Testing Exclusion and Shape Restrictions in Potential Outcomes Models

arXiv:2512.20851 · econ.EM · Submitted 2025-12-24 · 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: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Testing Exclusion and Shape Restrictions in Potential Outcomes Models".

Warren: Here's the gist of the paper. Exclusion and shape restrictions are central to defining causal effects and interpreting estimates in potential outcomes models, where this paper develops a general framework for characterizing sharp testable implications of general support restrictions on potential response functions using a novel graph-based representation of the model.

Ruth: This framework offers a unified, constructive method for deriving all observable implications arising from the modeling assumptions.

Warren: The observed data involves an exogenous input Z in Z (e.g., instruments or treatments) and an endogenous response R in R, which may encompass endogenous treatments, mediators, or outcomes.

Ruth: The model is defined by the potential response vector R* = (R*(x)) x in X in RX, indexed by counterfactual inputs x in X, and a known mapping T: RX times Z to R such that the observed response satisfies R = T(R*, Z).

Paper summary: Ruth: To start, this paper, Testing Exclusion and Shape Restrictions in Potential Outcomes Models, sets out to develop a general framework for characterizing sharp testable implications arising from general support restrictions on potential response functions using a novel graph-based representation of the model.

Warren: Essentially, they are taking assumptions about exclusion and shape restrictions in potential outcomes models—things we use to define causal effects—and turning them into something we can actually test with data.

Iris: The core claim is that this framework provides a unified and constructive method for deriving all observable implications of the modeling assumptions. They illustrate this approach in several popular settings, including instrumental variables, treatment selection, mediation, and interference.

Ruth: They focus on how the observed data comes from an exogenous input Z—say an instrument or a treatment—and an endogenous response R that depends on potential outcomes R star. The model is defined by the potential response vector R star indexed by counterfactual inputs X and a known mapping T that links it to the observed response.

Warren: The paper introduces a graph where vertices represent these potential responses combined with the input Z, and if two vertices are disconnected in this graph, it means those two events cannot happen jointly under the modeling assumptions.

Iris: That graph structure is crucial because it encodes the compatibility structure of the model, which helps determine if a joint distribution of potential outcomes that matches our observed data can actually exist.

Ruth: The theoretical results establish necessary and sufficient conditions for that joint distribution to exist, with theorem one stating that for any maximal independent set in the potential response graph, an inequality involving the probability of observing those responses is necessary.

Warren: Then theorem two broadens this, saying that for any subset of vertices you pick, there's a corresponding inequality involving lambda G and the size of that subset.

Iris: And under assumption three, which requires the graph to be perfect and every maximal clique to correspond to a specific support point, those inequalities from theorem one become sharp. This makes it constructive.

Ruth: It’s a unified approach that allows practitioners to empirically assess common modeling assumptions in a very broad class of potential outcome models, moving beyond case-by-case testing.

Conclusion: Warren: So what does this paper mean for us when we look at the title, Testing Exclusion and Shape Restrictions in Potential Outcomes Models, and who did the authors on it? It’s about giving us a general tool to test those fundamental modeling assumptions.

Ruth: It means we can move past just testing IV validity or mediation in isolation and start using this framework to see how these restrictions interact across different models systematically.

Iris: The implication is that practitioners gain a constructive way to derive all the observable implications stemming from their assumptions, which is a big step toward robust inference in causal modeling.

Warren: It’s about turning abstract modeling constraints into concrete, testable conditions that we can check against our empirical data using this graph theory approach.

Ruth: Basically, it gives us a way to rigorously assess the validity of the structure of our potential response functions without having to test every single possible scenario individually.

Boston University · University of Chicago

econ.EM

Submitted: 2025-12-24

Updated: 2026-11-10

Code: https://github.com/hkaido0718/SupportRestriction

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

Importance score: 88/100

The gist: Exclusion and shape restrictions are central to defining causal effects and interpreting estimates in potential outcomes models…

Key concepts

Potential Response Function
This is the vector of all possible outcomes ($\mathbf{R}^*$) that could occur for every possible input ($x$). It represents what would have happened under different scenarios, such as different treatments or exposures, before any actual observation occurred.
Graph-Based Representation
The model's compatibility structure is encoded in a graph where vertices represent specific potential outcomes. If two vertices are disconnected, it means the corresponding outcomes cannot happen together under the model's assumptions.
Maximal Independent Set (MIS)
An MIS is a set of vertices in the graph where no two vertices are connected by an edge. Theorem 1 shows that for a valid probability distribution to exist, specific inequalities involving the probabilities of outcomes within such sets must hold.
Regular Support Restrictions
This assumption simplifies the analysis by requiring the graph to be perfect and ensuring that every maximal clique corresponds exactly to a single possible support point in the potential response space.

