2026 年的期刊文章 · 第 9 页

共 1756 篇,按发表时间排,新的在前,每页 40 篇。这一页是 2026-08-27 到 2026-08-19。

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

JASA Other · 发表于 2026-08-27

Dose-Exposure-Response Modeling: Methods and Practical Implementation, 2nd ed

Sidi Wang

摘要

Jixian Wang. Boca Raton, FL: Chapman & Hall/CRC Press, 2026, xvi + 345 pp., $ 166.99(H), ISBN: 978-1-032-59625-9.This is the second edition of Jixian Wang’s book on exposure-response modeling, now ... *(abstract via OpenAlex)*

JASA Survey & Sampling · 发表于 2026-08-27

Complete Sample Likelihood Estimation and Inference for Single and Multi-Stage Sample Surveys

Robert G. Clark, Francis K. C. Hui

摘要

In the analysis of sample surveys, accounting for informative sampling design is critical to ensure estimates of model parameters such as regression coefficients are not systematically biased. One popular solution is maximum pseudo-likelihood estimation, which incorporates weights given by functions of the selection probabilities. In this article, we propose an alternative method for analytical estimation and inference in sample surveys known as complete sample likelihood or CSL. This approach jointly postulates a population regression model for the responses, and a model for the selection probabilities conditional on the response and measured covariates. CSL can offer potentially substantial reductions in bias and standard errors compared to maximum pseudo-likelihood estimation, while also avoiding a major drawback of existing “plug-in" sample likelihood techniques which require burdensome resampling techniques to perform inference. We develop general forms for CSL and theoretically justify their usage in both single- and multi-stage sample surveys, before deriving several tractable cases involving generalized linear (mixed) models commonly seen in practice with sample surveys. Simulation studies and an application to a two-stage sample survey of smoking status in New Zealand demonstrate the superior finite sample performance of CSL, and its ability to facilitate deployment of likelihood-based techniques and diagnostic tools. *(abstract via OpenAlex)*

JASA Testing & Uncertainty Causal Inference Clinical Trials · 发表于 2026-08-27,收录于 2026-06-26

Anytime-Valid Inference in Linear Models with Applications to Regression-Adjusted Causal Inference

Michael Lindon, Dae Woong Ham, Martin Tingley, Iavor Bojinov

摘要

(verbatim from arXiv:2210.08589 — tandfonline.com blocked direct fetching, so this is the preprint abstract, not the official JASA page; online date could not be confirmed)*: "Linear models are foundational tools in statistics and ubiquitous across the applied sciences. However, conventional statistical inference -- such as $t$-tests and $F$-tests -- are only valid at fixed sample sizes, making them unsuitable for sequential settings such as online A/B testing. We develop an anytime-valid theory of inference for the linear model, introducing sequential analogues of classical tests and confidence sets that provide Type-I error control and coverage guarantees uniformly over all sample sizes. Our construction is based on likelihood ratios of invariantly sufficient statistics, yielding simple closed-form expressions of ordinary least squares estimators and standard errors. The resulting tests are optimal in the GROW/REGROW sense for both frequentist and Bayesian alternative hypotheses. We then relax the linear model assumptions to provide heteroskedasticity-robust asymptotic sequential tests and confidence sequences, which enable sequential regression-adjusted inference for causal estimands in randomized controlled experiments. This formally allows experiments to be continuously monitored for significance, stopped early, and safeguards against statistical malpractices in data collection. We demonstrate the practical utility of our approach through simulations and applications to real A/B test data from Netflix.

JASA High-Dimensional Latent Variables & Mixtures · 发表于 2026-08-27

A General Latent Embedding Approach for Modeling Non-uniform High-dimensional Sparse Hypergraphs with Multiplicity

Shihao Wu, Gongjun Xu, Ji Zhu

摘要

Recent research has shown growing interest in modeling hypergraphs, which capture polyadic interactions among entities beyond traditional dyadic relations. However, most existing methodologies for hypergraphs face significant limitations, including their heavy reliance on uniformity restrictions for hyperlink orders and their inability to account for repeated observations of identical hyperlinks. In this work, we introduce a novel and general latent embedding approach that addresses these challenges through the integration of latent embeddings, vertex degree heterogeneity parameters, and an order-adjusting parameter. Theoretically, we investigate the identifiability conditions for the latent embeddings and associated parameters, and we establish the convergence rates of their estimators along with asymptotic distributions. Computationally, we employ a projected gradient ascent algorithm for parameter estimation. Comprehensive simulation studies demonstrate the effectiveness of the algorithm and validate the theoretical findings. Moreover, an application to a co-citation hypergraph illustrates the advantages of the proposed method. *(abstract via OpenAlex)*

SiM Survival Analysis Nonparametric & Semiparametric Latent Variables & Mixtures · 发表于 2026-08-26

The Cox–Aalen Rate Model for Recurrent Events With an Informative Terminal Event

Mingzu Yue, Xin Chen, Jiaqi Chen, Liuquan Sun

摘要

Recurrent event data arise frequently in many clinical and observational studies. In this article, we propose a Cox–Aalen rate model to analyze recurrent event data with a terminal event, where the covariate effects are additive, time‐varying, and dependent on a latent variable nonparametrically. The association between recurrent and terminal events is fully nonparametric, and a proportional hazards model is specified for the terminal event. To estimate the model parameters, an estimating equation approach is developed and kernel‐smoothing techniques are employed to estimate the conditional expectations in the estimating equation. The asymptotic properties of the proposed estimators are derived, and the finite sample performance is evaluated through simulation studies. An application to medical cost data for chronic heart failure patients is presented.

