2026 年的期刊文章 · 第 25 页

共 1756 篇,按发表时间排,新的在前,每页 40 篇。这一页是 2026-04-27 到 2026-04-14。

按期刊、按主题看 →

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 页

SiM High-Dimensional · 发表于 2026-04-27,收录于 2026-05-01

Penalized GEE for Complex Carry‐Over in Repeated‐Measures Crossover Designs

Nelson Alirio Cruz, Oscar Orlando Melo, Kalliopi Mylona

摘要

Crossover designs are commonly employed in clinical and behavioral research, yet the statistical models used to analyze them often rely on unrealistic assumptions–either ignoring carry‐over effects or modeling them as simple and homogeneous across treatment sequences. However, carry‐over effects are frequently complex, varying by treatment order and interaction, and until now, no statistical methodology had been formally established to estimate such complex effects. This paper introduces a penalized semiparametric Generalized Estimating Equations (GEE) approach designed to estimate first order complex carry‐over effects in crossover designs with repeated measurements. We first derive identifiability conditions under which complex carry‐over effects become estimable. We then provide theoretical guarantees–building on an extension of the sandwich variance formula–showing that the proposed penalized estimator achieves asymptotic normality for the functional components and shrinks negligible carry‐over effects toward zero, thereby enabling their practical identification. Through simulation studies and application to real data, the methodology demonstrates improved estimation accuracy when complex carry‐over effects are present, outperforming models that assume simple or no carry‐over. This work represents the first rigorous and generalizable approach for modeling complex carry‐over effects in repeated‐measures crossover designs.

SiM Survival Analysis Causal Inference · 发表于 2026-04-27,收录于 2026-05-17

Mediation Analysis of Failure Time Data Under Interval Censoring

Yanpeng Shi, Shuwei Li

摘要

Motivated by disentangling the direct and indirect genetic effects on the Alzheimer's disease progression, we propose a mediation approach for analyzing interval-censored failure time data. Under the counterfactual framework with sequential ignorability, generalized linear and spline-based proportional hazards (PH) models are employed to depict path-specific effects (PSEs) of an exposure on the interested failure time mediated through multiple mediators. In particular, the PSEs are derived on the basis of both survival probability and restricted mean survival time, bypassing the restrictive rare outcome assumption inherent in the PH model. A sieve maximum likelihood method is proposed to estimate all unknown parameters in the PH model. Simulation studies demonstrate the satisfactory empirical performances of the proposed mediation method. A real-world data analysis quantifies the mediation effects of Apolipoprotein E- ϵ $$ \epsilon $$ 4 allele on the risk of developing the Alzheimer's disease.

SiM Other · 发表于 2026-04-27,收录于 2026-05-01

Delayed Exposures and Pre‐Exposure Periods in Self‐Controlled Case Series Studies

Heather Whitaker, Yonas Ghebremichael Weldeselassie, Paddy Farrington

摘要

A key assumption of the self‐controlled case series (SCCS) method is that exposures should not depend on the event of interest. However, treatments such as vaccines may be deferred after an adverse health event. One suggestion to handle such delayed exposures is to include a pre‐exposure window in the SCCS model. We study the impact of such adjustments and of exposure deferment on the SCCS relative incidence estimates. We obtain explicit results in a simplified setting, and investigate more realistic scenarios by simulation. We develop some practical recommendations for sensitivity analyses: when the delayed exposures remain within the observation period, no adjustment is needed. When exposures are delayed beyond the end of the observation period, an adjustment may be required. In some circumstances the SCCS model for event‐dependent exposures should be used rather than the standard SCCS model. These options are illustrated with three practical examples.

JRSSB Causal Inference Latent Variables & Mixtures · 发表于 2026-04-27

Causal K-means clustering

Kwangho Kim, Jisu Kim, Edward H Kennedy

摘要

Causal effects are often characterized at the population level, which can mask important heterogeneity across latent subgroups. Since the subgroup structure is unknown, identifying and evaluating subgroup specific effects is substantially more challenging than standard population level analysis. We address this problem by proposing Causal k-Means Clustering, a framework that uses k-means clustering ideas to recover unknown subgroup structure from individual level causal contrasts. The problem differs fundamentally from ordinary clustering because the objects to be clustered are unknown counterfactual functions. We first study a simple plug in estimator that is readily implemented with standard algorithms and establish its rate of convergence. We then develop a bias-corrected estimator using semiparametric efficiency theory and double machine learning and show that it attains fast root-n rates and asymptotic normality in large nonparametric models. The proposed methods are especially useful for modern outcome wide studies with multiple treatment levels, and the framework extends naturally to clustering based on more general pseudo-outcomes, including partially observed outcomes and other unknown functionals. We study finite sample performance in simulations and illustrate the method with an application to mobile supported self-management for chronic low back pain.

SiM Genetics & Omics Bayesian Methods Latent Variables & Mixtures · 发表于 2026-04-26,收录于 2026-05-01

Simultaneous Representation Learning of Multi‐Omics and Clinical Outcome Data via a Supervised Knowledge‐Guided Bayesian Factor Model

Qiyiwen Zhang, Changgee Chang, Chong Jin, Li Shen, Qi Long

摘要

With the advent of high‐throughput techniques, multi‐omics data and various clinical outcomes have been collected for a range of diseases. Multi‐omics data play a crucial role in uncovering complex biological processes, yet simultaneous representation learning of such high‐dimensional, heterogeneous multi‐modality data along with clinical outcomes remains limited. To address this gap, we propose a supervised knowledge‐guided Bayesian factor model for integrative analysis of multi‐omics and clinical outcome data. The proposed method simultaneously extracts an informative low‐dimensional representation and predicts one or more clinical outcomes of interest. The two‐level adaptive shrinkage in the novel hierarchical priors allows for the identification of both active modalities and features, resulting in a biologically meaningful structural identification of the high‐dimensional data. Moreover, the method is robust to noisy edges in biological graphs that do not align with ground truth. Finally, the proposed method can handle different data types including both continuous and categorical data. Extensive simulation studies and real data analyses of Alzheimer's disease (AD) data demonstrate the advantages of the proposed approach over existing methods. Notably, our analysis of multi‐omics and imaging phenotype data from ADNI provides meaningful insights into the underlying biological mechanisms of AD.

