Nonparametric bounds for vaccine effects in randomized trials

Rachel Axelrod, Uri Obolski, Daniel Nevo · 2025

Read Nonparametric bounds for vaccine effects in randomized trials by Rachel Axelrod, Uri Obolski, Daniel Nevo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-25296.txt.

Introduction

Rachel Axelrod, Uri Obolski, and Daniel Nevo study vaccine trials. Their 2025 statistics paper tackles broken blinding in trials. Blinding hides who got vaccine or placebo. When blinding breaks, behavior can change outcomes. People may act differently if they guess their status. Then measured efficacy mixes immune and behavior effects.

Recent work proposed new causal targets for this problem. Those targets need strong assumptions to identify exactly. One assumption rules out unmeasured common causes. Personality traits might violate that assumption in practice. The three authors relax that strong assumption. They derive nonparametric bounds for vaccine effects instead. Bounds give a range rather than one number. They build bounds with linear programming and monotonicity. They test the bounds on synthetic data. One example draws on a COVID-19 trial.

Catalog record

CREATORRachel Axelrod, Uri Obolski, Daniel Nevo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE140 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-25296

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Nonparametric bounds for vaccine effects in randomized trials Rachel Axelrod, Uri Obolski, Daniel Nevo Abstract Vaccine randomized trials are typically designed to be blinded, ensuring that the estimated vaccine efficacy (VE) reflects the immunological effect of the vaccine. When blinding is broken, however, the estimated VE reflects not only the immunological effect but also behavioral effects stemming from participants’ awareness of their treatment status. Recent work has proposed alternative causal estimands to the standard VE to address this issue, but their point identification results require a strong assumption: the absence of unmeasured common causes of infection risk and participants’ belief about whether they received the vaccine. Personality traits, for example, may plausibly violate this assumption. We relax this assumption and derive nonparametric causal bounds for different types of VE. We construct these bounds using two approaches: linear programming-based and monotonicity-based methods. We further consider several possible causal structures for vaccine trials and show how the nonparametric bounds differ across these scenarios. Finally, we illustrate the performance of the proposed bounds using fully synthetic data and a semi-synthetic data example based on a COVID-19 vaccine trial. Keywords: Partial identification, nonparametric bounds, randomized trials, causal inference 1 Introduction In vaccine trials, broken blinding , i.e., when participants become aware of their treatment assignment, has recently been recognized as a factor that may cause the immunological effect, typically the target of the trial, to be non-identifiable ( Gabriel et al., 2025 ) . Moreover, the estimated vaccine efficacy (VE) in such trials has been recently shown to be a biased estimator of the true immunological effect ( Stensrud et al., 2024 ; Obolski et al., 2024 ) . In the presence of broken blinding, the estimated VE reflects not only the immunological effect but also behavioral changes induced by participants’ awareness of their treatment status. For example, after COVID-19 vaccination, individuals who felt more protected were found to increase social contacts, reduce mask use, and attend larger gatherings, potentially elevating their exposure to SARS-CoV-2 ( Trogen and Caplan, 2021 ; Buckell et al., 2023 ) . Since examples of broken blinding in vaccine trials have been documented ( Lazarus et al., 2021 ) , it is crucial to discuss how the true VE can be estimated in such trials. As reported in Lazarus et al., 2021 , a possible reason for the blinding process to fail is that the active vaccine causes more local reaction than placebo, which could unmask the active vaccine versus placebo allocation. Another reason is that the treatment may be effective and reveal itself through noticeable improvements, allowing participants to deduce their actual assignment ( Boutron et al., 2005 ) . Alternative causal estimands have recently been proposed to study behavioral and immunological effects arising from broken blinding in vaccine trials. One such estimand corresponds to the VE that would be observed had participants known whether they were vaccinated or not ( Stensrud et al., 2024 ; Obolski et al., 2024 ) . Importantly, such an estimand is not only relevant for broken-blinding scenarios but also reflects vaccine effectiveness in real-world settings, where people are aware of their vaccination status and may adjust their behavior accordingly. A recent point-identification strategy of the suggested estimands builds on the idea of measuring trial participants belief about their vaccination status. However, these point-identification results require an arguably strong assumption: that there are no unmeasured common causes of infection risk and of belief about the treatment received. A key example of such a common cause in vaccine studies is personality traits, which may violate this assumption. For instance, optimism may lead participants to believe they received the vaccine even when they did not (“wishful thinking”, Bang, 2016 ), and the same optimistic disposition may also make them more outgoing, thereby increasing their chances of exposure to infectious agents. To avoid making the assumption of no unmeasured common causes, we develop in this paper nonparametric causal bounds for different types of VE. Nonparametric bounds have been proposed in a variety of contexts. Robins, 1989 and Manski, 1990 independently derived the first informative nonparametric bounds on causal effects in studies with imperfect compliance, where the standard randomization assumption cannot be invoked. Later, Horowitz and Manski, 2000 derived nonparametric bounds for randomized experiments with missing data, relying only on observed confounders and making no assumptions about the missingness mechanism. Other contexts in which nonparametric bounds have been used include, for example, mediation analysis ( Robins and Richardson, 2010 ; VanderWeele, 2010 ; Gabriel et al., 2023a ) and outcome-dependent sampling with unmeasured confounding ( Gabriel et al., 2022 ) . We construct nonparametric bounds using two different approaches: linear programming (LP)-based and monotonicity-based methods. The LP-based approach originates in the seminal paper of Balke and Pearl, 1994 , whose framework enabled bounding the average treatment effect in studies with imperfect compliance. They recognized that deriving such bounds is equivalent to solving an LP optimization problem, with the target causal parameter and the constraints implied by observed probabilities expressed as linear combinations of the same underlying parameters. In the binary treatment and binary outcome setting, Balke and Pearl further derived closed-form symbolic solutions, providing analytical expressions without the need for iterative optimization. Their results were later extended to the more general instrumental variables case (see Swanson et al., 2018 , for a review). The original Balke–Pearl bounds can be further tightened by imposing monotonicity assumptions ( Balke and Pearl, 1994 ; Kuroki and Cai, 2008 ) , and more recent work has generalized the framework to accommodate multiple scenarios ( Gabriel et al., 2023b ; Sachs et al., 2023 ) . While LP-based bounds are sharp (i.e., the tightest possible bounds under the stated assumptions), they can be wide in practice, especially when the data provide limited information about the joint distribution of unmeasured variables. As an alternative, we consider a second approach based on monotonicity assumptions, namely that both the belief variable and the outcome are monotonically related to the unmeasured variable.. To this end, we build on ideas similar to those of VanderWeele, 2008 , who established sign determination results for causal effects in the presence of unmeasured confounding in observational studies. We consider several possible causal structures for a vaccine trial. Each structure consists of particular assumptions about the relationships between key trial variables, which in turn yield different nonparametric bounds for the different VEs. We begin with a baseline setting that assumes no unmeasured common causes between belief and risk of infection, thereby permitting point identification ( Stensrud et al., 2024 ) . Next, we relax this assumption by allowing an unmeasured common cause of belief and risk of infection. We further relax the assumptions by permitting the unmeasured common cause to also directly affect adverse events (AEs). For each structure, we first provide real-world examples, supported by existing studies and then present the corresponding LP-based and monotonicity-based bounds for the different VEs. As in illustrative example, we consider the ENSEMBLE2 trial, a randomized, double-blind, placebo-controlled study evaluating a COVID-19 vaccine in adults ( \geq 18 years) across multiple international sites ( Hardt et al., 2022 ) . In a prespecified safety subset, solicited local and systemic AEs occurred in 57.3% of vaccine

Read the complete text