SiM Genetics & Omics · 发表于 2026-08-26

ReMeDy : A Flexible Statistical Framework for Region‐Based Detection of DNA Methylation Dysregulation

Suvo Chatterjee, Siddhant Meshram, Ganesan Arunkumar, Fasil Tekola‐Ayele, Arindam Fadikar

摘要

Region‐based epigenome‐wide association studies have demonstrated improved statistical power and biological interpretability compared with probe‐wise analyses of DNA methylation data. However, most existing region‐based methods characterize methylation dysregulation primarily through changes in mean methylation levels associated with a phenotype of interest. Substantial evidence indicates that phenotype‐associated methylation alterations may also manifest through changes in methylation variability or through joint shifts in mean and variability. Despite this, no existing statistical framework jointly models mean–variance methylation changes in a region‐based manner. We propose ReMeDy, a flexible statistical framework that uses a hierarchical likelihood approach within a generalized linear model setting to identify differentially methylated regions, variably methylated regions, and regions exhibiting joint differential and variable methylation at a genome‐wide scale. Unlike existing models, ReMeDy operates directly on biologically defined co‐methylated regions, allowing it to naturally capture spatial correlation inherent in DNA methylation array data, while avoiding reliance on heuristic, user‐defined tuning parameters such as smoothing spans and kernel bandwidths that can substantially influence results and introduce subjectivity. Through extensive simulation studies and comprehensive benchmarking against popular models, we demonstrate that ReMeDy maintains false discovery and Type‐I error rates at nominal levels while achieving consistently higher statistical power across a wide range of realistic scenarios. Application to population‐level DNA methylation data further shows that ReMeDy identifies biologically meaningful regions and pathways implicated in complex human diseases that are not captured by conventional mean‐based analyses alone. ReMeDy is implemented as an open‐source R package and is freely available at https://github.com/SChatLab/ReMeDy .

SMMR Testing & Uncertainty Genetics & Omics · 发表于 2026-08-26

Combining dependent p-values with transformation using empirical distribution of correlated data and its application for genomic data

Junsik Kim, Junyong Park

摘要

Combining dependent p -values is a critical challenge in large-scale hypothesis testing, with applications in genome-wide association studies, transcriptomics, environmental studies, and meta-analyses. Existing methods—such as those based on transforming p -values into heavy-tailed distribution or estimating the correlation matrix of test statistics—often fail to control Type I error under complex dependency structures or rely heavily on impractical assumptions of dependency structures. To address these issues, we develop a new method consisting of two procedures: First, we propose an iterative algorithm to estimate the empirical null distribution function of dependent data. The proposed algorithm incorporates imputations of data simulated from the estimated null distribution. Second, we generate modified p -values based on the estimated empirical null distribution and show that these modified p -values are decorrelated. Combining these modified p -values provides more accurate Type I error control compared to existing methods. In addition, it improves statistical power through the strategy of imputation, while maintaining robustness across various dependency structures. Extensive numerical studies and real-world applications demonstrate the effectiveness of the proposed method in improving both Type I error control and testing power.

SMMR Causal Inference · 发表于 2026-08-26

A propensity score-integrated win ratio method for placebo borrowing

Yurong Chen, Michael Sonksen, Tuo Wang, Yingdong Feng, Joon Jin Song

摘要

Leveraging external control data has been used to enhance the efficiency of clinical trials, especially in rare diseases where recruitment is often challenging. However, directly pooling data from different studies without appropriate adjustments can lead to biased results when populations differ across these studies. In addition to limited sample sizes, trials for rare diseases commonly assess treatment efficacy through multiple clinical endpoints using composite endpoints. The win ratio has gained attention for composite endpoint analysis, as it enables prioritized comparisons that account for the relative clinical importance of each endpoint. Motivated by these two challenges, we proposed novel propensity score (PS)-integrated win ratio methods to incorporate external control data. Specifically, two PS-based weighting approaches, PS-ratio and PS-difference, are proposed to adjust for between-study baseline covariate differences, thereby mitigating the risk of potential bias. Simulation studies and real-world case analysis demonstrate that the proposed methods consistently improve statistical power while maintaining proper Type I error control. This framework offers a robust and practical solution for composite endpoint analysis using external controls, with particular relevance to rare disease trials and regulatory decision-making.

SiM Clinical Trials Testing & Uncertainty · 发表于 2026-08-25

Family‐Wise Error Rate Control in Clinical Trials With Overlapping Populations

Remi Luschei, Werner Brannath

摘要

We consider clinical trials with multiple, overlapping patient populations that test multiple treatment policies specifically tailored to these populations. Such designs may lead to multiplicity issues, as false statements will affect several populations. For type I error control, often the family‐wise error rate (FWER) is controlled, which is the probability to reject at least one true null hypothesis. If the joint distribution of the test statistics is known, the FWER level can be exhausted by determining critical values or adjusted ‐levels. The adjustment is typically done under the common ANOVA assumptions. However, the performed tests are then only valid under the rather strong assumption of homogeneous null effects, that is, when the null hypothesis applies to all subpopulations and their intersections. We show that under cancelling null effects, when heterogeneous effects cancel out in some or all subpopulations, this procedure does not provide FWER control. We also suggest different alternatives and compare them in terms of FWER control and their power.

SiM Longitudinal Data Diagnostics & Prediction Causal Inference · 发表于 2026-08-25,收录于 2026-10-02

Dynamic Predictions and Predictimands for Salvage Therapy in Recurrent Prostate Cancer Using Joint Models

Owens et al.

摘要

Prostate cancer patients with biochemical recurrence (BCR) face a decision of whether to start salvage therapy (ST), which may reduce the probability of metastatic progression at the cost of side effects. To inform the decision to start ST at or after BCR, models that make counterfactual predictions incorporating treatment are highly desirable. However, estimation of such models using observational data requires care due to time‐varying confounding by the longitudinal biomarker prostate‐specific antigen (PSA). Moreover, a careful definition of the estimands of interest, referred to as “predictimands”, is required due to the possibility of delayed initiation of treatment after biochemical recurrence. In this study, we utilize the framework of joint longitudinal and survival models to tackle these issues, estimating a model for pre‐ST PSA trajectories and risk of metastasis that incorporates the effect of ST, from a dataset of 2075 patients with BCR. We define relevant predictimands for a new patient after BCR under three scenarios: Immediately treated, never treated, and treatment under a dynamic regime, where ST is started when PSA is observed to exceed a pre‐specified threshold. We propose a Monte Carlo scheme for computing these predictimands, adapting previous work on dynamic predictions from joint models to account for treatment timing. This methodology is applied to an example patient and validated in a simulation study. This methodology could be adapted to a wide variety of applications requiring counterfactual predictions in the presence of time‐varying treatments and biomarkers. Code to implement such analyses is available in the R package JMbayes2.

