Design-based nested instrumental variable analysis

Bo Zhang; Michael O. Harhay; Xinran Li; Zhe Chen

arxiv: 2511.21992 · v2 · submitted 2025-11-27 · 📊 stat.ME · stat.AP

Design-based nested instrumental variable analysis

Zhe Chen , Xinran Li , Michael O. Harhay , Bo Zhang This is my paper

Pith reviewed 2026-05-17 05:44 UTC · model grok-4.3

classification 📊 stat.ME stat.AP

keywords nested IVdesign-based inferencepair-of-pairs designalways-compliersswitcherscompliance subgroupsinstrumental variablestreatment effect estimation

0 comments

The pith

A pair-of-pairs nested IV design identifies always-complier and switcher average treatment effects using design-based inference.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a pair-of-pairs nested instrumental variable design in which each matched stratum contains four units divided into two pairs. This setup allows researchers to identify and estimate average treatment effects separately for always-compliers, who take the treatment under both weaker and stronger encouragements, and switchers, who take it only under the stronger one. The method uses design-based inference that accounts for a partly biased randomization of the intensity of the instrumental variable. Such an approach matters because many real-world encouragements come in different strengths, and distinguishing these subgroups reveals why overall intention-to-treat effects may stay flat even when compliance rates rise, as illustrated in a cancer screening trial.

Core claim

In a nested IV structure where compliance with the weaker IV implies compliance with the stronger IV, the pair-of-pairs design randomizes IV assignment within pairs and employs a partly biased randomization scheme to handle non-random assignment of IV intensity within strata. This permits design-based inference for the sample average treatment effect among always-compliers and among switchers.

What carries the argument

The pair-of-pairs nested IV design, which structures four units into two matched pairs per stratum to support within-pair IV randomization while using a partly biased scheme for IV intensity assignment.

If this is right

Separate estimates of treatment effects become available for always-compliers and switchers.
The method remains valid in simulations with small samples and few switchers.
Application to the PLCO trial shows about 52% always-compliers and 27% switchers, with a trend toward benefit only among always-compliers.
This framework explains stable intention-to-treat effects despite rising compliance after 1997.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar designs could extend to other medical screening or policy interventions with graduated encouragement levels.
Researchers might use the separated estimates to better target interventions to specific compliance groups.
Further work could test the method's robustness when the nested compliance assumption is mildly violated.

Load-bearing premise

That individuals who comply under the weaker instrumental variable also comply under the stronger one, and that the partly biased randomization scheme correctly adjusts for non-random assignment of IV intensity within each stratum.

What would settle it

Finding individuals who comply with the weaker IV but fail to comply with the stronger IV would contradict the nested compliance structure and invalidate the separation of always-compliers from switchers.

Figures

Figures reproduced from arXiv: 2511.21992 by Bo Zhang, Michael O. Harhay, Xinran Li, Zhe Chen.

read the original abstract

Two binary instrumental variables (IVs) are nested if individuals who comply under one binary IV also comply under the other. This situation often arises when the two IVs represent different intensities of encouragement or discouragement to take the treatment, with one stronger than the other. In a nested IV structure, treatment effects can be identified for two latent subgroups: always-compliers and switchers. Always-compliers are individuals who comply even under the weaker IV, while switchers are those who do not comply under the weaker IV but do under the stronger IV. We introduce a novel pair-of-pairs nested IV design, where each matched stratum consists of four units organized in two pairs. We develop design-based inference for the always-complier sample average treatment effect and switcher sample average treatment effect. In a nested IV analysis, IV assignment is randomized within each IV pair; however, whether a study unit receives the weaker or stronger IV may not be randomized. To address this complication, we then propose a novel partly biased randomization scheme and study design-based inference under this new scheme. Using extensive simulation studies, we demonstrate the validity of the proposed method even in challenging scenarios with small sample sizes and a low proportion of switchers. Applying the nested IV framework, we estimated that 52.2% (95% CI: 50.4%-53.9%) of participants enrolled at the Henry Ford Health System in the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial were always-compliers, while 26.7% (95% CI: 24.5%-28.9%) were switchers. Among always-compliers, flexible sigmoidoscopy was associated with a trend toward a decreased colorectal cancer rate. No effect was detected among switchers. This offers a richer interpretation of why no increase in the intention-to-treat effect was observed after 1997, even though the compliance rate rose.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces a pair-of-pairs matching design for nested IVs and claims design-based inference for always-complier and switcher effects under a partly biased randomization scheme for IV intensity.

