Sharp Testable Implications of Encouragement Designs
summary
The gist
This paper derives sharp, closed-form testable implications for potential outcome models featuring continuous or discrete outcomes and multi-valued treatments and instruments where each instrument…
In short
This research develops sharp, closed-form tests for potential outcome models called 'encouragement designs.' These models assume that each instrument value encourages at most one unique treatment choice. The derived inequalities provide a powerful way to check if observed data violates these behavioral restrictions and pinpoint exactly where the violation occurs.
Key concepts
- Encouragement Design
- A setting where the key assumption is that each possible value of an instrument only encourages individuals toward at most one specific treatment choice. This assumption is a crucial identifying feature in many real-world empirical studies.
- Testable Implications
- Mathematical inequalities derived from the model that can be directly tested using observed data distributions. These implications are sharp, meaning any data distribution violating them must violate the underlying behavioral restrictions.
- Sharpness (Constructive Proof)
- The proof shows that the derived inequalities are not just necessary conditions but also sufficient. This means any observed data satisfying these inequalities can actually be generated by a valid model within the restricted class, making the test complete.
- Potential Outcome Model
- A framework used to study decisions where individuals face multiple potential treatments based on an instrument. The paper analyzes models with continuous or discrete outcomes and multi-valued treatments and instruments.
Terminology used across episodes
This episode discusses
- Sharp Testable Implications of Encouragement Designs · Paper Radio
- Inference for Treatment Effects Conditional on Generalized Principal Strata using Instrumental Variables
- On the Identifying Power of Generalized Monotonicity for Average Treatment Effects
- Identification in Multiple Treatment Models under Discrete Variation
- Treatment Effects with Targeting Instruments
- Testing Mechanisms
- When Should We (Not) Interpret Linear IV Estimands as LATE?
- Generalized Encouragement-Based Instrumental Variables for Counterfactual Regression
The paper
Sharp Testable Implications of Encouragement Designs · Read on arXiv
Yuehao Bai, Shunzhuang Huang, Max Tabord-Meehan
University of Southern California · Booth School of Business University of Chicago · University of Toronto
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "Sharp Testable Implications of Encouragement Designs".
Warren: Here's the gist of the paper. This paper derives sharp, closed-form testable implications for potential outcome models featuring continuous or discrete outcomes and multi-valued treatments and instruments where each instrument value encourages at most one unique treatment choice, which serves as the key identifying assumption in several prominent recent empirical papers Core Model Setup The analysis focuses on a setting called an encouragement design, where the restriction is that "each value of the instrument encourages towards at most one unique treatment choice" The paper introduces notation for multi-valued treatments and instruments, specifically considering two forms of the instrument support: (1) Z = 0, J0,..., J − 1 and (2) Z = 0, J0,..., J − 1 where Z = 0 is interpreted as the “base state” of the instrument The observed choice D is related to potential treatment choices and the instrument through an equation involving a sum over z in Z.
Ruth: Derivation of Testable Implications The central result is Theorem 3.1, which characterizes necessary conditions for the distribution P to be consistent with the class Q1, stating that for z(j) ∈ Z(j), 0 ≤ j ≤ J − 1, X 0≤j≤J−1 P(D = jZ = z(j)) ≤ 1.
Warren: This inequality is a direct consequence of the restriction in (3).
Ruth: Furthermore, Corollary 3.1 provides a simplification for J0 > 0, stating that the inequalities are equivalent to the statement that for 0 ≤ j ≤ J − 1 and k ∈ Z such that j ≠ k, P(D = jZ = k) ≤ P(D = jZ = 0).
Paper summary: Ruth: Now we’ve looked at the math behind these potential outcome models and encouragement designs, so we're wrapping up with a look at this paper by Yuehao Bai et al. on "Sharp Testable Implications of Encouragement Designs." It boils down to having these sharp, closed-form testable implications for when you have multi-valued treatments and instruments where each instrument level only encourages one specific choice.
Warren: That’s the core idea they’re pushing. They prove that those inequalities are not just theoretical; they're actually necessary conditions for the data to fit within that specific class of restrictions <ref:2411.09808#pg2>.
Iris: And they also proved they're sharp, which means any data that satisfies those inequalities is definitely compatible with those behavioral restrictions <ref:2411.09808#pg3>.
Ruth: That testable part is what matters. You can use these inequalities to immediately check if your observed data violates the assumption about how instruments encourage choices in a way that makes sense <ref:2411.09808#pg2>.
Warren: And they give you a constructive argument, so if you find a violation, they show you exactly which substitution pattern is causing it, which is really helpful for identification <ref:2411.09808#pg3>.
Iris: It connects this to real-world problems where people choose between different programs or treatments based on what's offered at the time <ref:2411.09808#pg3>.
Ruth: This isn't just abstract math; it gives us a way to check the logic of these behavioral assumptions right against the numbers we get from surveys or experiments <ref:2411.09808#pg2>.
Warren: It also shows that this setup is equivalent to a fairly general random utility model, which means the default choice you see in models like that arises naturally <ref:2411.09808#pg3>.
Iris: So, we've seen how they build this tool and what it says about those behavioral assumptions. Now we look at how researchers can actually use these implications to test their own hypotheses on incentive structures <ref:2411.09808#pg3>.
Conclusion: Ruth: So, we’ve looked at the math behind these potential outcome models and encouragement designs, now we're at the conclusion of this paper by Bai and colleagues on "Sharp Testable Implications of Encouragement Designs."
Warren: It boils down to having sharp, closed-form testable implications for when you have multi-valued treatments and instruments where each instrument level only encourages one specific choice.
Iris: That’s the core idea. They prove that these inequalities are not just theoretical; they're actually necessary conditions for the data to fit within that specific class of restrictions.
Ruth: And they also proved they're sharp, which means any data that satisfies those inequalities is definitely compatible with those behavioral restrictions.
Warren: That’s the testable part. You can use these inequalities to immediately check if your observed data violates the assumption about how instruments encourage choices in a way that makes sense.
Iris: And they give you a constructive argument, so if you find a violation, they show you exactly which substitution pattern is causing it, which is really helpful for identification.
Ruth: It connects this to real-world problems where people choose between different programs or treatments based on what's offered at the time.
Warren: This isn't just abstract math; this gives us a way to check the logic of those behavioral assumptions right against the numbers we get from surveys or experiments.
Iris: It also shows that this setup is equivalent to a fairly general random utility model, which means the default choice you see in models like that arises naturally.
Ruth: So, we've seen how they build this tool and what it says about those behavioral assumptions. Now we look at how researchers can actually use these implications to test their own hypotheses on incentive structures.
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