SiM Causal Inference Bayesian Methods Machine Learning · 发表于 2026-08-25

Bayesian Machine Learning for Estimating Optimal Dynamic Treatment Regimes With Ordinal Outcomes

Xinru Wang, Tanujit Chakraborty, Bibhas Chakraborty

摘要

Dynamic treatment regimes (DTRs) are sequences of decision rules designed to tailor treatments based on patients' treatment history and evolving disease status. Ordinal outcomes frequently serve as primary endpoints in clinical trials and observational studies. However, constructing optimal DTRs for ordinal outcomes has been underexplored. This article introduces a Bayesian machine learning (BML) framework to estimate optimal DTR with ordinal outcomes. To deal with potential nonlinear associations between outcomes and predictors, we first introduce ordinal Bayesian additive regression trees (OBART), a Bayesian tree‐based model that integrates the latent continuous variable framework within the traditional Bayesian additive regression trees (BART). We then incorporate OBART into the BML (named BML‐OBART) to estimate optimal DTRs based on ordinal data and quantify the associated uncertainties of the estimated parameters. Extensive simulation studies were conducted to evaluate the performance of the proposed method in comparison with existing methods. We demonstrate the application of the proposed BML method using data from a smoking cessation trial and provide the OBART R package along with R code for the implementation of BML‐OBART.

SiM Other · 发表于 2026-08-25

An Assessment of Reliability Estimation Methods for Binomial Health Care Quality Measures

Laura Aume, William White, Po‐Hsu Chen, Jeffrey Geppert

摘要

We evaluated the performance of commonly used methods for estimating the reliability of binomial health care quality measures using simulated datasets spanning a range of performance score means and variances, numbers of entities, and patient sample sizes. For each simulation, reliability was estimated for all selected methods and compared with the known true reliability derived from the simulation parameters, with methods assessed on their accuracy and precision. Logistic regression with reliability estimated on the outcome scale demonstrated the highest accuracy and precision among all methods evaluated. The widely used Adams beta‐binomial method performed poorly, although a modification recommended by Nieser and Harris substantially improved its performance. These approaches are applicable only to binomial measures. Among methods that can be applied to both binomial and continuous measures, permutation resampling of the Spearman rank correlation coefficient was the most accurate and precise, outperforming other commonly used approaches. Overall, for binomial quality measures, logistic regression on the outcome scale is the preferred method for reliability estimation, followed closely by the modified beta‐binomial approach, while for non‐binomial measures, permutation‐based Spearman rank correlation appears to be the most suitable method.

SiM Bayesian Methods · 发表于 2026-08-25

A Bayesian Spatiotemporal Model for Joint Estimation of Brain Activation and Connectivity in fMRI Studies

Parisa Naseri, Yuliya Shapovalova, Ioan Gabriel Bucur, Charlotte Cambier van Nooten, Giancarlo Valente, Tom Heskes

摘要

Task‐based functional magnetic resonance imaging (fMRI) experiments play a crucial role in modern data‐driven neuroscience research. These studies often aim to understand how external stimuli trigger activation in specific brain regions and to explore functional connectivity patterns among predefined regions, commonly referred to as regions of interest (ROIs). Accurately estimating both brain activation and inter‐regional connectivity is challenging due to complex spatiotemporal correlations and low signal‐to‐noise ratios inherent in fMRI data. This paper introduces a joint spatiotemporal Bayesian framework that simultaneously models activation and connectivity across multiple subjects while estimating the hemodynamic response function (HRF) for each region. Spatial dependencies are captured via an unweighted graph‐Laplacian prior on regression and autoregressive coefficients, and region‐specific random effects are modeled using a Bayesian Gaussian graphical model to reflect connectivity among ROIs. We evaluate the performance of the model through simulation studies, demonstrating robust estimation under realistic low signal‐to‐noise conditions. The approach is then applied to a multisubject motor task dataset from the Human Connectome Project (HCP), mapping brain motor areas associated with specific movements (e.g., finger, toe, tongue) and assessing their activation and lateralization in response to visual cues. The model is further evaluated on the Individual Brain Charting (IBC) dataset, a high‐resolution 3T fMRI dataset designed for fine‐grained cognitive mapping across multiple tasks. Finally, the model is validated through comparisons with the classical general linear model (GLM), which is commonly used in the neuroscience community, highlighting the advantages of our Bayesian approach in jointly capturing activation and connectivity patterns.

SiM Survival Analysis Bayesian Methods · 发表于 2026-08-25

A Bayesian Log‐Cauchy Mixture Cure Fraction Model for Heavy‐Tailed Survival Data: Implementation via OpenBUGS

Khalid Salah

摘要

Bayesian estimation implemented via OpenBUGS is used to establish the novelty and practical relevance of the log‐Cauchy distribution within a mixture cure fraction modeling framework. Mixture cure fraction models are widely used in survival analysis to accommodate the presence of long‐term survivors; however, standard parametric assumptions often fail when survival data exhibit heavy‐tailed behavior and substantial right censoring. In such settings, conventional distributions such as the exponential, Weibull, gamma, and lognormal may lead to biased estimation and inadequate uncertainty quantification. In this paper, we propose a Bayesian mixture cure fraction model based on the log‐Cauchy distribution to flexibly model survival times of uncured subjects in the presence of extreme observations. Bayesian inference is conducted using Markov chain Monte Carlo (MCMC) methods implemented in OpenBUGS, allowing full posterior inference for model parameters through explicit specification of the log‐Cauchy likelihood. A comprehensive simulation study is performed to assess the finite‐sample performance of the proposed model under varying cure fractions and censoring levels. The results demonstrate accurate parameter estimation, stable MCMC convergence, and reliable posterior coverage, even in challenging scenarios characterized by high censoring and long‐term survival. Comparative analyses further show that the proposed model consistently outperforms competing parametric cure models in terms of estimation accuracy and model fit, particularly under heavy‐tailed survival settings. The practical utility of the model is further illustrated through application to a real breast cancer dataset, where the proposed approach provides improved model fit and interpretable inference compared with conventional parametric alternatives. Overall, the results confirm that the Bayesian log‐Cauchy mixture cure model, implemented via OpenBUGS, provides a robust, flexible, and computationally efficient framework for analyzing survival data with cure fractions and heavy‐tailed characteristics, with clear relevance for medical and epidemiological research.