read the letter

The core contribution is a matching structure that organizes units into strata of four, split into two pairs, to support design-based estimation of the always-complier SATE and the switcher SATE in nested IV settings. This targets encouragement designs where one IV is stronger than the other, which shows up in trials like cancer screening. They also propose a partly biased randomization scheme to handle cases where assignment of the weaker versus stronger IV within a stratum is not fully random, then derive estimators and variance formulas under that scheme. Simulations cover small samples and low switcher proportions, and the PLCO application estimates compliance shares and finds a trend toward lower colorectal cancer rates among always-compliers but none among switchers. That gives a more granular reading of the overall null ITT result after 1997 despite higher compliance. The matching and the explicit handling of intensity assignment are the parts that look new relative to standard IV or nested IV work. The simulations and real-data breakdown are the practical strengths. The main soft spot is the partly biased scheme itself. The abstract notes that IV intensity assignment may not be randomized, so the scheme is introduced to restore design-based properties. If that scheme requires specifying a probability model for the intensity choice (even if based on observed covariates), the procedure mixes randomization with modeling assumptions. That weakens the exact finite-sample guarantees, especially for the smaller switcher group. The nested compliance assumption is also required to label the latent subgroups, and it is a substantive restriction rather than a design feature. This paper is for causal inference researchers who analyze encouragement trials or similar designs with varying IV strengths. Readers working on design-based methods or medical applications would find the matching construction and the empirical illustration useful. It deserves peer review because the setup addresses a recurring practical problem and the simulations provide some reassurance on finite-sample behavior, even though the theoretical status of the biased component needs closer checking.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a pair-of-pairs nested IV design in which each matched stratum consists of four units organized into two pairs. It develops design-based estimators and inference for the always-complier sample average treatment effect and the switcher sample average treatment effect under nested compliance, using a novel partly biased randomization scheme to accommodate non-random assignment of weaker versus stronger IV intensity within strata. The approach is supported by simulation studies under small samples and low switcher proportions, together with an application to the PLCO cancer screening trial that reports estimated proportions of always-compliers (52.2%) and switchers (26.7%) and their respective effects on colorectal cancer incidence.

Significance. If the claimed exact finite-sample design-based properties hold under the partly biased scheme, the work would extend design-based IV methods to nested encouragement structures and enable subgroup-specific inference without outcome modeling. The reported simulation coverage and the empirical decomposition of the ITT effect into compliance-group contributions would constitute a practical contribution to causal inference in matched designs.

major comments (2)

[§3] §3 (partly biased randomization scheme): the scheme is introduced to handle non-random assignment of IV intensity within strata, yet the manuscript does not provide an explicit probability measure over all possible assignments that is known from the design alone. If the intensity assignment instead depends on a parametric or semi-parametric model (even if only for the purpose of weighting), the resulting variance estimators for the switcher SATE would no longer be purely design-based; this directly undermines the central claim of exact finite-sample validity, especially when the switcher proportion is low as in the reported simulations.
[Simulation studies] Simulation section (tables reporting coverage for switcher SATE): the data-generating processes must explicitly implement the partly biased scheme as a known randomization distribution rather than as a fitted model; otherwise the reported coverage rates cannot confirm the design-based properties asserted for the switcher estimator.

minor comments (2)

[Abstract and §2] The abstract states that IV assignment is randomized within each pair but does not specify the exact matching procedure used to form the pair-of-pairs strata; a brief description in the main text would improve reproducibility.
[§2] Notation for the two latent subgroups (always-compliers and switchers) is introduced without an accompanying table that maps the four possible compliance types under the two nested IVs; adding such a table would clarify the identification argument.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive comments on our manuscript. Below, we provide point-by-point responses to the major comments and indicate the revisions we plan to make.

read point-by-point responses

Referee: [§3] §3 (partly biased randomization scheme): the scheme is introduced to handle non-random assignment of IV intensity within strata, yet the manuscript does not provide an explicit probability measure over all possible assignments that is known from the design alone. If the intensity assignment instead depends on a parametric or semi-parametric model (even if only for the purpose of weighting), the resulting variance estimators for the switcher SATE would no longer be purely design-based; this directly undermines the central claim of exact finite-sample validity, especially when the switcher proportion is low as in the reported simulations.

Authors: The partly biased randomization scheme is defined directly from the design, with the probability of assigning the stronger versus weaker IV within each stratum set to a fixed, known value that does not depend on any outcome model or compliance model. This probability measure is known a priori and is used to derive the exact finite-sample unbiasedness and variance expressions for both the always-complier and switcher SATE estimators. We acknowledge that the current presentation could be more explicit about the full probability space, and we will revise Section 3 to include a formal statement of the randomization distribution over all possible assignments. revision: yes
Referee: Simulation section (tables reporting coverage for switcher SATE): the data-generating processes must explicitly implement the partly biased scheme as a known randomization distribution rather than as a fitted model; otherwise the reported coverage rates cannot confirm the design-based properties asserted for the switcher estimator.