SiM Survival Analysis Causal Inference · 发表于 2026-04-24,收录于 2026-05-01

Win Ratio as an Effect Size Measure Under Non‐Proportional Hazards: A Comparison With Difference in Restricted Mean Survival

Yuan Wu, Xiaofei Wang, Zhiguo Li

摘要

When the proportional hazards assumption does not hold, the hazard ratio can misrepresent treatment effects in survival analysis. We evaluate the win ratio, originally proposed for prioritizing multiple outcomes, as an effect size measure for a single survival outcome under non‐proportional hazards, and compare it with the difference in restricted mean survival time (RMST). We perform bootstrap‐based inference for the win ratio under both right‐ and interval‐censoring using plug‐in estimators based on nonparametric maximum likelihood estimators or spline‐based sieve maximum likelihood estimators of the survival functions. We also study stratified win ratio to mitigate confounding. Extensive simulations are conducted to assess and compare the performances of the win ratio and the difference in RMST under various types of alternatives encountered in practice. The simulation results show that the win ratio‐based tests outperform RMST‐based tests when treatment benefits arise early, whereas RMST is more sensitive to late‐onset effects, and stratified win ratio maintains nominal type I error in the presence of confounding, unlike unstratified win ratio. As an illustration, we analyze right‐censored and interval‐censored progression‐free survival in patients with multiple myeloma treated with two different regimens. The results of this article support reporting the win ratio, along with the difference in RMST, when the proportional hazards assumption is doubtful, offering complementary clinical interpretability and robustness across censoring mechanisms and treatment effect patterns.

SiM Survival Analysis · 发表于 2026-04-24,收录于 2026-05-01

Sequential Multiple Assignment Randomized Trials Based on Restricted Mean Survival Time

Jianhong Pan, Shijie Yu, Minggang Yin, Yuxuan Yang, Chongyang Duan

摘要

Sequential multiple assignment randomized trials (SMART) designs, which are used to evaluate adaptive treatment strategies (ATSs), involve multiple stages of patient randomization based on intermediate outcomes. While offering greater flexibility and personalization in treatment, these designs are often analytically complex and resource‐intensive—especially when survival outcomes are involved, given the extended follow‐up times and potential violations of proportional hazards assumptions. To address these challenges, we propose a statistical inference framework based on restricted mean survival time (RMST). RMST does not rely on the proportional‐hazards assumption and serves as a robust summary for survival data. The framework includes fixed‐ and dynamic‐weight RMST estimators, their variance–covariance structures, confidence intervals, and hypothesis tests for pairwise and global comparisons among ATSs. We also integrate interim analyses into SMART designs using RMST and develop a type I error‐control method that accommodates the lack of the independent‐increments property. Extensive simulations demonstrate good estimator performance and show that interim designs can reduce either sample size or trial duration. In summary, this study offers an efficient and practical framework for SMART trials with survival outcomes.

SMMR Spatial & Time Series Bayesian Methods · 发表于 2026-04-24

Quantifying the effects of air pollution on respiratory ill health treated in primary care when the locations of the populations at risk are partially unknown

Qiangqiang Zhu, Duncan Lee, Oliver Stoner

摘要

Most air pollution and health studies focus on severe outcomes such as hospitalisations and deaths, overlooking the impact that air pollution may have on non-hospitalised respiratory ill health treated in primary care. This paper presents a new study investigating the effects of NO 2 , PM 10 and PM 2.5 on the prescription rates of respiratory medications in Scotland between 2016 and 2020 at a monthly resolution. To enhance the spatial accuracy of the exposure estimates, air pollution predictions at a 1 km 2 resolution are realigned to General Practioner (GP) surgery patient populations by accounting for where patients are likely to live rather than just where the GP surgery is. A Bayesian spatio-temporal conditional autoregressive model is utilised to account for spatial and temporal dependencies in the data, and this paper proposes two novel spatial neighbourhood matrices to better represent the spatial closeness among the patient populations registered at each GP surgery. These matrices improve model performance in capturing spatial correlation compared to standard distance-based approaches, such as using K -nearest neighbours approach. The results of the study suggest that particulate matter pollution has a significant impact on prescription rates for inhaled corticosteroids that are taken to prevent the symptoms of respiratory ill health, while NO 2 demonstrates no such association.

JASA Spatial & Time Series · 发表于 2026-04-24

Change Point Analysis: Theory and Application

Abhishek Kaul

SiM Bayesian Methods Latent Variables & Mixtures Genetics & Omics · 发表于 2026-04-24,收录于 2026-05-01

A Tutorial on Bayesian Multi‐Study Factor Analysis With Applications in Nutrition and Genomics

Mavis Liang, Blake Hansen, Alejandra Avalos‐Pacheco, Roberta De Vito

摘要

High‐dimensional data are crucial in biomedical research. Integrating such data from multiple studies is a critical process that relies on the choice of advanced statistical models, enhancing statistical power, reproducibility, and scientific insight compared to analyzing each study separately. Factor analysis (FA) is a core dimensionality reduction technique that models observed data through a small set of latent factors. Bayesian extensions of FA have recently emerged as powerful tools for multi‐study integration, enabling researchers to disentangle shared biological signals from study‐specific variability. In this tutorial, we provide a practical and comparative guide to seven advanced Bayesian integrative factor models: Perturbed Factor Analysis (PFA), Bayesian Factor Regression with non‐local spike‐and‐slab priors (MOM‐SS), Subspace Factor Analysis (SUFA), Bayesian Multi‐study Factor Analysis (BMSFA), a variational‐inference implementation of BMSFA (CAVI), Bayesian Latent Analysis through Spectral Training (BLAST), and Bayesian Combinatorial Multi‐study Factor Analysis (Tetris). To contextualize these methods, we also include two benchmark approaches: Standard FA applied to pooled data (Stack FA), and FA applied separately to each study (Ind FA). We evaluate all methods through extensive simulations, assessing computational efficiency, accuracy in estimation of loadings, and the number of factors. To bridge theory and practice, we present a full analytical workflow—with detailed R code—demonstrating how to apply these models to real‐world datasets in nutrition and genomics. This tutorial is designed to guide applied researchers through the landscape of Bayesian integrative factor analysis, offering insights and tools for extracting interpretable, robust patterns from complex multi‐source data. Simulation code, R package “bmfaToolkits” and a user‐friendly guidebook can be found at https://github.com/Mavis‐Liang/Bayesian_integrative_FA_tutorial .