Terminology

Summary

Exclusion and shape restrictions are central to defining causal effects and interpreting estimates in potential outcomes models, where this paper develops a general framework for characterizing sharp testable implications of general support restrictions on potential response functions using a novel graph-based representation of the model. p. 0 This framework offers a unified, constructive method for deriving all observable implications arising from the modeling assumptions. p. 0

The observed data involves an exogenous input Z in Z (e. p. 1g., instruments or treatments) and an endogenous response R in R, which may encompass endogenous treatments, mediators, or outcomes. p. 1 The model is defined by the potential response vector R* = (R*(x)) x in X in RX, indexed by counterfactual inputs x in X, and a known mapping T: RX times Z to R such that the observed response satisfies R = T(R*, Z). p. 1 Support restrictions are typically of the form R* in RX almost surely, for a known set R* RX. p. 1

A graph-based representation encodes the compatibility structure of the model. p. 1 If two vertices v r,z and v r',z' are disconnected in this graph G, then the events R = r when Z = z and R = r' when Z = z' cannot occur jointly under the modeling assumptions.

The core theoretical results establish necessary and sufficient conditions for the existence of a joint distribution Q of R* supported on R* that induces the observed probability distribution P. p. 5 Theorem 1 states that if Assumptions 1–2 hold, then for any maximal independent set I of the potential response graph G, the inequality X v r,z in I P(R = r Z = z) 1 is necessary for such a distribution Q to exist. p. 3 Theorem 2 generalizes this by stating that if Assumptions 1–2 hold, then for any subset V of the vertices of G, the inequality X v r,z in V P(R = r Z = z) G(V) is necessary and sufficient for the existence of Q. p. 5 Under Assumption 3 (Regular Support Restrictions)—which requires G to be perfect and every maximal clique to correspond to a support point r* in R* —Theorem 3 establishes that the MIS inequalities from Theorem 1 are sharp. p. 3

The analysis extends to continuous covariates through a sequence of potential response graphs G n and a generalized theorem (Theorem 4) based on partitions A n,k. p. 3 Furthermore, Theorem 6 provides a necessary and sufficient condition for the existence of Q based on maximal cliques (C* G) and maximal independent sets (IG n) across a sequence of graphs. p. 5

Computational aspects involve two methods for constructing G: checking pairwise incompatibility (complexity O Y 2 D 2 Z squared) or enumerating all r* in R* (complexity O(Y D D Z)). p. 6 Inference procedures utilize moment inequalities derived from the half-space representation of conditional distributions, and a test based on vertex representation involving solving a quadratic program to define critical values. p. 4 The framework accommodates continuous covariates by discretizing them into bins when testing such implications. p. 4

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. General framework for deriving sharp testable implications: The system can derive all observable implications of modeling assumptions using a graph-based representation of the model, which facilitates interpretation through efficient graph algorithms and provides a unified method for testing assumptions across various settings like instrumental variables and mediation.

  2. Computation of testable implications: The system can efficiently compute maximal independent sets (MIS) of the potential response graph to derive necessary moment inequalities, ensuring these inequalities are sharp under regularity conditions.

  3. Handling continuous outcomes: The system can extend the analysis to settings where the response vector R is continuous by constructing sequences of potential response graphs and using partitions to derive general analogues of testable implications.

  4. Model comparison and sensitivity analysis: The system can compare multiple models with nested support restrictions by showing that adding a stronger restriction either tightens existing inequalities or introduces new ones, allowing practitioners to evaluate the impact of changing modeling assumptions on implied restrictions.

  5. Inference for continuous covariates: The system can perform inference using methods like Chernozhukov et al.'s approach for testing sharp implications when covariates include continuous components by formulating testable implications as an intersection bound involving a nuisance parameter.

Sources

Related papers