Inference After Model Refutation: Uniform Post-Relaxation Inference for Partially Identified Models
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: "Inference After Model Refutation".
Warren: Here's the gist of the paper. When data reject a model, researchers must decide which identifying assumptions to discard.
Ruth: If multiple repairs restore compatibility, and the repair itself is unknown, inference under any single repair might exclude structural values supported by others.
Warren: The proposed Post-Relaxation Projection (PR) addresses this by carrying one simultaneous confidence event through every minimal repair.
Ruth: For a rectangular band defined on the repair-specific restrictions, PR captures exactly those post-relaxation targets compatible with that band.
Title and authors: Ruth: So we’re back to the title, "Inference After Model Refutation: Uniform Post-Relaxation Inference for Partially Identified Models." It sets up the whole problem we just discussed about model rejection and multiple repairs.
Warren: It means the paper is focused on figuring out how to do inference when you have several equally good ways to fix your model, but you don't know which one is actually right.
Iris: The focus seems to be on creating a uniform way to make inferences that works across all those possible repairs instead of just picking one.
Ruth: Right. It’s not just about finding *a* valid conclusion; it’s about making sure the conclusion holds regardless of which repair you happened to select from the menu.
Warren: The authors are proposing this Post-Relaxation Projection method as a way to handle that uncertainty systematically, carrying one simultaneous confidence event through every minimal repair.
Iris: And they build a rectangular band on the restrictions specific to each repair, and this captures precisely what targets are compatible with that band.
Ruth: The difficulty they highlight is learning the set of population minimum repairs from the data itself, because that’s a discrete object you have to learn from the same sample.
Warren: That selection step can delete crucial components of your target even if the next step is exact, so they are worried about that failure mode.
Iris: So they introduce three criteria—repair recovery concerns, entire-target coverage concerns, and structural-value coverage concerns—to guide their approach to avoiding that failure.
Ruth: And PR specifically targets the entire-target coverage concern first, which seems like the most comprehensive way to ensure you don't miss anything.
The paper's summary: Warren: To summarize what they did, they developed Post-Relaxation Projection to handle model refutation by considering all minimal repairs simultaneously.
Iris: They show that PR carries one simultaneous confidence event through each repair and builds a rectangular band based on the specific restrictions of that repair.
Ruth: This method captures exactly the post-relaxation targets compatible with that band, which is key when you have multiple candidates for model fixes.
Warren: They also derived limits on learning the repair itself, its exact perturbation radius, oracle equivalent adaptation, and versions for instrumental variables without a first-stage relevance bound.
Iris: This is important because it addresses the question of how much information you need about the repair to perform inference reliably.
Ruth: It means they’ve also found limits on learning the repair, how big its error can be, and what kind of adaptation you need for your instrumental variables.
Warren: They are looking at this in cross-country data on institutions and income where the data don't settle which historical instrument to drop yet every admissible minimum repair implies a positive effect.
Iris: This is the scenario where multiple repairs support different structural parameters, making it hard to decide which one you should trust.
Ruth: So, the paper essentially says that even when the data doesn't tell you which instrument to drop, every viable minimum repair still suggests a positive effect survives sizable direct effects of that instrument.
The paper's improvements: Warren: The authors suggest several improvements to this initial approach to make it more robust and practical in real-world settings.
Iris: They propose Resolution-Adaptive Inference, which uses an oracle procedure for a certified layer and Post-Relaxation Projection otherwise.
Ruth: So when the data gives you a high resolution on feasibility bounds, the system can switch to that exact oracle inference instead of the more conservative projection.
Warren: When uncertainty is high, like with low resolution, it switches over to this more conservative, data-driven projection method.
Iris: This allows them to adapt their confidence procedures based on how statistically resolved their feasibility bounds are using the oracle procedure for a certified layer and PR otherwise.
Ruth: It sounds like a smart way to manage uncertainty—using the best tool available depending on what the data can actually tell you.