SiM Clinical Trials · 发表于 2026-04-24,收录于 2026-05-01

A Novel Group Sequential Design for Sequential Multiple Assignment Randomized Trial

Xueqing Liang, Shijie Yu, Minggang Yin, Siyu Zhu, Yixin Luo, Yuxuan Yang, Chongyang Duan

摘要

The sequential multiple assignment randomized trial (SMART) design offers an innovative approach for efficiently selecting optimal adaptive treatment strategies (ATSs) from a set of candidate treatment strategies. By incorporating interim monitoring, the group sequential SMART design reduces both the trial duration and sample size while maintaining overall efficiency. However, existing SMART designs that incorporate interim analysis rely on global tests, which are inadequate for addressing the challenges of selecting the optimal subset of ATSs and performing early termination of ineffective strategies. In this study, we propose a novel group‐sequential SMART design to improve the efficiency of clinical trials in identifying the most optimal subset of ATSs in a way that aligns with the nuances of clinical practice. This design allows for the early termination of less efficacious treatments, reducing the required sample size and increasing the likelihood of accurately identifying the optimal ATSs. The effectiveness of the proposed approach is demonstrated through simulation results.

JASA Bayesian Methods · 发表于 2026-04-22,收录于 2026-04-24

Translating Predictive Distributions into Informative Priors

Andrew A. Manderson, Robert J. B. Goudie

摘要

When complex Bayesian models exhibit implausible behaviour, one solution is to assemble available information into an informative prior. Challenges arise as prior information is often only available for the observable quantity, or some model-derived marginal quantity, rather than directly pertaining to the (usually latent) parameters in our model. We propose a method for translating available prior information, in the form of an elicited distribution for the observable or model-derived marginal quantity, into an informative joint prior. Our approach proceeds given a parametric class of prior distributions with as yet undetermined hyperparameters, and minimises the difference between the supplied elicited distribution and corresponding prior predictive distribution. We employ a global, multi-stage Bayesian optimisation procedure to locate optimal values for the hyperparameters. Three examples illustrate our approach: a cure-fraction survival model, where censoring implies that the observable quantity is a priori a mixed discrete/continuous quantity; a setting in which prior information pertains to R2 – a model-derived quantity; and a nonlinear regression model.” *(abstract via OpenAlex)*

JASA Bayesian Methods Nonparametric & Semiparametric · 发表于 2026-04-22,收录于 2026-04-24

Nonparametric Prior Learning in Differential Equation Modeling

Junxiong Jia, Deyu Meng, Zongben Xu, Fang Yao

摘要

This article addresses Bayesian inference related to partial differential equations (PDEs), particularly nonparametric regression constrained by PDEs. To effectively encode prior information, we propose a novel framework that learns a prediction function of the prior distribution from historical training datasets. We introduce hyper-prior and hyper-posterior distributions and derive a generalization error estimate, which accommodates data-dependent priors by extending the concept of differential privacy. Some mild conditions are given to validate the error estimate, where various typical PDEs such as diffusion and Darcy flow equations can be integrated. We thus formulate an infinite-dimensional optimization problem to obtain the point estimate of the hyper-posterior. Numerical examples demonstrate the performance of our proposed method in learning the prediction function of priors. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work. *(abstract via OpenAlex)*

SiM Causal Inference Bayesian Methods Survival Analysis · 发表于 2026-04-22,收录于 2026-05-29

Mediation Analysis With Bayesian Nonlinear Joint Models: Evaluation of the Treatment Causal Pathways Between Tumor Growth Kinetics and Overall Survival

Georgios Kazantzidis, Francois Mercier, Virginie Rondeau

摘要

Understanding the mechanisms through which anti‐cancer treatments influence survival is central to improving drug development and evaluation. In this work, we develop a Bayesian joint modeling framework for mediation analysis to quantify the extent to which tumor size dynamics mediate the effect of treatment on overall survival (OS). Our model integrates a nonlinear longitudinal sub‐model describing tumor growth inhibition (TGI) with a parametric survival model, and supports several biologically motivated link functions between tumor size and survival. We explore alternative parameterizations of the treatment effect on tumor shrinkage and regrowth rates and assess their impact on mediation quantities. The approach is implemented using a Bayesian hierarchical joint model. Marginal effects are computed to characterize the proportion of treatment effect (PTE), natural direct effect (NDE), and natural indirect effect (NIE) over time. Through simulations under varying mediation levels and sample sizes, we evaluate bias, coverage, and identifiability of mediation quantities. Application to the IMBrave150 Phase III study suggests that the mediated proportion of the treatment effect varies over time, indicating that tumor dynamics partially but not fully capture the treatment mechanism of action. This framework provides a robust tool to investigate mediation mechanisms in oncology and may contribute to the validation of tumor size as a surrogate endpoint for survival.

JASA Other · 发表于 2026-04-22,收录于 2026-04-24

Maximum Binomial Likelihood Method for Multivariate Mixture Data

Tao Yu, Jing Qin, Pengfei Li

摘要

Multivariate mixture data analysis presents numerous challenges and constitutes a vital area of interest in the fields of statistics and data science. Research into multivariate mixture structures holds relevance across diverse application domains and plays a pivotal role in the advancement of artificial intelligence and machine learning. In this article, we focus on nonparametric estimation techniques for multivariate mixture data. Specifically, we assume a known number of subpopulations and propose a binomial likelihood method, along with an efficient numerical algorithm, to estimate the mixing proportions and cumulative distribution functions of these subpopulations without relying on parametric assumptions. Through extensive numerical experiments, we demonstrate three key advantages of our approach: (a) Our method eliminates the need for tuning parameters. (b) It does not require the assumption of continuous component density functions. (c) Our method consistently delivers stable performance. Under mild regularity conditions, we provide theoretical proofs for the asymptotical properties of our estimators. To illustrate the practical performance of our method, we include a real-data example. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work. *(abstract via OpenAlex)*

JRSSB Optimization & Learning Theory · 发表于 2026-04-22

Gaussianized design optimization for covariate balance in randomized experiments

Wenxuan Guo, Tengyuan Liang, Panos Toulis

摘要

Achieving covariate balance in randomized experiments enhances the precision of treatment effect estimation. However, existing methods often require heuristic adjustments based on domain knowledge and are primarily developed for binary treatments. This paper presents Gaussianized Design Optimization, a novel framework for optimally balancing covariates in experimental design. The core idea is to Gaussianize the treatment assignments: we model treatments as transformations of random variables drawn from a multivariate Gaussian distribution and convert the design problem into a nonlinear continuous optimization over Gaussian covariance matrices. Compared to existing methods, our approach offers significant flexibility in optimizing covariate balance across a diverse range of designs and covariate types. Adapting the Burer–Monteiro approach for solving semidefinite programmes, we introduce first-order local algorithms for optimizing covariate balance, improving upon several widely used designs. Furthermore, we develop inferential procedures for constructing design-based confidence intervals under Gaussianization and extend the framework to accommodate continuous treatments. Simulations demonstrate the effectiveness of Gaussianization in multiple practical scenarios.