SiM Survival Analysis · 发表于 2026-08-24

Shape‐Based Partially Linear Single‐Index Cox Model for Alzheimer's Disease Conversion

Xuqiao Li, Qiuyan Zhou, Shengxian Ding, Wenliang Pan, Rongjie Liu, Ying Yan, Chao Huang

摘要

Alzheimer's disease (AD) is the major cause of dementia in the elderly, and characterizing the time to conversion to AD is crucial in preventative treatment. While existing statistical methods have proven effective in modeling AD conversion involving various clinical, genetic, and neuroimaging predictors, limited research has explored scenarios where these predictors are shapes derived from the shape space, a nonlinear Hilbert space. In addition, the linear relationship assumption in existing methods may be violated, leading to substantial efficiency losses in real‐world applications. To address these challenges, we propose a shape‐based partially linear single‐index Cox (SPLS‐Cox) model that accommodates both scalar and shape predictors. This new development is motivated by establishing the likelihood of conversion to AD in 372 patients with mild cognitive impairment (MCI) enrolled in the Alzheimer's Disease Neuroimaging Initiative, leveraging the early shape‐based markers of conversion extracted from the brain white matter region, corpus callosum (CC). These 372 MCI patients were followed over 48 months, during which 161 progressed to AD. Our SPLS‐Cox model establishes both the estimation procedure and the pointwise confidence band. Simulation studies are conducted to evaluate the finite‐sample performance of our SPLS‐Cox. The real application reveals that the CC contour shape is a significant predictor for AD conversion.

JASA Other · 发表于 2026-08-24,收录于 2026-06-19

Channeling Multimodality Through a Unimodalizing Transport: Warp-U Sampler and Stochastic Bridge Sampling Estimator

Not available

摘要

Monte Carlo integration is a powerful tool for scientific and statistical computation, but faces significant challenges when the integrand is a multi-modal distribution, even when the mode locations are known. This work introduces novel Monte Carlo sampling and integration estimation strategies for the multi-modal context by leveraging a generalized version of the stochastic Warp-U transformation (Wang et al., 2022). We propose two flexible classes of Warp-U transformations, one based on a general location-scale-skew mixture model and a second using neural ordinary differential equations. We develop an efficient sampling strategy called Warp-U sampling, which applies a Warp-U transformation to map a multi-modal density into a uni-modal one, then inverts the transformation with injected stochasticity. In high dimensions, our approach relies on information about the mode locations, but requires minimal tuning and demonstrates better mixing properties than conventional methods with identical mode information. To improve normalizing constant estimation once samples are obtained, we propose a stochastic Warp-U bridge sampling estimator, which we demonstrate has higher asymptotic precision per CPU second compared to the original approach proposed by Wang et al. (2022). We also establish the ergodicity of our sampling algorithm. The effectiveness and current limitations of our methods are illustrated through simulation studies and an application to exoplanet detection. *(abstract via OpenAlex)*

JASA Bayesian Methods Meta-Analysis & Data Integration High-Dimensional · 发表于 2026-08-24

Bayesian Transfer Learning for Enhanced Estimation and Inference

Daoyuan Lai, Oscar Hernan Madrid Padilla, Xiang Li, Tian Gu

摘要

Transfer learning enhances model performance in a target population with limited samples by leveraging knowledge from related studies. While many works focus on improving predictive performance, challenges persist in statistical inference. Bayesian approaches naturally provide uncertainty quantification for parameter estimates; however, existing Bayesian transfer learning methods are typically limited to single-source scenarios or require individual-level data. We introduce TRansfer leArning via guideD horseshoE prioR (TRADER), a novel approach enabling multi-source transfer through pre-trained models in high-dimensional linear regression. TRADER shrinks target parameters toward an adaptively weighted average of source estimates, while remaining robust to differences in source scale and correlation. Theoretical investigation shows that TRADER achieves faster posterior contraction rates than standard continuous shrinkage priors when sources are well-aligned with the target, while preventing negative transfer from heterogeneous sources. The finite-sample marginal posterior behavior of TRADER is established. Extensive simulations show that TRADER achieves inference performance no worse than using target data alone, performs comparably to a competing frequentist method despite using only summary-level source data, and offers substantial computational advantages. Application to a high-dimensional genetic dataset further shows TRADER's effectiveness in inference under strong multicollinearity. Supplementary materials for this article, including a standardized reproducibility guide detailing all materials required to replicate the results, are available online. *(abstract via PubMed)*

SiM Spatial & Time Series · 发表于 2026-08-23

Local Linear Estimation for Covariate‐Dependent Coefficients Model in Disease Mapping

Yexuan Jiang, Pei‐Sheng Lin, Jun Zhu, Feng‐Chang Lin

摘要

Spatial regression effects may depend on covariates in disease‐mapping models. For example, the association between infectious disease incidence and risk factors may vary according to climatic factors, such as time, season, and temperature. In this study, we aim to develop a spatial model that accounts simultaneously for the spatial and varying effects of covariates believed to modulate the spatial association. We employ a local linear estimation method to estimate covariate‐dependent coefficients in a model for excess zero counts. The local linear estimator effectively smooths covariate‐dependent coefficient estimation. Comprehensive simulation studies were conducted to evaluate the performance of the local linear estimators, and reported dengue cases in villages in Kaohsiung City from January 2014 to December 2015 are used to inform our proposed method for practical use in real‐world applications.

JASA Meta-Analysis & Data Integration Nonparametric & Semiparametric Causal Inference · 发表于 2026-08-21,收录于 2026-09-24

Transfer Learning of CATE with Kernel Ridge Regression

Seok-Jin Kim, Hongjie Liu, Molei Liu, Kaizheng Wang

摘要

(from arXiv preprint 2502.11331, an earlier working-paper draft of this now-published paper, since tandfonline.com blocked direct access — content may differ slightly from the final published wording) "The proliferation of data has sparked significant interest in leveraging findings from one study to estimate treatment effects in a different target population without direct outcome observations. However, the transfer learning process is frequently hindered by substantial covariate shift and limited overlap between (i) the source and target populations, as well as (ii) the treatment and control groups within the source. We propose a novel method for overlap-adaptive transfer learning of conditional average treatment effect (CATE) using kernel ridge regression (KRR). Our approach involves partitioning the labeled source data into two subsets. The first one is used to train candidate CATE models based on regression adjustment and pseudo-outcomes. An optimal model is then selected using the second subset and unlabeled target data, employing another pseudo-outcome-based strategy. We provide a theoretical justification for our method through sharp non-asymptotic MSE bounds, highlighting its adaptivity to both weak overlaps and the complexity of the CATE function. Extensive numerical studies confirm that our method achieves superior finite-sample efficiency and adaptability. We conclude by demonstrating the effectiveness and superiority of our approach on two real-world datasets.

