Targeted maximum likelihood estimation of vaccine effectiveness and immune correlates in test-negative design studies with missing data
Leah I. B. Andrews*, Lars van der Laan, Peter B. Gilbert · 2026
Read Targeted maximum likelihood estimation of vaccine effectiveness and immune correlates in test-negative design studies with missing data by Leah I. B. Andrews*, Lars van der Laan, Peter B. Gilbert (2026) 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-2605-21793.txt.
Introduction
Leah Andrews and co-authors present a targeted likelihood approach for vaccine effectiveness and immune correlates in test-negative studies with missing data. The test-negative design enrolls people seeking tests and compares cases by vaccination or immune-marker level, cutting healthcare-seeking confounding but not other confounding or gaps. Their method uses a semiparametric logistic model aiming at a causal conditional risk ratio, yielding an efficient estimator with flexible control. Plasmode simulations test finite-sample behavior, and COVID-19 vaccine results and antibody markers are assessed from Moderna cohorts.
Catalog record
| CREATOR | Leah I. B. Andrews*, Lars van der Laan, Peter B. Gilbert |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 108198 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.21793
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Targeted maximum likelihood estimation of vaccine effectiveness and immune correlates in test-negative design studies with missing data Leah I. B. Andrews*, Lars van der Laan, Peter B. Gilbert Abstract The test-negative design (TND) is a resource-efficient observational study design that can assess vaccine effectiveness and exposure-proximal immune correlates of disease. The TND enrolls symptomatic individuals seeking diagnostic testing and compares case status by an exposure variable, such as vaccination status or immune marker level, that is measured at testing. While the TND reduces confounding by healthcare-seeking behavior, other sources of confounding may remain. TND studies may also have missing data in the exposure variable due to incomplete records or two-phase sampling designs. We present a targeted maximum likelihood estimation approach involving a semiparametric logistic regression model that targets a causal conditional risk ratio of symptomatic disease in the healthcare-seeking population. Under causal and missing at random assumptions, our method produces an efficient, asymptotically linear estimator that provides flexible, data-driven confounding control and valid causal inference when analyzing TND studies with missing exposure variable data. We evaluate our method’s finite sample properties using plasmode simulations of a two-phase TND immune correlates study. We also apply our method to assess COVID-19 vaccine effectiveness and antibody marker correlates of COVID-19 from TND study cohorts derived from the Moderna Coronavirus Efficacy phase 3 trial. Keywords: test-negative design, targeted maximum likelihood estimation, vaccine effectiveness, immune correlates, missing data, two-phase sampling * Corresponding author : Leah I. B. Andrews (email address: [email protected], ORCID iD: 0000-0002-3418-9468). 1 Introduction The test-negative design (TND) is a practical observational study design that has been used to monitor post-marketing vaccine effectiveness against pneumococcal disease, influenza, rotavirus, cholera, pertussis, and most recently, COVID-19 [ 1 , 2 , 3 , 4 , 5 ] . The TND enrolls individuals who obtain diagnostic testing for a target pathogen and meet a symptom definition. Vaccine effectiveness is assessed by comparing vaccination status between individuals who test positive (cases) and individuals who test negative (noncases), after adjusting for potential confounders [ 6 , 7 ] . TND studies are resource-efficient because they recruit cases and noncases identically, collect participants’ information in one visit, and obtain a high proportion of cases. The TND also reduces bias from healthcare-seeking behavior (i.e., seeking medical attention when ill). Individuals with healthcare-seeking behavior are more likely to obtain frequent or intensive healthcare services that increase their probability of being diagnosed, engaging in risk-reducing behaviors (e.g., hand-washing, vaccination), and participating in studies, which induces bias in observational studies since healthcare-seeking behavior is difficult to measure [ 7 , 8 ] . The TND reduces this bias by limiting enrollment to symptomatic individuals who obtain testing and assuming they have similar healthcare-seeking behavior [ 6 , 7 ] . A novel application of the TND is to study exposure-proximal immune correlates of disease (i.e., studying the association of immune markers present at the time of potential pathogen exposure with a clinical disease endpoint) [ 9 , 10 , 11 , 12 , 13 ] . Exposure-proximal immune correlates can provide greater insights about immunological mechanisms of protection than immune markers measured at fixed time points and can help validate surrogate endpoints for inferring vaccine efficacy. In a TND immune correlates study, a participant contributes their specimen sample (which can be assayed later for immune markers), diagnostic test, and additional covariates in a single visit. To interpret immune markers as exposure-proximal immune correlates, TND studies must either assume that participants’ immune marker levels are identical at testing and potential pathogen exposure or account for rapid memory immune responses following natural infection [ 9 , 13 ] . The TND is especially advantageous for studying rare severe outcomes and cell-mediated immune markers that are rarely studied in randomized, controlled clinical trials (RCTs) and prospective cohort studies because their measurement requires resource-intensive storage of peripheral blood mononuclear cells [ 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 ] . Despite its advantages, the TND is an observational study design that is subject to confounding from characteristics like age, comorbidities, SARS-CoV-2 infection history, vaccination history, region, and calendar date [ 7 , 4 , 2 , 21 ] . To account for confounding, most TND studies fit an ordinary logistic regression adjusted for confounders [ 22 , 23 , 24 , 25 ] , which assumes a parametric model that may induce bias if misspecified. Alternatively, some TND studies match by covariates and fit a conditional logistic regression [ 22 , 26 , 27 ] , which may not adequately control for confounding, may be less efficient than covariate adjustment through regression, and is more resource-intensive when studying multiple symptom definitions [ 28 , 29 ] . Only recently have TND methods incorporated machine-learning methods to provide flexible confounding control [ 30 ] . TND studies may also have missing data in the exposure variable, which must be considered to ensure generalizable results. Missing vaccination status may arise in a TND vaccine effectiveness study from lack of vaccine records or unverified self-report [ 31 , 32 , 3 ] . Most TND studies require known vaccination status in the study eligibility criteria and conduct complete-case analyses [ 31 , 33 ] , which may induce bias if individuals missing vaccination status have different characteristics than those with vaccination status observed [ 34 ] . Since measuring immune markers is expensive and labor-intensive, TND immune correlates studies may apply two-phase sampling, a common cost-effective and efficient design in immune correlate studies [ 35 , 28 , 36 , 37 , 38 ] . Under this design, demographic and clinical information, diagnostic tests, and specimen samples are collected from all enrolled individuals at testing (i.e., phase one). Once test results are available, immune markers are measured from a subset of phase one individuals sampled into phase two with sampling probabilities dependent on case status and covariates. Although the missingness framework is known, analyses must account for the missingness patterns to obtain an unbiased estimate. With increasing reliance on the TND for post-marketing vaccine evaluation, it is essential to develop more robust statistical methods for TND analyses than ordinary logistic regression. Targeted maximum likelihood estimation approaches use semiparametric theory and machine-learning algorithms to produce less biased estimates under fewer assumptions than parametric models [ 39 , 40 ] . These approaches also respect model constraints and may have better finite sample properties than one-step estimation and estimating equation approaches [ 41 , 42 ] . van der Laan and Gilbert [ 43 ] developed a targeted maximum likelihood estimator (TMLE) to study strain-specific relative COVID-19 vaccine efficacy in case-only designs. This method assumes a semiparametric logistic regression model that allows for flexible, data-driven confounding adjustment and missing cause of failure information (e.g., viral genotype) in test-positive cases. In this work, we adapt the TMLE developed by van der Laan and Gilbert [ 43 ] to a TND setting to study vaccine effectiveness and exposure-proximal immune correlates. We extend the identifying assumptions to target a causal conditional risk ratio of symptomatic disease in the healthcare-seeking population and account for missing data in the exposure variable (vaccination status or exposure-proximal immune marker level). Our