JASA Machine Learning Optimization & Learning Theory Bayesian Methods · 发表于 2026-04-22,收录于 2026-04-24

Bayesian Optimization for Branching and Nested Hyperparameters in Deep Learning

Jiazhao Zhang, Chung-Ching Lin, Ying Hung

摘要

Hyperparameter optimization plays a crucial role in the success of neural networks as hyperparameters directly control the behavior and performance of the training algorithms. To obtain efficient tuning, Bayesian optimization based on Gaussian process is widely used. Despite numerous applications in deep learning, the existing methods rely on a convenient but restrictive assumption that the tuning parameters are independent of each other. However, tuning parameters with conditional dependence are common in practice. In this article, we focus on two types of them: branching and nested parameters. Nested parameters refer to those tuning parameters that exist only within a particular setting of another tuning parameter, and a parameter that contains other nested parameters is referred to as a branching parameter. To capture the conditional dependence between branching and nested parameters, a unified Bayesian optimization framework is proposed. The sufficient conditions are rigorously derived to guarantee the validity of the kernel function, and the asymptotic convergence of the proposed optimization framework is proven under the continuum-armed-bandit setting. The new model, which accounts for the dependent structure among input variables through a new kernel function, provides a higher prediction accuracy and a better optimization efficiency in simulations and real data applications in neural networks. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work. *(abstract via OpenAlex)*

JASA Latent Variables & Mixtures Bayesian Methods · 发表于 2026-04-22,收录于 2026-04-24

Adaptive Partition Factor Analysis

Elena Bortolato, Antonio Canale

摘要

Factor Analysis has traditionally been utilized across diverse disciplines to extrapolate latent traits that influence the behavior of multivariate observed variables. Historically, the focus has been on analyzing data from a single study, neglecting the potential study-specific variations present in data from multiple studies. Multi-study factor analysis has emerged as a recent methodological advancement that addresses this gap by distinguishing between latent traits shared across studies and study-specific components arising from artifactual or population-specific sources of variation. In this paper, we extend the current Bayesian methodologies by introducing novel shrinkage priors for the latent factors, thereby accommodating a broader spectrum of scenarios—from the absence of study-specific latent factors to models in which factors pertain only to small subgroups nested within or shared between the studies. For the proposed construction we provide conditions for identifiability of factor loadings and guidelines to perform straightforward posterior computation via Gibbs sampling. Through comprehensive simulation studies, we demonstrate that our proposed method exhibits competing performance across a variety of scenarios compared to existing methods, yet providing richer insights. The practical benefits of our approach are further illustrated through applications to bird species co-occurrence data and ovarian cancer gene expression data. *(abstract via OpenAlex)*

JASA Machine Learning · 发表于 2026-04-22,收录于 2026-04-24

A Physics-Informed Spatiotemporal Deep Learning Framework for Turbulent Systems

Luca Menicali, Andrew P. Grace, David H. Richter, Stefano Castruccio

摘要

Fluid thermodynamics underpins atmospheric dynamics, climate science, industrial applications, and energy systems. However, direct numerical simulations (DNS) of such systems can be computationally prohibitive. To address this, we present a novel physics-informed spatiotemporal surrogate model for Rayleigh-Bénard convection (RBC), a canonical example of convective fluid flow. Our approach combines convolutional neural networks, for spatial dimension reduction, with an innovative recurrent architecture, inspired by large language models, to model long-range temporal dynamics. Inference is penalized with respect to the governing partial differential equations to ensure physical interpretability. Since RBC exhibits turbulent behavior, we quantify uncertainty using a conformal prediction framework. This model replicates key physical features of RBC dynamics while significantly reducing computational cost, offering a scalable alternative to DNS for long-term simulations. *(abstract via OpenAlex)*

SiM Survival Analysis · 发表于 2026-04-20,收录于 2026-04-24

Novel Influence Diagnostics in Multistate Models for Breast Cancer

Valeria Leiva‐Yamaguchi, Alejandra Tapia, Manuel Galea, Oscar M. Rueda

摘要

Multistate models were developed to model survival data where several midpoints and endpoints are of interest; and they have been particular successful in modeling dynamics of cancer. As in any statistical model, identification of influential observations is an essential task, as they can significantly affect the validity of inferred parameters and conclusions drawn from the data. The local influence approach is a set of methods designed to detect the effect of small perturbations of the model or data on the inference, allowing for deeper data analysis. In this paper, we derive local influence methods for multistate models and illustrate their use with a breast cancer dataset. In particular, we develop and implement local influence diagnostic techniques based on a suitable estimation equation. For simplicity, we restrict our consideration to the Multistate Proportional Hazards model, using different case‐weight perturbation strategies.

SiM Other · 发表于 2026-04-20,收录于 2026-04-24

Extending the Median Odds Ratio ( MOR ), the Interval Odds Ratio ( IOR ), and the Proportion of Opposed Odds Ratios ( POOR ) for Use With 3‐Level Multilevel Logistic Regression Models

Peter C. Austin, George Leckie

摘要

Multilevel data occur frequently in health services, population and public health, and epidemiologic research (e.g., patients nested within hospitals). Multilevel logistic regression models allow one to account for the clustering of individuals within higher‐level units when estimating the association of individual‐level and cluster‐level characteristics with individual‐level binary outcomes. Odds ratios for cluster‐level variables generated from a multilevel logistic regression model have an interpretation that is conditional on the random effects, compared to the marginal interpretation of odds ratios from logistic regression models that do not incorporate cluster‐specific random effects. The interval odds ratio (IOR) and the proportion of opposed odds ratios (POOR) are two summary measures that have been proposed for quantifying the magnitude and heterogeneity of the association of cluster‐level variables with individual‐level binary outcomes. The median odds ratio (MOR) is a measure of the between‐cluster heterogeneity of a binary outcome. The IOR, the POOR, and the MOR were developed for use with two‐level multilevel logistic regression models (e.g., patients nested within hospitals). In applied research, data with a three‐level multilevel structure are common (e.g., hospitalized patients cared for by the attending physician who in turn practice within hospitals). We mathematically derived extensions of the MOR, the IOR, and the POOR for use with three‐level multilevel logistic regression models. We then illustrated the application and interpretation of the MOR, the IOR, and the POOR by analyzing mortality in patients hospitalized with acute myocardial infarction, where patients are nested within physicians who in turn are nested within hospitals.

