pith. sign in

arxiv: 2604.13276 · v2 · pith:JQN5T4FGnew · submitted 2026-04-14 · 📊 stat.ME · math.ST· stat.TH

Addressing Confounding by Indication Through (Un)Measured Centre Characteristics in Learn-As-you-GO(LAGO) Trials

Minh Thu Bui , Christopher T. Longenecker , Ante Bing , Donna Spiegelman , Allison R. Webel , Hayden B. Bosworth , Judith J. Lok This is my paper

Pith reviewed 2026-05-20 23:57 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.TH
keywords LAGO designfixed center effectsconfounding by indicationadaptive clinical trialsmulticomponent interventionsgeneralized linear modelslogistic regression
0
0 comments X

The pith

Fixed center effects in LAGO trials remove bias from both measured and unmeasured site characteristics that confound intervention packages and outcomes.

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

The paper establishes that in Learn-As-you-Go adaptive trials, where multicomponent intervention packages are updated across stages and centers join multiple stages, adding fixed effects for each center to the outcome regression models blocks all confounding paths from center traits. This approach preserves the asymptotic consistency and normality of estimators for the intervention effect without any need to measure or model the specific confounding center characteristics explicitly. The theory covers both continuous outcomes in generalized linear models and binary outcomes in logistic regression, and it remains valid even when the number of centers is small. A sympathetic reader would care because large-scale implementation trials routinely face this form of confounding by indication, which can otherwise invalidate conclusions about which package produces the best results.

Core claim

In LAGO trials, center characteristics can predict both the chosen intervention package and the observed outcomes, creating confounding by indication. Including fixed center effects in the regression model for the outcome ensures that the estimators for the intervention effect are consistent and asymptotically normal, regardless of whether the center traits are measured or unmeasured. The same fixed-effects construction supplies valid confidence intervals, hypothesis tests for the overall intervention effect, and a constrained optimization procedure that identifies the lowest-cost package achieving a target mean outcome.

What carries the argument

fixed center effects included as indicators in the generalized linear model or logistic regression for the outcome, which block all confounding paths from center characteristics

If this is right

  • Point and interval estimators for the intervention effect remain consistent and asymptotically normal.
  • Hypothesis tests for the overall intervention effect achieve correct size and power.
  • Constrained optimization yields the intervention package that minimizes cost while meeting a target outcome mean.
  • The same guarantees hold for both continuous and binary outcomes and for small numbers of centers.

Where Pith is reading between the lines

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

  • Trial designers could run LAGO studies across multiple sites without collecting detailed data on every possible center trait.
  • The fixed-effects device might be adapted to other sequential or cluster-randomized adaptive designs that face similar site-level confounding.
  • With few centers the method still works, suggesting it is practical for trials that cannot recruit dozens of sites.

Load-bearing premise

The outcome regression model is correctly specified when it conditions on the intervention package and the fixed center indicators, so that these terms fully capture the conditional mean.

What would settle it

In data or simulations that contain a strong unmeasured center characteristic affecting both package assignment and outcome, the estimated intervention effect after adding center fixed effects would differ materially from the effect obtained when that characteristic is also measured and included.

Figures

Figures reproduced from arXiv: 2604.13276 by Allison R. Webel, Ante Bing, Christopher T. Longenecker, Donna Spiegelman, Hayden B. Bosworth, Judith J. Lok, Minh Thu Bui.

Figure 1
Figure 1. Figure 1: Distribution of ∆SBP (left) and residuals [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Total (top) and marginal (bottom) costs of the cubic cost function for each intervention component, component [PITH_FULL_IMAGE:figures/full_fig_p057_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Total (top) and marginal (bottom) costs of the cubic cost function for each intervention component, nurse [PITH_FULL_IMAGE:figures/full_fig_p062_3.png] view at source ↗
read the original abstract

The Learn-As-you-Go (LAGO) design is an adaptive clinical trial design that allows modifications to multicomponent intervention packages across stages. Centers participate in more than one stage, as is common in large-scale implementation trials. In LAGO trials, center characteristics may act as confounders, predicting both the intervention package and the outcomes. We extend the LAGO theory by introducing fixed center effects to control for confounding by indication through measured and unmeasured center characteristics. Conditioning on center characteristics by including fixed center effects ensures asymptotic results hold without requiring explicit characterization of unmeasured confounders. Our methods apply even with small numbers of centers. LAGO theory is established for continuous outcomes following a generalized linear model and binary outcomes following a logistic regression model, unifying theory across outcome types. Point- and interval estimators are derived, and consistency and asymptotic normality are established. Valid hypothesis tests for the overall intervention effect are provided, and the optimal intervention package minimizing cost subject to a target outcome mean is obtained via constrained optimization.

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.

Referee Report

1 major / 1 minor

Summary. The paper extends Learn-As-you-Go (LAGO) adaptive trial designs for multicomponent interventions by incorporating fixed center effects to control for confounding by indication arising from measured and unmeasured center characteristics. Centers participate across stages, and the extension derives point and interval estimators under generalized linear models for continuous outcomes and logistic regression for binary outcomes. It claims to establish consistency, asymptotic normality, and valid tests for the overall intervention effect, while also solving a constrained optimization problem to find the cost-minimizing intervention package that achieves a target outcome mean. The methods are asserted to apply even when the number of centers is small and fixed.