JASA Other · 发表于 2026-08-21

Selecting the number of components in PCA via random signflips

David Hong, Yue Sheng, Edgar Dobriban

摘要

Principal component analysis (PCA) is a foundational tool in modern data analysis, and a crucial step in PCA is selecting the number of components to keep. However, classical selection methods (e.g., scree plots, parallel analysis, etc.) lack statistical guarantees in the increasingly common setting of large-dimensional data with heterogeneous noise, i.e., where each entry may have a different noise variance. Moreover, it turns out that these methods, which are highly effective for homogeneous noise, can fail dramatically for data with heterogeneous noise. This paper proposes a new method called signflip parallel analysis (FlipPA) for the setting of approximately symmetric noise: it compares the data singular values to those of “empirical null” matrices generated by flipping the sign of each entry randomly with probability one-half. We develop a rigorous theory for FlipPA, showing that it has nonasymptotic type I error control and that it consistently selects the correct rank for signals rising above the noise floor in the large-dimensional limit (even when the noise is heterogeneous). We also rigorously explain why classical permutation-based parallel analysis degrades under heterogeneous noise. Finally, we illustrate that FlipPA compares favorably to state-of-the-art methods via numerical simulations and an illustration on data coming from astronomy. *(abstract via OpenAlex)*

JASA Networks & Graphs High-Dimensional · 发表于 2026-08-21

Penalized Network Cross-Validation for Nested Models by Edge-Sampling

Bokai Yang, Yuanxing Chen, Yuhong Yang

摘要

In the network literature, a wide range of statistical models has been proposed to exploit structural patterns in the data. Therefore, model selection between different models is a fundamental problem. However, systematic theoretical understanding remains limited when comparisons involve different model classes. To address this challenging issue, we propose a penalized edge-sampling cross-validation framework for nested network model selection. By incorporating a model complexity penalty into the evaluation process, our method effectively mitigates the overfitting tendency of cross-validation and adapts to varying model structures. This framework supports comparisons among widely used models, including stochastic block models (SBMs), degree-corrected SBMs (DCBMs), and graphon models, providing the first consistency guarantees for model selection across these settings to the best of our knowledge. Empirical evaluations, including both simulated data and the “Political Books” network, demonstrate that our method yields stable and accurate performance across various scenarios. *(abstract via OpenAlex)*

JASA Causal Inference Longitudinal Data · 发表于 2026-08-21

Longitudinal weighted and trimmed treatment effects with flip interventions

Alec McClean, Alexander W. Levis, Nicholas Williams, Iván Díaz

摘要

Weighting and trimming are popular methods for addressing positivity violations in causal inference. While well-studied with single-timepoint data, standard methods do not easily generalize to longitudinal data and remain vulnerable to time-varying positivity violations. We extend weighting and trimming to longitudinal settings via stochastic “flip” interventions, which maintain the treatment status of subjects who would have received the target treatment, and flip others’ treatment to the target with probability equal to their weight (e.g., overlap weight, trimming indicator). In single-timepoint data, we show that flip interventions yield a broad class of weighted average treatment effects, providing a novel policy interpretation to these estimands. For longitudinal data, flip interventions enable interpretable weighting or trimming on time-varying covariates and, crucially, can ensure identifiability under arbitrary positivity violations. We derive a flexible estimator based on efficient influence functions when weights are smooth functions of propensity scores. Namely, we construct a new sequentially second-order estimator achieving root-n consistency and asymptotic normality under nonparametric conditions. Finally, we illustrate our methods with simulations and an analysis of the effect of union membership on earnings. *(abstract via OpenAlex)*

JASA Optimization & Learning Theory Meta-Analysis & Data Integration Nonparametric & Semiparametric · 发表于 2026-08-21

Locally Private Nonparametric Contextual Multi-armed Bandits with Transfer Learning

Yuheng Ma, Feiyu Jiang, Zifeng Zhao, Hanfang Yang, Yi Yu

摘要

Motivated by privacy concerns in sequential decision-making on sensitive data, we address the challenging problem of nonparametric contextual multi-armed bandits (MAB) under local differential privacy (LDP). Via a novelly designed LDP-compatible confidence bound, we propose an algorithm that achieves near-optimal regret performance, whose optimality is further supported by a first-seen minimax lower bound. We further consider the case of private transfer learning where auxiliary datasets are available, subject also to (heterogeneous) LDP constraints. Under the widely-used covariate shift framework, we propose a jump-start scheme, accompanied with a reweighted LDP-compatible estimator and confidence bound, which effectively combine information from heterogeneous auxiliary data. The minimax optimality of the algorithm is further established by a matching lower bound. Comprehensive experiments on both synthetic and real-world datasets validate our theoretical results and underscore the effectiveness of the proposed methods. *(abstract via OpenAlex)*

JASA Other · 发表于 2026-08-21

Incremental effects for continuous exposures

Kyle Schindl, Shuying Shen, Edward H. Kennedy

摘要

Causal inference problems often involve continuous treatments, such as dose, duration, or frequency. However, identifying and estimating standard dose-response estimands requires that everyone has some chance of receiving any level of the exposure (i.e., positivity). To avoid this assumption, we consider stochastic interventions based on exponentially tilting the treatment distribution by some parameter δ (an incremental effect); this increases or decreases the likelihood that a unit receives a given treatment level. We derive the efficient influence function and semiparametric efficiency bound for these incremental effects under continuous exposures. We then show that estimation depends on the size of the tilt, as measured by δ. In particular, we derive new minimax lower bounds illustrating how the best possible root mean squared error scales with an effective sample size of n/δ, instead of n. Further, we establish new convergence rates and bounds on the bias of double machine learning-style estimators. Our novel analysis gives a better dependence on δ compared to standard analyses by using mixed supremum and L2 norms. Finally, we define a “reflected” exponential tilt around any interior point and show that taking δ→∞ yields a new estimator of the dose-response curve across the treatment support. *(abstract via OpenAlex)*

SMMR Clinical Trials Meta-Analysis & Data Integration · 发表于 2026-08-21

Inclusive designs that allow the inclusion of a broader population in randomized controlled trials: An evaluation of randomization ratio and analysis strategy

Kim May Lee, Ziyan Wang, Richard Emsley, Nigel Stallard

摘要

In most multi-arm randomized controlled trials, all participants need to be eligible to be randomized to all the arms. This means some participants are excluded from participating in multi-arm studies. We propose to relax this by considering a ‘differential randomization’ approach that allows the inclusion of participants who are eligible for some but not all arms, as in some innovative trial approaches. We refer to a multi-arm design that employs differential randomization as an ‘inclusive design’ because it maximizes participant inclusion. Considering superiority comparisons for a normal endpoint, we evaluate the performance of some analysis methods for a three-arm inclusive design by simulation studies. Methods include pairwise or overall regression analysis, an averaging approach that pools summary statistics using prevalence rate, and a meta-analysis type framework. We compare the statistical power when the inclusive design employs different treatment allocation schemes. We find that using equal allocation ratio within each subpopulation leads to a higher disjunctive power than using equal allocation across arms, when the many-to-one comparisons are analysed using the corresponding pairwise data separately as opposed to using all data in a single regression analysis. We find that differential covariate–outcome relations among patients can affect the property of the inference.