SiM Spatial & Time Series Bayesian Methods · 发表于 2026-04-20,收录于 2026-04-24

BAR12 : Bayesian Autoregressive Phase 1‐2 Design for Cell Therapy Trials With Manufacturing Changes

Cheng‐Han Yang, Peter F. Thall, David Marin, Sheferaw Y. Belay, Ruitao Lin

摘要

An early phase dose‐finding trial of a new cell therapy may involve one or more manufacturing modifications made during the trial, known as “tweaks,” to improve the cell product quality. For example, a tweak may change the cell culture duration, cytokine cocktail, or donor criteria. Ideally, a tweak can be done without changing the treatment so much that the trial must be restarted as if the treatment were entirely new. However, a statistical design still should account for changes in the dose–response distribution due to the tweak. For such settings, we propose a B ayesian A uto R egressive Phase 1‐2 (BAR12) design that accounts for manufacturing tweaks during a phase 1‐2 trial by using a first order autoregressive model with spike‐and‐slab priors having components corresponding to pre‐ and post‐tweak distributions of toxicity and efficacy parameters. Simulations under a broad range of dose‐response functions were conducted to compare BAR12 to conventional phase 1‐2 designs that either assume the tweak had no effect and use all available data, or ignore the pre‐tweak data. The simulations show that BAR12 has superior operating characteristics, including higher probabilities of correct dose selection, allocation of more patients to optimal doses, and more efficient monitoring for identifying unsafe or ineffective doses.

SiM High-Dimensional Latent Variables & Mixtures · 发表于 2026-04-19,收录于 2026-04-24

Penalized Estimation in Finite Mixtures of Multivariate Regression Models via the EM‐PGM Algorithm

Heeyeon Kang, Sunyoung Shin

摘要

In many areas of medical and life sciences research, multiple diagnostic criteria and phenotypic outcomes are assessed simultaneously to characterize diseases and biological traits. These multivariate outcomes are often associated with high‐dimensional covariates and subject to underlying population heterogeneity, presenting challenges for analysis and interpretation. To address these issues, we study a finite mixture of multivariate regression models (mvFMR), which jointly models multiple outcomes while capturing latent subgroup structure. By accounting for both outcome multiplicity and data heterogeneity, mvFMR improves predictive performance and enhances interpretability for multivariate outcomes in complex datasets. We adopt a penalized maximum likelihood approach to estimate mvFMR, which enables scalability and variable selection in high‐dimensional settings. We further develop the EM‐PGM algorithm, an efficient estimation procedure that combines the expectation‐maximization (EM) framework with the proximal gradient method (PGM) to handle high‐dimensionality and the non‐differentiability of the penalty function. Simulation studies demonstrate that mvFMR with EM‐PGM outperforms alternative methods in terms of estimation accuracy, sparsity recovery, and computational efficiency. Applications to analyses of diabetes diagnosis data and Cancer Cell Line Encyclopedia data illustrate the practical utility of our penalized mvFMR.

Biometrics Causal Inference Functional Data · 发表于 2026-04-17,收录于 2026-06-05

Personalized treatment design in the context of functional confounding

Zhixian Yang, Peijun Sang, Yixin Han, Bei Jiang, Linglong Kong, Xingcai Zhou

摘要

One of the primary goals of individualized treatment rule (ITR) methodology is to identify optimal decision rules using clinical predictors. While functional data has become increasingly available in biomedical research, there has been limited work on incorporating functional data into ITR estimation, particularly in observational studies. In this paper, we propose a novel approach that integrates outcome-weighted learning (OWL) with reproducing kernel Hilbert space to determine optimal treatment regimes involving functional data. Furthermore, to address the issue of data piling, we employ the distance-weighted discrimination classifier instead of traditional support vector machines. We establish the theoretical consistency of the decision functional estimator with its risk bound. Extensive simulations and the analysis of the Alzheimer’s Disease Neuroimaging Initiative dataset demonstrate the superior performance of our method compared to existing OWL approaches. The results highlight critical factors in Alzheimer’s Disease progression and reveal limitations of the original OWL method in this context.

SiM Survival Analysis Testing & Uncertainty · 发表于 2026-04-17,收录于 2026-04-24

Meta‐Analysis of Median Survival Times With Inverse‐Variance Weighting

Sean McGrath, Cheng‐Han Yang, Jonathan Kimmelman, Omer Ozturk, Russell Steele, Andrea Benedetti

摘要

We consider the problem of meta‐analyzing outcome measures based on median survival times. Primary studies with time‐to‐event outcomes often report estimates of median survival times and confidence intervals based on the Kaplan‐Meier estimator. However, outcome measures based on median survival are rarely meta‐analyzed, as standard inverse‐variance weighted methods require within‐study standard errors that are typically not reported. In this article, we consider an inverse‐variance weighted approach to meta‐analyze median survival times that estimates the within‐study standard errors from the reported confidence intervals. We show that this method consistently estimates the standard error of median survival when applied to confidence intervals constructed by the Brookmeyer–Crowley method. We conduct a series of simulation studies evaluating the performance of this approach at the study level (i.e., for estimating the standard error of median survival) and the meta‐analytic level (i.e., for estimating the pooled median, difference of medians, and ratio of medians) for commonly used confidence intervals for median survival, including the Brookmeyer–Crowley method and nonparametric bootstrap. We find that this approach often performs comparably to a benchmark approach that uses the true within‐study standard errors for meta‐analyzing median‐based outcome measures when within‐study sample sizes are moderately large (e.g., above 50). However, when the effective sample sizes are small, the method can yield biased estimates of within‐study standard errors. We illustrate an application of this approach in a meta‐analysis evaluating survival benefits of being assigned to experimental arms versus comparator arms in randomized trials for non‐small cell lung cancer therapies.