Warren: It also discusses uniform reduced-form inference without a relevance bound, allowing estimation even when first stages are weak.
Iris: This means they can get a set that is uniformly valid without any lower bound on the magnitude of p i j including cases where p i j equals zero.
Ruth: That prevents the system from getting overly conservative just because it’s unsure about how relevant an instrument is.
Conclusion: Warren: So to wrap up, the main implication is that this paper provides a uniform way to get inferences when models are rejected by data and multiple repairs are possible.
Iris: It shows that even with uncertainty about which repair you pick, you can still establish a robust conclusion because every admissible minimum repair implies something.
Ruth: The Post-Relaxation Projection method gives us an interval conclusion that is robust to uncertainty over the menu of repairs.
Warren: The RST data they use demonstrate this idea, showing that the conclusion holds under either one of those two possibilities.
Iris: It really shows how institutions matter for income under every repair that the data cannot rule out, and this minimum effect is economically large when instruments are held close to exact validity.
Ruth: The final word on "Inference After Model Refutation: Uniform Post-Relaxation Inference for Partially Identified Models" is that it’s a method for robust inference in these tough identification scenarios.
Ulrich Hounyo
University at Albany, State University of New York
econ.EM, stat.ME
Submitted: 2026-11-11
Updated: 2026-11-11
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 68/100
The gist: When data reject a model, researchers must decide which identifying assumptions to discard.
Key concepts
- Post-Relaxation Projection (PR)
- A method that handles model rejection by carrying one simultaneous confidence event through every minimal repair. It captures exactly those post-relaxation targets compatible with a band defined by the restrictions of each individual repair, ensuring robustness when the best repair is unknown.
- $R^ ext{*} (P)$
- A discrete population object learned from the same data used for inference. Selecting a specific repair and then inferring can fail if that selection step deletes a component of the target set. PR avoids this selection by retaining every repair that could still be a population-minimum on an event, ensuring all supported values are kept.
- Bandwise Completeness ($C_{PR}^n$)
- This establishes that the canonical PR method covers all necessary targets within a specific band defined by the repair menu. If any set $D$ contains the targets for every primitive function in the band, it must contain $C_{PR}^n$, providing a guarantee of comprehensive coverage under certain conditions.
Terminology
Summary
When data reject a model, researchers must decide which identifying assumptions to discard. p. 0 If multiple repairs restore compatibility, and the repair itself is unknown, inference under any single repair might exclude structural values supported by others. p. 0 The proposed Post-Relaxation Projection (PR) addresses this by carrying one simultaneous confidence event through every minimal repair. p. 0 For a rectangular band defined on the repair-specific restrictions, PR captures exactly those post-relaxation targets compatible with that band. p. 0 A key difficulty is that R(P) is a discrete population object learned from the same data used for inference; selecting a repair and then inferring can fail if the selection step deletes a component of the target. p. 2 The paper introduces three criteria: Repair recovery concerns (PP P 0), entire-target coverage concerns (PP P 0), and structural-value coverage concerns (PP P 0). p. 2 PR targets entire-target coverage. p. 3
Theorem 1 defines the post-relaxation target as theta MR(P) = [R in R(P) theta R(P)]. p. 3 PR avoids a repair-selection step by retaining every repair R that could still be population-minimum on an event and taking the union of their corresponding structural outer sets, ensuring that every population-minimum repair and its supported values are retained. p. 3 The construction maintains the finite repair menu, fixed ex ante costs, and assumes repairability (that for every P in P, at least one R in R satisfies theta R(P) not equal to). p. 1
Theorem 4 establishes bandwise completeness of canonical PR on En(P), stating C PR n = [f in B rep n T(f)], and consequently, any set D theta satisfying T(f) D for every f in B rep n n must contain C PR n. p. 3 Theorem 5 shows that under the primitive repair setup and Assumption 3, on En(P), R(P) R b n and theta MR(P) C PR n, leading to n P in P PP 0 PP 04. p. 3