JASA High-Dimensional · 发表于 2026-08-21,收录于 2026-09-24

High-dimensional Spectrum Matching: Covariance Inference Based on Universality

Not retrieved — tandfonline page blocked; no confidently-matching preprint located

摘要

We propose a new spectrum matching framework based on universality for high-dimensional covariance inference. The method constructs reference distributions for a broad class of spectral statistics by generating an auxiliary Gaussian matrix whose covariance spectrum matches that of the target covariance, with validity justified by universality. This framework accommodates both known and unknown covariance settings and adapts to different inferential targets, including covariance testing, spike detection, and simultaneous inference for linear functionals of the covariance matrix. Theoretical guarantees are established through anisotropic local laws for general covariance-type random matrices beyond the proportional-growth regime. Numerical studies and a real-data application demonstrate the practical effectiveness of the proposed methodology in high-dimensional settings. *(abstract via OpenAlex)*

JASA Latent Variables & Mixtures Longitudinal Data · 发表于 2026-08-21

Beyond Overall Diagnosis: A Longitudinal Multi-Domain Latent Structure Model for Domain-Specific Parkinson’s Disease Progression

Zijian Ye, Yuanjia Wang, Yu Gu, Kai Kang

摘要

Parkinson’s disease (PD) progresses heterogeneously, with motor, cognitive, and other non-motor symptoms often worsening at different times. Conventional models based on overall diagnoses or composite scores can obscure this asynchronous progression and limit individualized prognosis and care. Clinically, a domain-specific progression time marks when a patient meaningfully worsens in one symptom domain, helping clinicians identify when focused monitoring or care may be needed. We propose a longitudinal multi-domain latent structure model that groups PD-related markers into clinically distinct domains and represents progression within each domain through a shared latent disease state, while allowing latent disease states to differ across domains. The unobserved timing of the domain-specific progression is inferred from repeated clinical measurements whose distributions differ before and after progression, and baseline risk factors are related to this timing through a Cox proportional hazards model. We establish theoretical guarantees for recovering the underlying grouping of clinical markers into domains. Applied to longitudinal measurements of 30 retained MDS-UPDRS items from 849 PPMI participants, the method identifies five clinically interpretable symptom domains and individualized domain-specific progression trajectories. Faster progression is associated with male sex in selected domains and with smaller baseline putamen volume across domains, whereas education shows no clear association. *(abstract via OpenAlex)*

JASA Testing & Uncertainty · 发表于 2026-08-21

A Simple and Adaptive Confidence Interval when Nuisance Parameters Satisfy an Inequality

Gregory Fletcher Cox

摘要

Inequalities may appear in many models. They can be as simple as assuming a parameter is nonnegative, possibly a regression coefficient or a treatment effect. This paper focuses on models with one inequality and proposes an inequality-imposed confidence interval (IICI) that has particularly attractive features. The IICI is simple in that it does not require simulations or tuning parameters. Also, the IICI is adaptive to the slackness of the inequality, uniformly valid, and never longer than the usual confidence interval. We demonstrate the IICI in two empirical applications. The first empirical application considers a regression when a coefficient is known to be nonpositive. The second empirical application considers an instrumental variables regression when the endogeneity of a regressor is known to be nonnegative. *(abstract via OpenAlex)*

SiM Longitudinal Data Latent Variables & Mixtures · 发表于 2026-08-20

Using Latent Representations to Link Disjoint Longitudinal Data for Mixed‐Effects Regression

Clemens Schächter, Maren Hackenberg, Michelle Pfaffenlehner, Félix B. Tambe‐Ndonfack, Thorsten Schmidt, Astrid Pechmann, Janbernd Kirschner, Jan Hasenauer, Harald Binder

摘要

Many rare diseases offer limited established treatment options, leading patients to switch therapies when new medications emerge. To analyze the impact of such treatment switches within the low sample size limitations of rare disease trials, it is important to use all available data sources. This, however, is complicated when the use of measurement instruments changes during the observation period, for example when instruments are adapted to specific age ranges. The resulting disjoint longitudinal data trajectories complicate the application of traditional modeling approaches like mixed‐effects regression. We tackle this by mapping observations of each instrument to an aligned low‐dimensional temporal trajectory, enabling longitudinal modeling across instruments. Specifically, we employ a set of variational autoencoder architectures to embed item values into a shared latent space for each time point. Temporal disease dynamics and treatment switch effects are then captured through a mixed‐effects regression model applied to latent representations. To enable statistical inference, we present a novel statistical testing approach that accounts for the joint parameter estimation of mixed‐effects regression and variational autoencoders. The methodology is applied to quantify the impact of treatment switches for patients with spinal muscular atrophy. Here, our approach aligns motor performance items from different measurement instruments for mixed‐effects regression and maps estimated effects back to the observed item level to quantify the treatment switch effect. Our approach allows for model selection as well as for assessing effects of treatment switching. The results highlight the potential of modeling in joint latent representations for addressing small data challenges.

Biometrika Nonparametric & Semiparametric · 发表于 2026-08-20,收录于 2026-09-03

Tight differencing in spectral density estimation with centrosymmetric kernels

Y. Wang, K. W. Chan

摘要

(verbatim, from OUP article page): "Mean-robust estimation of spectral density and long-run variance is crucial for many statistical inference procedures. However, existing methods often degrade when serially dependent data exhibit volatile, time-varying trends, or sudden jumps, particularly in small samples. While differencing and kernel averaging are standard tools for achieving mean robustness and consistency, they are not inherently compatible. Combining them can compromise optimality. Specifically, tight differencing, an operation of taking small-lag differences to enhance local de-trending, introduces strong correlations that distort the high-order properties of kernel-averaged estimators. To resolve this incompatibility, we introduce a novel class of centrosymmetric kernels explicitly designed to integrate with tight differencing. We demonstrate that the optimal tight difference sequence for serially dependent data differs from classical sequences designed for independent data. Notably, these proposed optimal sequences are data-independent and can be applied directly without pre-fitting. Finally, the proposed estimators are demonstrated to be useful across various statistical inference tasks, including tests for stationarity and white noise.