Biometrics Causal Inference Missing Data & Measurement Error · 发表于 2026-04-17,收录于 2026-06-05

Measurement error-robust causal inference via constructed instrumental variables

Caleb H Miles, Linda Valeri, Brent Coull

摘要

Measurement error can often be harmful when estimating causal effects. Two scenarios in which this is the case are in the estimation of (a) the average treatment effect when confounders are measured with error, and (b) the natural indirect effect when the exposure and/or confounders are measured with error. Methods adjusting for measurement error typically require external data or knowledge about the measurement error distribution. Here, we propose methodology not requiring any such information. Instead, we show that when the outcome regression is linear in the error-prone variables, consistent estimation of these causal effects can be recovered using constructed instrumental variables (IVs) under certain conditions. These variables, which are functions of only the observed data, behave like IVs for the error-prone variables. Using data from a study of the effects of prenatal exposure to heavy metals on growth and neurodevelopment in Bangladeshi mother–infant pairs, we apply our methodology to estimate (a) the effect of lead exposure on birth length while controlling for maternal protein intake, and (b) lead exposure’s role in mediating the effect of maternal protein intake on birth length. Protein intake is calculated from food journal entries, and is suspected to be highly prone to measurement error.

SiM Bayesian Methods Spatial & Time Series · 发表于 2026-04-17,收录于 2026-04-24

Joint Bayesian Nowcasting of Severe Acute Respiratory Illness and COVID‐19 Positives in Brazil

Alba Halliday, Oliver Stoner, Theo Economou, Leonardo Soares Bastos

摘要

Surveillance of severe acute respiratory illness (SARI) in Brazil provides an early warning system for respiratory outbreaks, including COVID‐19, and statistical nowcasting methods are vital for correcting long and variable reporting delays that would otherwise hinder timely decision‐making. Coherent joint prediction of SARI and COVID‐positive SARI in Brazil could improve outbreak response planning and hospital resource management, but most existing nowcasting methods only target one outcome at a time. Furthermore, existing approaches usually focus solely on reconstructing recent incidence rather than forecasting future trends, which could support more proactive risk mitigation. Here, we propose a Bayesian hierarchical framework for joint nowcasting and short‐term forecasting of two outcomes, where one is a strict subset of the other. Building on the generalized‐Dirichlet‐multinomial (GDM) method for correcting delayed reporting, a new beta‐binomial component links SARI and COVID‐positive counts, while separate conditional GDM components capture their distinct delay patterns. To allow for changes over time in the level of disease and delay distributions, flexible latent effects are included in all model components. Using national surveillance data from 2021 to 2024, we conduct a 20‐date rolling prediction experiment across Brazil's 27 federative units. Compared to a well‐established Bayesian nowcasting approach, our joint model achieves about one‐third lower mean absolute error and continuous ranked probability score for contemporaneous nowcasts, with the largest gains in high‐incidence regions. Meanwhile, energy scores indicate improved calibration for joint forecasts of total and COVID‐positive SARI relative to comparable independent models.

Biometrics Spatial & Time Series Survival Analysis Bayesian Methods · 发表于 2026-04-17,收录于 2026-06-05

A zero-inflated hierarchical generalized transformation model to address non-normality in spatially-informed cell-type deconvolution

Hunter J Melton, Jonathan R Bradley, Chong Wu

摘要

Oral squamous cell carcinomas (OSCC), the predominant head and neck cancer, pose significant challenges due to late-stage diagnoses and low five-year survival rates. Spatial transcriptomics offers a promising avenue to decipher the genetic intricacies of OSCC tumor microenvironments. In spatial transcriptomics, Cell-type deconvolution is a crucial inferential goal; however, current methods fail to consider the high zero-inflation present in OSCC data. To address this, we develop a novel zero-inflated version of the hierarchical generalized transformation model (ZI-HGT) and apply it to the Conditional AutoRegressive Deconvolution (CARD) for cell-type deconvolution. The ZI-HGT serves as an auxiliary Bayesian technique for CARD, reconciling the highly zero-inflated OSCC spatial transcriptomics data with CARD’s normality assumption. The combined ZI-HGT + CARD framework achieves enhanced cell-type deconvolution accuracy and quantifies uncertainty in the estimated cell-type proportions. We demonstrate the superior performance through simulations and analysis of the OSCC data. Furthermore, our approach enables the determination of the locations of the diverse fibroblast population in the tumor microenvironment, critical for understanding tumor growth and immunosuppression in OSCC.

SMMR Meta-Analysis & Data Integration · 发表于 2026-04-16

The applicability to systematic reviews of common effect, random effects and fixed effects approaches to meta-analysis

Richard J Stevens

摘要

Systematic reviewers planning quantitative meta-analysis usually choose between fixed effect meta-analysis, and random effects meta-analysis. An alternative method is called fixed effects (note the s in the name). This method has the unique property that the target estimand is defined by the variances of the studies found by the systematic review. This article considers each in relation to the quantitative analysis of data to be obtained by systematic review. For clarity, I refer to the traditional fixed effect method as the common effect method and the newer approach as fixed effects (plural). Case studies illustrate the post hoc nature of the fixed effects (plural) method, in which the population under study is determined by the data rather than by the protocol. Mathematical analysis shows that unlike common effect and random effects methods, the fixed effects (plural) method requires an additional, and unrealistic, assumption about the data obtained in systematic reviews. A simulation study demonstrates that confidence intervals from fixed effects (plural) meta-analysis do not account for the post hoc nature of the method. Fixed effects (plural) meta-analysis is neither a slot-in replacement for the common effect method nor for the random effects method of meta-analysis. Of the three methods considered here, the common effect method and the random effects method are potentially valid for the quantitative analysis of systematic reviews.

Biometrika Survival Analysis Functional Data High-Dimensional · 发表于 2026-04-16

Functional principal component analysis forsparse censored data

Caitrin Murphy, Eric Laber, Rhonda Merwin, Brian Reich, Jake Koerner

摘要

Summary: Functional principal component analysis is a key tool in the study of functional data, driving both exploratory analyses and feature construction for use in formal modelling and testing procedures. However, existing methods do not apply when functional observations are censored; for example, when the measurement instrument only supports recordings within a prespecified interval, thereby truncating values outside this range to the nearest boundary. A naïve application of existing methods, without correction for instrument-induced censoring, introduces bias into the estimators of the mean, covariance and functional principal component scores. We extend the functional principal component analysis framework to accommodate noisy and potentially sparse censored functional data. Local loglikelihood maximization is used to recover smooth estimates of the mean and covariance surface that are representative of the latent process’s mean and covariance functions. The covariance-smoothing procedure yields a positive semidefinite covariance surface, computed without the need to retroactively remove negative eigenvalues in the covariance-operator decomposition. Additionally, we construct a predictor of the scores, conditional on the censored functional data, and demonstrate its use in the generalized functional linear model. Convergence rates for the proposed estimators are established. In simulation experiments, the proposed method yields improved predictive performance and lower bias than existing alternatives. We illustrate its practical value in a study aimed at classifying eating disorder diagnoses in individuals with type 1 diabetes, using censored functional blood glucose data.