Authors: Our simulation data-generating processes do implement the partly biased scheme by generating the IV intensity assignments according to the fixed design probability, independent of any fitted model. The reported coverage rates therefore reflect the design-based properties under this known randomization distribution. To address the referee's point, we will expand the simulation description to explicitly detail the randomization mechanism used in the DGP. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained from novel design

full rationale

The paper defines a novel pair-of-pairs nested IV design with four units per stratum and introduces a partly biased randomization scheme to handle non-random IV intensity assignment. Design-based inference for always-complier and switcher SATEs is then derived directly from the known randomization distribution under this scheme. No load-bearing step reduces by construction to a fitted parameter, self-citation, or tautological redefinition; the estimands and variance formulas follow from the explicit randomization probabilities stated in the design. The nested compliance assumption is a substantive identifying restriction rather than a circular definitional move. The approach remains self-contained against external benchmarks with no evidence of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The method rests on the nested compliance assumption and on the validity of the partly biased randomization scheme for handling non-random IV intensity assignment. No free parameters or invented entities are mentioned in the abstract.

axioms (2)

domain assumption Nested compliance: individuals who comply under the weaker IV also comply under the stronger IV.
Stated directly in the abstract as the definition of the nested IV structure that allows identification of always-compliers and switchers.
domain assumption The partly biased randomization scheme correctly adjusts for non-random assignment of weaker vs. stronger IV within strata.
Introduced to address the complication that IV intensity may not be randomized; required for the design-based inference to be valid.

pith-pipeline@v0.9.0 · 5648 in / 1482 out tokens · 24329 ms · 2026-05-17T05:44:34.693406+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We introduce a novel pair-of-pairs nested IV design... partly biased randomization scheme... design-based inference for ACO-SATE and SW-SATE
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 1... Theorem 2... Theorem 3... Theorem 4 (inference under M_Γ)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

[1]

( 1 2 , 1 2 ,1,0) (0a,1 a,1 b,0 b) (0,1, 1 2 , 1

work page
[2]

( 1 2 , 1 2 ,1,0) (1a,0 a,0 b,1 b) (1,0, 1 2 , 1

work page
[3]

( 1 2 , 1 2 ,0,1) (0a,1 a,0 b,1 b) (0,1, 1 2 , 1

work page
[4]

1− Γ−1 Γ + 1 2# ¯τ(t1,t2) i,SW . Averaging over the four treatment assignment cases of (T i1, Ti2), E[Ui,Γ] = 1 4 X t1∈{1,−1} X t2∈{1,−1} E[U (t1,t2) i,Γ ] ≤ 1 4

( 1 2 , 1 2 ,0,1) (1b,0 b,1 a,0 a) ( 1 2 , 1 2 ,1,0) (1,0, 1 2 , 1 2) (1b,0 b,0 a,1 a) ( 1 2 , 1 2 ,0,1) (1,0, 1 2 , 1 2) (0b,1 b,1 a,0 a) ( 1 2 , 1 2 ,1,0) (0,1, 1 2 , 1 2) (0b,1 b,0 a,1 a) ( 1 2 , 1 2 ,0,1) (0,1, 1 2 , 1 2) 37 Taking the expectation conditional on the potential outcomesF, we have E[Vi | F] = E{ 2X j=1 2X k=1 {Z b ijkDijk −(1−Z b ijk)Dij...

work page 2020

[1] [1]

( 1 2 , 1 2 ,1,0) (0a,1 a,1 b,0 b) (0,1, 1 2 , 1

work page

[2] [2]

( 1 2 , 1 2 ,1,0) (1a,0 a,0 b,1 b) (1,0, 1 2 , 1

work page

[3] [3]

( 1 2 , 1 2 ,0,1) (0a,1 a,0 b,1 b) (0,1, 1 2 , 1

work page

[4] [4]

1− Γ−1 Γ + 1 2# ¯τ(t1,t2) i,SW . Averaging over the four treatment assignment cases of (T i1, Ti2), E[Ui,Γ] = 1 4 X t1∈{1,−1} X t2∈{1,−1} E[U (t1,t2) i,Γ ] ≤ 1 4

( 1 2 , 1 2 ,0,1) (1b,0 b,1 a,0 a) ( 1 2 , 1 2 ,1,0) (1,0, 1 2 , 1 2) (1b,0 b,0 a,1 a) ( 1 2 , 1 2 ,0,1) (1,0, 1 2 , 1 2) (0b,1 b,1 a,0 a) ( 1 2 , 1 2 ,1,0) (0,1, 1 2 , 1 2) (0b,1 b,0 a,1 a) ( 1 2 , 1 2 ,0,1) (0,1, 1 2 , 1 2) 37 Taking the expectation conditional on the potential outcomesF, we have E[Vi | F] = E{ 2X j=1 2X k=1 {Z b ijkDijk −(1−Z b ijk)Dij...

work page 2020