AoS Causal Inference Nonparametric & Semiparametric · 发表于 2026-08-20

Nuisance Function Tuning and Sample Splitting for Optimally Estimating a Doubly Robust Functional

Sean McGrath, Rajarshi Mukherjee

摘要

Estimators of doubly robust functionals typically rely on estimating two complex nuisance functions, such as the propensity score and conditional outcome mean for the average treatment effect functional. We consider the problem of how to estimate nuisance functions to obtain optimal rates of convergence for a doubly robust nonparametric functional that has witnessed applications across the causal inference and conditional independence testing literature. For several plug-in estimators and a first-order bias-corrected estimator, we illustrate the interplay between different tuning parameter choices for the nuisance function estimators and sample splitting strategies on the optimal rate of estimating the functional of interest. For each of these estimators and each sample splitting strategy, we show the necessity to either undersmooth or oversmooth the nuisance function estimators under low regularity conditions to obtain optimal rates of convergence for the functional of interest. Unlike the existing literature, we show that plug-in and first-order bias-corrected estimators can achieve minimax rates of convergence across all Hölder smoothness classes of the nuisance functions by careful combinations of sample splitting and nuisance function tuning strategies. We complement these results with numerical simulations illustrating the impact of different nuisance function tuning and sample splitting strategies." *Publication/online date: date unknown (IMS Future Papers list gives no per-item date; formal AoS issue not yet assigned)*

Biometrics Nonparametric & Semiparametric · 发表于 2026-08-20,收录于 2026-09-17

Generalized Local Kendall's τ: a novel framework for uncovering nonlinear local dependence

Zaixin Huang, Zhengjun Zhang

摘要

Identifying patient subgroups is increasingly vital for advancing personalized and precision medicine, where patient heterogeneity plays a central role. Subgroups can be revealed by examining how predictors depend on each other. However, traditional global dependence measures, such as Kendall's τ, often overlook localized and nonlinear relationships, potentially missing critical patterns that define subpopulations. Existing local dependence measures are typically constrained to linear structures or lack flexibility in capturing complex local behaviors. We propose a novel, intuitive, and highly flexible framework, the Generalized Local Kendall's τ, for assessing nonlinear local dependence. This method allows local regions to take diverse shapes (e.g., rectangles, squares, extreme-value neighborhoods), thereby unifying global, local, and tail dependence within a flexible and interpretable structure. The framework includes U-statistic-based and copula model-based estimators, introducing four new quantile dependence measures and novel tail dependence coefficients. We develop two visualization tools—Type I and Type II local Kendall's τ surfaces—applied to Archimedean copulas to aid interpretation. Comparisons show that global Kendall's τ often differs markedly from its local versions, emphasizing the added value of local analysis. Simulation results indicate that when the model is correctly specified, the copula-based estimator exhibits significantly superior performance. We illustrate the framework's practical utility using Parkinson's disease and COVID-19 datasets, showing its ability to detect subgroup-specific dependence patterns vital for precision medicine." *Publication/online date: 20 August 2026*

Biometrics Clinical Trials · 发表于 2026-08-20,收录于 2026-09-17

Blinded sample size recalculation in randomized controlled trials with analysis of covariance

Takumi Kanata, Yasuhiro Hagiwara, Koji Oba

摘要

In randomized controlled trials, covariate adjustment can improve statistical power and reduce the required sample size compared with unadjusted estimators. Analysis of covariance (ANCOVA) is often used to adjust for baseline covariates when outcomes are continuous. To design a sample size based on ANCOVA, it is necessary to prespecify the association between the outcome and baseline covariates, as well as among the baseline covariates themselves. However, determining these parameters at the design stage is challenging. Although it may be possible to adaptively assess these during the trial and recalculate the required sample size, existing sample size recalculation methods assume that the joint distribution of the outcome and baseline covariates is multivariate normal, which is not always the case in practice. In this study, we propose a blinded sample size recalculation method for the ANCOVA estimator based on the asymptotic relative efficiency under minimal distributional assumptions, thus accommodating arbitrary model misspecification. The proposed method is able to achieve the nominal power and reduce the required sample size without inflating the type I error rate in the final analysis. We conducted simulations to evaluate the performance of the proposed method under various scenarios. Applied to an HIV clinical trial, the proposed method reduced the required sample size by up to one-third compared to the unadjusted analysis, despite occasional slight increases." *Publication/online date: 20 August 2026*

JASA Other · 发表于 2026-08-19,收录于 2026-05-17

Seminal Ideas and Controversies in Statistics (by Roderick J. A. Little)

Jonathan P Williams

摘要

1 IntroductionThis book truly is essential reading for all statisticians, PhD students in particular. Historical context is generally lacking in statistics PhD programs, and so this book fills an i... *(abstract via OpenAlex)*

Biometrics Meta-Analysis & Data Integration Longitudinal Data · 发表于 2026-08-19

Random-effects meta-analysis via generalized linear mixed models: A Bartlett-corrected approach for few studies

Keisuke Hanada, Tomoyuki Sugimoto

摘要

Random-effects models are central to meta-analysis, yet the between-study variance is often underestimated when the number of studies is small. In such settings, confidence intervals become unduly narrow and fail to attain the nominal coverage probability. Although several small-sample corrections, including the Bartlett correction, have been developed under the normal–normal model, corresponding methodology for generalized linear mixed models (GLMMs) remains limited. This study proposes a unified framework for random-effects meta-analysis within the GLMM framework that relies exclusively on aggregate data and accommodates outcomes following any distribution in the exponential family, including the binomial, Poisson, and gamma distributions. To improve interval estimation with few studies, we develop a profile likelihood method with a simplified Bartlett correction (PLSBC), which refines the chi-squared approximation of the profile likelihood ratio statistic without requiring higher-order derivatives. We show theoretically that the proposed estimators preserve the consistency and asymptotic normality of the maximum likelihood estimators. Simulation studies demonstrate that the PLSBC yields nearly unbiased estimates and maintains nominal coverage across a variety of outcome types. Applications to three published meta-analyses with binomial, Poisson, and gamma outcomes indicate that the proposed approach provides robust and interpretable inference with few studies. The PLSBC therefore offers a practical and broadly applicable framework for random-effects meta-analysis when the number of studies is limited.

