping for the best, but they can test the assumptions rigorously.
Warren: So the focus now shifts from just doing the estimation to actively testing whether those assumptions are actually true before you trust the result.
Conclusion: Ruth: So to wrap up, this paper on instrumental variables with treatment clusters shows that you can identify weighted averages of local average treatment effects under specific assumptions at the unclustered level.
Warren: That’s right. The main point is that cluster-level instrumental variable methods yield a weighted average of LATEs if you stick to certain assumptions at the finer level.
Iris: It really boils down to this: you get an identification, but that identification only holds if those unclustered assumptions are true.
Ruth: And they gave us testable conditions, which is huge because it means we can actually check if those assumptions are valid using data available at the unclustered level through conditional moment inequalities.
Warren: It shows that even with these clusters, the work isn't completely arbitrary; there’s a structure you have to follow.
Iris: The implication is that researchers need to be very careful about what they assume when they aggregate treatments into clusters, because those assumptions aren't automatically guaranteed.
Ruth: It means if you want to use these cluster methods for causal inference, you have to prove the identification conditions are met before you trust your result.
Warren: And the paper also extends this concept into fuzzy regression discontinuity designs, showing that even in those cases, as long as local continuity holds, the TSLS estimand still has a causal interpretation.
Iris: That extension is helpful because it shows the framework isn't just limited to simple binary treatments; it’s more flexible.
Ruth: It gives us a way to think about how policy encouragement affects treatment uptake by providing those bounds on the effect weights using corollary one.
Warren: So if you are running a strategy, that means you can quantify exactly how much policy nudging is helping or hindering the actual adoption rate of a treatment.
Iris: And they also provide falsification tests for when unclustered data is available, which lets us know exactly when this identification method falls apart.
Ruth: It’s about moving from just getting an estimate to understanding the precise conditions under which that estimate is valid for causal claims.
Warren: So the big picture here is that cluster-level analysis isn't a magic shortcut; it’s a sophisticated tool, but you have to be rigorous about checking the assumptions at every level.
Iris: I think what this paper does best is give us the tools—the moment inequalities and falsification tests—to make that rigor possible.
Ruth: Exactly. So we’ve looked at how to interpret cluster-level IVs, what they mean for policy application, and the specific tests you need to run.
Warren: It leaves us with a clear idea: if you want to claim a causal effect from clustered data, you have to prove the identification structure first.
Iris: Before we move on, I just want to stress that the success of this whole method hinges entirely on verifying those unclustered assumptions correctly.
Ruth: Right. So next up, we’re going to look at how these identification-aware MCMC methods can help us handle even more complex problems in statistical inference.
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19