SiM Other · 发表于 2026-04-16,收录于 2026-04-24

Correcting for a Baseline Difference in Group Comparisons on a Binary Outcome: Covariate Adjustment or Analysis of Change?

Gerard J. P. Van Breukelen

摘要

Effects of treatments or exposures are evaluated by comparing a treated or exposed group with a control group after treatment or exposure. A baseline group difference can be accounted for by covariate adjustment or by analyzing change from baseline. For quantitative outcomes these two methods can give contradictory results (Lord's paradox) and especially covariate adjustment is questionable for nonrandomized group comparisons. This paper explores analogous methods for the case of a binary outcome, specifically logistic regression of the outcome on group and baseline, ordinal regression of change from baseline on group, and mixed logistic regression and generalized estimating equations ( GEE ) for repeated measures. The methods were compared mathematically, conceptually in terms of causal diagrams and estimands, and numerically on fictitious scenarios that varied in whether groups differed at baseline and/or in change over time on the logodds scale. The methods were also compared on a smoking prevention study among school children. The scenario results were similar to those published for quantitative outcomes: Logistic regression of outcome on group and baseline gave almost the same results as mixed logistic regression and GEE without parameter for a baseline group difference. Ordinal regression of change from baseline gave almost the same results as mixed logistic regression and GEE allowing for a baseline group difference. These (near‐)equivalences are in line with two mathematical proofs in this paper. Further, in data sets with a baseline group difference, covariate adjustment and change analysis led to contradictory conclusions. The results from the smoking prevention study confirmed the above results and showed Lord's paradox.

JRSSB Meta-Analysis & Data Integration Latent Variables & Mixtures · 发表于 2026-04-16

Byzantine-tolerant distributed learning of finite mixture models

Qiong Zhang, Yan Shuo Tan, Jiahua Chen

摘要

Traditional statistical methods need to be updated to work with modern distributed data storage paradigms. The split-and-conquer framework that learns models on local machines and averaging their parameter estimates is common. However, this does not work for the important problem of learning finite mixture models, because subpopulation indices on each local machine may be arbitrarily permuted (the ‘label switching problem’). Earlier work proposed mixture reduction (MR) to address this issue, offering an effective and efficient solution for aligning and aggregating local mixture components. Building upon this foundation, this paper considers the additional challenge of Byzantine failure, where a fraction of local machines may transmit arbitrarily erroneous information. We introduce distance-filtered mixture reduction (DFMR), a Byzantine-tolerant framework that enhances MR by adding a distance-based filtering mechanism to identify and exclude corrupted local estimates before the MR aggregation. This integration allows DFMR to maintain MR’s efficiency while achieving strong robustness against Byzantine failure. We provide theoretical justification for DFMR, proving its optimal convergence rate and asymptotic equivalence to the global maximum likelihood estimate under standard assumptions. Numerical experiments on simulated and real-world data validate the effectiveness of DFMR in achieving robust and accurate aggregation in the presence of Byzantine failure.

Biometrics Causal Inference Bayesian Methods High-Dimensional · 发表于 2026-04-16,收录于 2026-07-24

Bayesian shrinkage priors for penalized synthetic control estimators in the presence of spillovers

Not available

摘要

Synthetic control (SC) methods are widely used to estimate the effects of policy interventions, especially those targeting specific geographic regions, referred to as units. These methods construct a weighted combination of untreated units, forming a “synthetic” control that approximates the counterfactual outcomes of the treated unit had the intervention not occurred. Although neighboring areas are often selected as controls due to their similarity in observed and unobserved characteristics, their proximity can lead to spillover effects, where the intervention indirectly impacts control units, potentially biasing causal estimates. To address this challenge, we introduce a Bayesian SC framework with utility-based shrinkage priors. Our approach extends traditional penalization techniques (i.e., horseshoe, spike-and-slab) by incorporating a utility function that combines covariate similarity and spatial distance. This provides a metric that guides the data-driven selection of control units based on their relevance and spillover risk, which is assumed to increase with spatial proximity. Rather than outright excluding neighboring units, the method balances bias and variance by reducing the importance of potentially contaminated controls by spillovers. We evaluate the proposed method through simulation studies at varying spillover levels and apply it to assess the impact of Philadelphia’s 2017 beverage tax on the sales of sugar-sweetened and artificially sweetened beverages in mass merchandise stores.

SMMR Bayesian Methods Clinical Trials · 发表于 2026-04-16

Bayesian sample size determination using robust commensurate priors with interpretable discrepancy weights

Lou E Whitehead, James MS Wason, Oliver Sailer, Haiyan Zheng

摘要

Randomized controlled clinical trials provide the gold standard for evidence generation in relation to the efficacy of a new treatment in clinical research. Relevant information from previous studies may be desirable to incorporate in the design and analysis of a new trial, with the Bayesian paradigm providing a coherent framework to formally incorporate prior knowledge. Many established methods involve the use of a discounting factor, sometimes related to a measure of ‘similarity’ between historical and the new trials. However, it is often the case that the sample size is highly nonlinear in those discounting factors. This hinders communication with subject-matter experts to elicit sensible values for borrowing strength at the trial design stage. Focussing on a method that can incorporate historical data from multiple sources, we highlight a particular issue of nonmonotonicity and explain why this causes issues with interpretability of discounting factors (hereafter referred to as ‘weights’). We propose a solution from which an analytical sample size formula is derived. We then propose a linearization technique such that the sample size changes uniformly over the weights. This leads to interpretable weights (as a percentage of information to borrow/discount) which could facilitate easier elicitation of expert opinion on their values.

JRSSB Spatial & Time Series Networks & Graphs · 发表于 2026-04-16

Autoregressive networks with dependent edges

Jinyuan Chang, Qin Fang, Eric D Kolaczyk, Peter W MacDonald, Qiwei Yao

摘要

We propose an autoregressive framework for modelling dynamic networks with dependent edges. It encompasses models that accommodate, for example, transitivity, degree heterogenenity, and other stylized features often observed in real network data. By assuming the edges of networks at each time are independent conditionally on their lagged values, the models, which exhibit a close connection with temporal exponential random graph models, facilitate both simulation and the maximum likelihood estimation (MLE) in a straightforward manner. Due to the possibly large number of parameters in the models, the natural MLEs may suffer from slow convergence rates. An improved estimator for each component parameter is proposed based on an iteration employing projection, which mitigates the impact of the other parameters. Leveraging a martingale difference structure, the asymptotic distribution of the improved estimator is derived without the assumption of stationarity. The limiting distribution is not normal in general, although it reduces to normal when the underlying process satisfies some mixing conditions. Illustration with a transitivity model was carried out in both simulation and a real network data set.

