A cautious use of auxiliary outcomes for decision-making in randomized clinical trials
Pith reviewed 2026-05-23 05:58 UTC · model grok-4.3
The pith
A Bayesian framework lets randomized trials use auxiliary outcomes like response rates for decisions while exactly controlling type I error without assuming how they relate to the primary outcome.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
We develop a Bayesian decision-theoretic framework that uses both primary and auxiliary outcomes for interim and final decision-making. The framework allows investigators to control standard frequentist operating characteristics, such as the type I error rate, and can be used with auxiliary outcomes from emerging technologies, such as circulating tumor assays. False positive rates and other frequentist operating characteristics are rigorously controlled without any assumption about the concordance between primary and auxiliary outcomes. Algorithms implement the approach and show that incorporating auxiliary information can lead to relevant efficiency gains.
What carries the argument
Bayesian decision-theoretic rules that construct interim and final decisions to enforce exact or approximate control of type I error and related operating characteristics for arbitrary joint distributions of primary and auxiliary outcomes.
If this is right
- Efficiency gains become available in trial duration or sample size when auxiliary data arrive earlier than the primary endpoint.
- The same control holds when auxiliary data come from new technologies such as circulating tumor assays.
- Algorithms exist that produce the required decision rules while preserving the frequentist guarantees.
- No modeling step that links the auxiliary outcome to the primary outcome is needed to keep error rates in check.
Where Pith is reading between the lines
- The separation of decision rules from any dependence model could let the same machinery apply to other pairs of delayed and fast endpoints in non-oncology settings.
- Regulatory acceptance might hinge on whether the algorithms can be pre-specified and verified by simulation for each concrete trial design.
- If the control holds in finite samples, the method could reduce the ethical cost of continuing trials after strong auxiliary signals appear.
Load-bearing premise
That decision rules and algorithms exist which maintain the target type I error rate for every possible joint distribution of the primary and auxiliary outcomes.
What would settle it
A concrete joint distribution of primary and auxiliary outcomes, together with a data-generating process under the null, for which the proposed decision procedure yields a type I error rate above the nominal level.
read the original abstract
Clinical trials often collect data on multiple outcomes, such as overall survival (OS), progression-free survival (PFS), and response to treatment (RT). In most cases, however, study designs only use primary outcome data for interim and final decision-making. In several disease settings, clinically relevant outcomes, for example OS, become available years after patient enrollment. Moreover, the effects of experimental treatments on OS might be less pronounced compared to auxiliary outcomes such as RT. We develop a Bayesian decision-theoretic framework that uses both primary and auxiliary outcomes for interim and final decision-making. The framework allows investigators to control standard frequentist operating characteristics, such as the type I error rate and can be used with auxiliary outcomes from emerging technologies, such as circulating tumor assays. False positive rates and other frequentist operating characteristics are rigorously controlled without any assumption about the concordance between primary and auxiliary outcomes. We discuss algorithms to implement this decision-theoretic approach and show that incorporating auxiliary information into interim and final decision-making can lead to relevant efficiency gains according to established and interpretable metrics.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript develops a Bayesian decision-theoretic framework for incorporating both primary and auxiliary outcomes into interim and final decision-making in randomized clinical trials. It asserts that the framework controls frequentist operating characteristics such as type I error rate (and others) for arbitrary joint distributions of the outcomes, without requiring any concordance assumptions between primary and auxiliary, while also yielding efficiency gains; algorithms for implementation are discussed.
Significance. If the central claim of exact or approximate frequentist control without concordance assumptions were to hold while still allowing nontrivial use of auxiliary data, the result would be significant for clinical trial design, particularly in settings where auxiliary outcomes (e.g., from circulating tumor assays) become available earlier than the primary. The paper references established metrics for efficiency gains and applicability to emerging technologies.
major comments (1)
- [Abstract] Abstract: The assertion that 'False positive rates and other frequentist operating characteristics are rigorously controlled without any assumption about the concordance between primary and auxiliary outcomes' cannot hold for nontrivial use of auxiliary data. For any decision rule whose rejection indicator depends on the realized auxiliary value at a fixed primary outcome y_p in the null support, an adversary can always choose a conditional distribution of the auxiliary given y_p such that P(reject | y_p) = 1, driving the type I error above the nominal level while preserving the null marginal on the primary. The only rules achieving sup_P(type I error) ≤ α over all joints are those that are deterministic functions of the primary outcome alone. This directly undermines the framework's stated ability to use auxiliary outcomes while maintaining rigorous control.
minor comments (1)
- The abstract states that 'We discuss algorithms to implement this decision-theoretic approach and show that incorporating auxiliary information... can lead to relevant efficiency gains' but the provided text contains no explicit algorithm descriptions, pseudocode, or simulation results that would allow verification of the claimed control or gains.
Simulated Author's Rebuttal
We thank the referee for their careful reading and for raising this important point about the scope of our frequentist control claims. We address the major comment below.
read point-by-point responses
-
Referee: [Abstract] Abstract: The assertion that 'False positive rates and other frequentist operating characteristics are rigorously controlled without any assumption about the concordance between primary and auxiliary outcomes' cannot hold for nontrivial use of auxiliary data. For any decision rule whose rejection indicator depends on the realized auxiliary value at a fixed primary outcome y_p in the null support, an adversary can always choose a conditional distribution of the auxiliary given y_p such that P(reject | y_p) = 1, driving the type I error above the nominal level while preserving the null marginal on the primary. The only rules achieving sup_P(type I error) ≤ α over all joints are those that are deterministic functions of the primary outcome alone. This directly undermines the framework's stated ability to use auxiliary outcomes while maintaining rigorous control.
Authors: We agree with the referee that the argument is correct: uniform control of the type I error (i.e., sup over all joints with fixed null marginal on the primary) cannot be achieved by any nontrivial dependence on the auxiliary outcome. Our intended meaning of 'without any assumption about the concordance' was that the procedure does not require the auxiliary and primary to be positively associated or to have concordant treatment effects; the control was meant to hold under a fixed but arbitrary joint distribution that is modeled or estimated without imposing concordance. However, the abstract phrasing is imprecise and can be read as claiming distribution-free control over arbitrary joints, which the referee's counterexample shows is impossible. We will revise the abstract, the description of the framework, and related sections to clarify that frequentist control holds for a given joint distribution (without requiring concordance within that distribution) and that the guarantee is not uniform over all possible joints. This revision will be incorporated in the next version of the manuscript. revision: yes
Circularity Check
No significant circularity; control claimed for arbitrary joints without reduction to fit or self-citation
full rationale
The paper's central claim is that a Bayesian decision-theoretic framework controls type I error and other frequentist characteristics for any joint distribution of primary and auxiliary outcomes, with no assumptions on concordance. The provided abstract and reader's summary contain no equations, fitted parameters, or self-citations that reduce this control to a post-hoc definition or construction. No patterns of self-definitional claims, fitted inputs called predictions, or ansatz smuggling appear. The derivation is presented as holding by the framework's design across arbitrary joints, making it self-contained against external benchmarks. A potential logical tension with the skeptic attack (that non-trivial auxiliary dependence would violate the supremum bound) concerns correctness rather than circularity in the derivation chain.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.