Significance. If the asymptotic results are valid, the work offers a unified framework for handling center-level confounding in adaptive implementation trials without requiring explicit modeling or measurement of unmeasured center traits. This could improve the reliability of effect estimation and optimization in large-scale trials where centers are reused across stages, particularly when full characterization of confounders is impractical.

major comments (1)
  1. [asymptotic theory for binary outcomes] The section deriving consistency and asymptotic normality for the logistic model (around the statements following the abstract's claim that 'consistency and asymptotic normality are established' and 'Our methods apply even with small numbers of centers'): the fixed-effects logistic regression estimator is subject to the incidental parameters problem when the number of centers K is treated as fixed while total sample size grows only through larger per-center sample sizes n_k. The center intercepts are inconsistent, and this inconsistency generally propagates to bias the intervention package coefficients even under correct conditional-mean specification. The manuscript does not appear to impose additional rate conditions (e.g., n_k growing sufficiently fast relative to K) or switch to conditional likelihood, so the claimed consistency does not follow for the small-K regime highlighted in §
minor comments (1)
  1. [Abstract] The abstract states that the approach 'unifies theory across outcome types,' but the manuscript would benefit from an explicit comparison of the differing regularity conditions or proof strategies required for the GLM versus logistic cases.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We are grateful to the referee for their detailed and insightful comments, which have helped us improve the clarity of our manuscript on extending LAGO trials with fixed center effects. Below, we provide a point-by-point response to the major comment.

read point-by-point responses

  1. Referee: [asymptotic theory for binary outcomes] The section deriving consistency and asymptotic normality for the logistic model (around the statements following the abstract's claim that 'consistency and asymptotic normality are established' and 'Our methods apply even with small numbers of centers'): the fixed-effects logistic regression estimator is subject to the incidental parameters problem when the number of centers K is treated as fixed while total sample size grows only through larger per-center sample sizes n_k. The center intercepts are inconsistent, and this inconsistency generally propagates to bias the intervention package coefficients even under correct conditional-mean specification. The manuscript does not appear to impose additional rate conditions (e.g., n_k growing sufficiently fast relative to K) or switch to conditional likelihood, so the claimed consistency does not follow

    Authors: We appreciate the referee raising this important point regarding potential bias from the incidental parameters problem in fixed-effects logistic regression. However, we respectfully disagree that this issue arises under the asymptotic regime considered in the manuscript. We treat the number of centers K as fixed (including the small-K case highlighted), while allowing the total sample size N = sum n_k to grow through increasing per-center sizes n_k → ∞. With K fixed, the total number of parameters (the K center intercepts plus the finite-dimensional intervention coefficient vector) does not grow with N. Under standard regularity conditions for maximum likelihood estimation in logistic regression (e.g., the observed information matrix being positive definite and sufficient within-center variation in the intervention packages), both the intervention coefficients and the center intercepts are consistent, and the estimator is asymptotically normal. The incidental parameters problem occurs when the number of nuisance parameters (here, centers) increases with sample size, which is not the case when K is held fixed. No additional rate conditions relating n_k to K are required. To improve clarity and directly address this concern, we have revised the relevant sections to explicitly state the fixed-K asymptotic regime and to explain why the standard MLE theory applies without further restrictions. revision: partial

Circularity Check

0 steps flagged

No circularity; derivation extends prior LAGO with independent fixed-effects estimators and standard asymptotics

full rationale

The paper introduces fixed center effects into the LAGO framework to block confounding paths, then derives point/interval estimators and proves consistency/asymptotic normality for GLM and logistic models under correct conditional-mean specification. These steps rely on standard fixed-effects GLM theory rather than redefining quantities in terms of previously fitted values from the same data. No self-definitional loops, fitted inputs relabeled as predictions, or load-bearing self-citations that reduce the central claims to unverified priors. The extension is self-contained against external statistical benchmarks for fixed-effects models with fixed K and growing per-center sample sizes.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on standard GLM and logistic regression assumptions plus the fixed-effects blocking property; no new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Outcomes follow a generalized linear model (continuous) or logistic regression (binary) conditional on intervention package and center indicators.
    Invoked to unify theory and derive asymptotic results across outcome types.

pith-pipeline@v0.9.0 · 5738 in / 1255 out tokens · 31111 ms · 2026-05-20T23:57:42.007349+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

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

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

2 extracted references · 2 canonical work pages

  1. [1]

    Y (1) ij − βA βℓ T A(1) j ℓ(1) j !# + KX k=2 J (k) X j=1 n(k) jX i=1 A(k,nk−) j ℓ(k) j !

    Thus, Var(U 2 2,n)→0. Hence, by Chebyshev’s inequality,U 2 2,n P →0. In conclusion, it follows that √n U(β ∗) has the same asymptotic distribution as U1 2,n = 1√n   J (1) X j=1 n(1) jX i=1 A(1) j ℓ(1) j ! ϵ(1) ij + J (2) X j=1 n(2) jX i=1 a(2) j ℓ(2) j ! ϵ(2) ij   , so that √n U(β ∗) converges to a normal distribution with mean 0 and variance J (1) ...

    work page 2000

  2. [2]

    29 Figure 2: Total (top) and marginal (bottom) costs of the cubic cost function for each intervention component, component 1 (left) and component 2 (right)

    The cubic cost function was calibrated such that its average marginal cost over the feasible intervention range approximately equals the constant marginal cost of the linear function. 29 Figure 2: Total (top) and marginal (bottom) costs of the cubic cost function for each intervention component, component 1 (left) and component 2 (right). D Additional res...

    work page