SMMR Longitudinal Data Clinical Trials Missing Data & Measurement Error · 发表于 2026-04-16

A practical review of response-adaptive randomization: Under-explored challenges and potential directions

Hao Mei, Xiaolin Xu, Hang Yang, Fan Wang, Yang Li

摘要

Response-adaptive randomization (RAR) dynamically adjusts allocation probabilities of sequentially enrolled patients based on accumulating response information. It has gained increasing attention in clinical trials for its ability to enhance statistical efficiency by accelerating estimation of treatment effects, improve ethical allocation by favoring superior interventions, and maintain robust parameter estimation (e.g. via variance or uncertainty reduction) while accounting for adaptive modifications. While both regulatory agencies and the broader clinical research community are making efforts to promote the application of RAR, its real-world implementation remains limited. In this article, we review the application of RAR in clinical practice from 2015 to 2024, identifying key challenges such as managing diverse patient outcome types, incorporating repeated measurements, and addressing missing data. We categorize various RAR methods and propose practical solutions to these challenges, providing insights and practical references for clinical trial practitioners. Additionally, we discuss the limitations of existing RAR methods and outline potential future research directions. Our review aims to bridge the gap between theory and practice, promoting the broader adoption of RAR in clinical trials while advancing the development of RAR theory and methodology.

Biometrika Nonparametric & Semiparametric Networks & Graphs High-Dimensional · 发表于 2026-04-15

Nonparametric estimators over metric graphs

Aldo Clemente, Eleonora Arnone, Jorge Mateu, Laura M Sangalli

摘要

This work discusses a theory of functional spaces over metric graphs that permits the definition of penalized likelihood methods for data observed over spatial supports that are graphs. Within the considered mathematical framework, we recover classical results in functional analysis, such as a Poincaré-type inequality. This, in turn, enables us to uplift, to the present setting, the theory of some fundamental penalized likelihood methods. Specifically, we present two important classes of statistical models: nonparametric regression and nonparametric density estimation, here defined for data observed over graphs. We derive theoretical results regarding the well-posedness of the associated estimation problems and the consistency of the estimators. We also demonstrate the performance of the defined estimators with respect to state-of-the-art alternatives.

JASA Survey & Sampling · 发表于 2026-04-15,收录于 2026-04-17

Double-Robust Small Area Estimation

Haiqiang Ma, Zhiyan Sheng, Jiming Jiang

摘要

In the context of robust small area estimation (SAE), there are two types of robustness considerations, robustness against model misspecification and robustness against outliers. We propose a method of SAE that has both types of robustness features. The method combines the idea of observed best prediction (OBP), which is known to be more robust against model misspecification than the traditional best linear unbiased prediction (EBLUP) method, and the method of density power divergence (DPD), which is known to be more robust against outliers than the EBLUP. The double robust predictor (DRP) is developed under an area-level model with normal or normal-mixture sampling errors, and under a unit-level model. Another advantage of the DRP method is that it provides an natural estimator of a tuning parameter involved in the DPD. We develop theory about the proposed method, and demonstrate empirical performance of the proposed DRP and its comparison to EBLUP, OBP, a robust version of the EBLUP, and predictors based on the DPD. A second-order unbiased estimator of the mean squared prediction error of the DRP is developed and its performance is evaluated. A real-data example is discussed. *(abstract via OpenAlex)*

SiM High-Dimensional · 发表于 2026-04-14,收录于 2026-04-17

Unsupervised Sparse Multi‐Task Learning With Application to Alzheimer's Disease

Hao Chen, Jiadong Ji, Dong Liu, Bofeng Yu

摘要

This article is motivated by the challenge of identifying interpretable brain functional connectivity biomarkers for Alzheimer's disease (AD) progression using high‐dimensional functional magnetic resonance imaging (fMRI) data, where predictors comprise both heterogeneous disease‐stage activation patterns and strongly similar functional connections. We propose a unified statistical method, called Sparse Multi‐task Adaptive Regularization Truncation (SMART), to simultaneously address three critical challenges: (1) High dimensionality is resolved through an ‐penalty () that selects sparse, stage‐distinct functional connections; (2) Disease‐stage heterogeneity is accommodated via an ‐penalty () that maintains stable activation patterns across tasks; (3) Connection collinearity is mitigated using a truncated penalty (TLP; ) that adaptively groups edges with similar cross‐task profiles without pre‐specified structure. SMART offers key advantages over existing methods: Its joint regularization naturally handles smooth activation patterns across stages, while the TLP's dual parameters ( for adaptive grouping threshold, for sparsity control) provide a principled trade‐off between biological fidelity and model complexity. Computationally, we develop a DC‐ADMM algorithm that transforms the optimization into tractable subproblems, establishing finite‐step convergence to KKT points. Comprehensive simulation studies and real‐data analysis of AD neuroimaging data demonstrate SMART's superior accuracy in connectivity biomarker identification, enhanced stability in feature selection, and improved interpretability for AD cohort studies. An accompanying R package, SMART, is available on GitHub ( https://github.com/LDstat/SMART ).

JASA Machine Learning Optimization & Learning Theory · 发表于 2026-04-14

Towards Better Statistical Understanding of Watermarking LLMs

Zhongze Cai, Shang Liu, Hanzhao Wang, Huaiyang Zhong, Xiaocheng Li

摘要

In this paper, we study the problem of watermarking large language models (LLMs). We consider the trade-off between model distortion and detection ability and formulate it as a constrained optimization problem based on the red-green list watermarking algorithm. We show that the optimal solution to the optimization problem enjoys a nice analytical property which provides a better understanding and inspires the algorithm design for the watermarking process. We develop an online dual gradient ascent watermarking algorithm in light of this optimization formulation and prove its asymptotic Pareto optimality between model distortion and detection ability. Such a result guarantees an averaged increased green list probability and henceforth detection ability explicitly (in contrast to previous results). Moreover, we provide a systematic discussion on the choice of the model distortion metrics for the watermarking problem. We justify our choice of KL divergence and present issues with the existing criteria of “distortion-free” and perplexity. Finally, we empirically evaluate our algorithms on extensive datasets against benchmark algorithms. *(abstract via OpenAlex)*

第 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 页