The procedure involves four steps: 1. p. 2 List repairs R with costs/partial order and restrictions b Rj(theta, P) at most 0 beforehand. p. 1 2. Compute a simultaneous (1-alpha) band for all primitive functions and feasibility indices, yielding b Rj,n at most b Rj,n and q R,n at most q P(P). p. 3 3. Keep repair R if q R,n at most 0 and q S,n > 0 for every cheaper S, defining the set R b n. p. 5 4. Report the set C PR n = S R in R b n, b Rj,n(theta) at most 0 j, and read off conclusions holding on the whole set. p. 4
Corollary 3 provides a tolerance- and cost-uniform IV conclusion envelope: if [B-n(r bar), B+n(r bar)] (beta, beta), then for every r in [0, r], structural bounds are nonbinding for every coefficient vector in the simultaneous rectangle, the delete-all repair is not PR-plausible for any strictly positive additive weight vector w, and C PR n, r, w [B-n(r), B+n(r)] r in [0, r], w in (0, infinity). p. 5
The paper concludes that PR is narrower than the worst case over every relaxation in the menu but broader than a single repair picked from the data. p. 1 The RST data illustrate both sides: PR keeps both, and the conclusion it supports is a lower bound that holds under either one.
The final conclusion emphasizes that institutions matter for income under every repair that the data cannot rule out, and this minimum effect is economically large when instruments are held close to exact validity. Conclusions are conditional on the repair menu and relaxing as little as data requires. p. 6 The tolerance has a direct economic reading: it bounds the direct effect of an instrument on income relative to its first-stage effect on the rule-of-law index. p. 6 The lower envelope is 0. p. 2311 at r = 0.42 and 0.131 at r = 0.60; rounded down, beta > 0.31 and beta > 0.13. The sign conclusion is not sensitive to simulation error in the critical value: at r = 0. p. 260 it would fail only if the critical value exceeded 2. p. 297. The positive sign does not depend on the relevance bound, and the reduced-form rectangle gives 0. p. 0172, corresponding to direct effects of about 0. p. 006 log points per standard deviation; this is the price of allowing exactly zero first stages. p. 6
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
-
Uniform Post-Relaxation Inference (PR) allows for inference under any one of several equally preferred repairs, stating that
the repair itself is unknown, and inference under any one of them can exclude structural values supported by the others.
This enables systems to report apositive effect
even when data do not settle which historical instrument to drop. -
Resolution-Adaptive Inference (RA) allows systems to adapt their confidence procedures based on the statistical resolution of feasibility bounds, as it uses
the oracle procedure for a certified layer and PR otherwise.
This means the system can switch between an exact oracle inference and a more conservative, data-driven projection when uncertainty is high. -
Uniform reduced-form inference without a relevance bound allows systems to perform IV estimation even when first stages are weak, as it provides
a set that is uniformly valid without any lower bound on |pij| , including at pij = 0.
This prevents the system from failing or becoming overly conservative due to uncertainty about instrument relevance. -
Order-Robust Post-Relaxation Projection (PR under a finite partial order) ensures that the resulting confidence set
is robust to uncertainty over the prespecified orders,
meaning the system's conclusion does not change if its scientific ranking of repairs is slightly unsettled. -
Certified conclusions provide a mechanism where
one entire-target event protects all prespecified universal conclusions read deterministically from Cn; no additional multiplicity correction is needed.
This allows for robust reporting of claims that hold under a common event, simplifying the final output by certifying all necessary universal conclusions simultaneously.
Sources
Related papers
- Stationarity with Occasionally Binding Constraints
- Backtesting High-Dimensional Expected Shortfall
- Higher-Order Neyman Orthogonality in Moment-Condition Models
- Online Bayesian Inference for a Dirichlet Process Mixture of Binary Logit Models
- Debiased Conditional Independence Tests for Causal Effects
- Sharp Testable Implications of Encouragement Designs