JASA Networks & Graphs · 发表于 2026-08-19

Network Regression and Supervised Centrality Estimation

Junhui Cai, Dan Yang, Ran Chen, Haipeng Shen, Linda Zhao, Wu Zhu

摘要

The centrality in a network is often used to measure nodes’ importance and model network effects on a certain outcome. Empirical studies widely adopt a two-stage procedure, which first estimates the centrality from the observed noisy network and then infers the network effect from the estimated centrality, even though it lacks theoretical understanding. We propose a unified modeling framework to study the properties of centrality estimation and inference and the subsequent network regression analysis with noisy network observations. Furthermore, we propose a supervised centrality estimation methodology, which aims to simultaneously estimate both centrality and network effect. We showcase the advantages of our method compared with the two-stage method both theoretically and numerically via extensive simulations and a case study in predicting currency risk premiums from the global trade network. *(abstract via OpenAlex)*

JASA Survival Analysis Machine Learning Networks & Graphs · 发表于 2026-08-19

Mini-batch Estimation for Deep Cox Models: Statistical Foundations and Practical Guidance

Lang Zeng, Weijing Tang, Zhao Ren, Ying Ding

摘要

The stochastic gradient descent (SGD) algorithm has been widely used to optimize deep Cox neural network (Cox-NN) by updating model parameters using mini-batches of data. We show that SGD aims to optimize the average of mini-batch partial-likelihood, which is different from the standard partial-likelihood. This distinction requires developing new statistical properties for the global optimizer, namely, the mini-batch maximum partial-likelihood estimator (mb-MPLE). We establish that mb-MPLE for Cox-NN is consistent and achieves the optimal minimax convergence rate up to a polylogarithmic factor. For Cox regression with linear covariate effects, we further show that mb-MPLE is n -consistent and asymptotically normal with asymptotic variance approaching the information lower bound as batch size increases, which is confirmed by simulation studies. Additionally, we offer practical guidance on using SGD, supported by theoretical analysis and numerical evidence. For Cox-NN, we demonstrate that the ratio of the learning rate to the batch size is critical in SGD dynamics, offering insight into hyperparameter tuning. For Cox regression, we characterize the iterative convergence of SGD, ensuring that the global optimizer, mb-MPLE, can be approximated with sufficiently many iterations. Finally, we demonstrate the effectiveness of mb-MPLE in a large-scale real-world application where the standard MPLE is intractable. *(abstract via PubMed)*

JASA Machine Learning · 发表于 2026-08-19

Low-Rank Contextual Reinforcement Learning from Heterogeneous Human Feedback

Seong Jin Lee, Will Wei Sun, Yufeng Liu

摘要

Reinforcement learning from human feedback (RLHF) has become a cornerstone for aligning large language models with human preferences. However, the heterogeneity of human feedback, driven by diverse individual contexts and preferences, poses significant challenges for reward learning. To address this, we propose a Low-rank Contextual RLHF (LoCo-RLHF) framework that integrates contextual information to better model heterogeneous feedback while maintaining computational efficiency. Our approach builds on a contextual preference model, leveraging the intrinsic low-rank structure of the interaction between user contexts and query-answer pairs to mitigate the high dimensionality of feature representations. Furthermore, we address the challenge of distributional shifts in feedback through our Pessimism in Reduced Subspace (PRS) policy, inspired by pessimistic offline reinforcement learning techniques. We theoretically demonstrate that our policy achieves a tighter sub-optimality gap compared to existing methods. Extensive experiments, ranging from synthetic simulations to an analysis of the real-world PersonalLLM benchmark, validate the effectiveness of LoCo-RLHF and demonstrate its superior performance in personalized settings. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work. *(abstract via OpenAlex)*

SiM Clinical Trials · 发表于 2026-08-19

Long Run Non‐Transitivity of Win Ratio and How to Rule It Out

Olga V. Demler, Anna Győrffy‐Kerekes

摘要

The win ratio (WR) has emerged as an increasingly used method for analyzing composite endpoints in randomized controlled trials (RCTs). Its growing popularity stems from its ability to accommodate a hierarchy of event types within a composite primary endpoint in an RCT. However, as a relatively new measure, the WR exhibits some surprising properties. Oakes and other authors have noted that for a given triplet of participants, there are realizations of event times and hierarchy of event types when “wins” form a loop. They term this property “intransitivity.” In this paper, we report and study a different phenomenon, a violation of long‐run transitivity of WRs, for example, when in the long run , and paradoxically , where denote the time‐to‐event in trial arms A, B, and C (or placebo, standard of care and novel treatment, respectively). We refer to it as Efron‐type “non‐transitivity.” Unlike Oakes‐type intransitivity, Efron‐type non‐transitivity is an inherent property of the underlying random variables. As such, it cannot be resolved by a longer observation period or larger sample size. Lumley and Gillen showed that transitivity can be guaranteed only for univariate statistics (e.g., mean, median, etc.), whereas head‐to‐head comparisons can be non‐transitive. WR and win proportion (WP) are head‐to‐head comparisons and may be non‐transitive, in which case the best treatment cannot be determined. Therefore, the non‐transitivity property challenges the utility of all head‐to‐head comparisons as efficacy measures in RCTs. In this paper, we explore the scope of this problem, focusing on identifying conditions that rule out non‐transitivity for WR, WP, and hazard ratio (HR). We apply our findings to establish transitivity in two case studies using published RCTs. This work highlights an important phenomenon that must be recognized when the WR is used as an efficacy measure in randomized controlled trials.

JASA Causal Inference · 发表于 2026-08-19,收录于 2026-04-24

Learning When the Concept Shifts: Confounding, Invariance, and Dimension Reduction

Kulunu Dharmakeerthi, YoonHaeng Hur, Tengyuan Liang

摘要

Practitioners often face the challenge of deploying prediction models in new environments with shifted distributions of covariates and responses. With observational data, such shifts are often driven by unobserved confounding, and can in fact alter the concept of which model is best. This article studies distribution shifts in the domain adaptation problem with unobserved confounding. We postulate a linear structural causal model to account for endogeneity and unobserved confounding, and we leverage exogenous invariant covariate representations to cure concept shifts and improve target prediction. We propose a data-driven representation learning method that optimizes for a lower-dimensional linear subspace and a prediction model confined to that subspace. This method operates on a non-convex objective—that interpolates between predictability and stability—constrained to the Stiefel manifold, using an analog of projected gradient descent. We analyze the optimization landscape and prove that, provided sufficient regularization, nearly all local optima align with an invariant linear subspace resilient to distribution shifts. This method achieves a nearly ideal gap between target and source risk. We validate the method and theory with real-world datasets to illustrate the tradeoffs between predictability and stability. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work. *(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