Monday, Aug 3: 2:00 PM - 3:50 PM
Contributed Posters
Thomas M. Menino Convention & Exhibition Center
Room: CC-Exhibit Hall A
Main Sponsor
Scientific and Public Affairs Advisory Committee
Presentations
Accurately assigning exposure scores to open-ended categories is critical in dose-response meta-analyses (DRMAs). However, the current usage of scoring methods is complete chaos, resulting in inconsistent, non-reproducible, and unreliable results. Through a systematic review, we summarized at least 11 commonly used methodologies. In addition, to identify current trends, we analyzed 158 DRMA publications from the past three years and listed the five most popular methods. We found that none of these methods were used with sufficient justification, and that even more surprisingly, over 39% of publications failed to report their scoring approach. In this work, we propose a novel recursive method that assigns exposure scores based on the sample sizes of categories, which provides a more unified, data-driven framework for exposure scoring, enhancing the accuracy, transparency, and reliability of DRMAs. We performed two case studies to highlight the method's practical utility and robustness. In addition, we further provide guidance on selecting scoring methods in various scenarios.
Keywords
Meta-analysis
Dose–response relationships
Grouped dose levels
Trend estimation
Speaker
Yida Gu, Beijing Normal-Hong Kong Baptist University
Co-Author(s)
Zhijian Li, Beijing Normal-Hong Kong Baptist University
Tiejun Tong, Hong Kong Baptist University
Xiao Ke, Shenzhen Technology University
Guoliang Tian, Southern University of Science and Technology
Yuedong Wang, Univ. of California At Santa Barbara
Predictions from machine learning (ML) models are increasingly used to impute missing data, but their naive use risks biased inference. Existing methods provide valid inference with ML imputations regardless of prediction quality to enhance efficiency, while they are limited to missing outcomes under a missing-completely-at-random assumption. We develop a novel framework for valid statistical inference in Z-estimation problems using ML imputations under a missing-at-random assumption and for general missingness patterns. The method stratifies data by distinct missingness patterns and constructs an estimator by appropriately weighting and aggregating pattern-specific information. We establish the asymptotic theory and provide a theoretical guarantee on efficiency dominance over weighted complete-case analyses (WCCA). Practically, the method affords simple implementations by leveraging existing WCCA software. Extensive simulations are carried out to validate theoretical results. An analysis of \textit{All of Us} data further shows the practical utility of the method. The paper concludes with a brief discussion on practical implications, and potential future directions.
Keywords
All of Us
Machine learning imputation
Missing data
Prediction-based inference
Z-estimation
Accurate ICU outcome prediction increasingly relies on longitudinal electronic health record data, yet clinical decisions require timely and interpretable models that avoid dependence on late-arriving measurements. We propose a Temporal Nested LASSO (TNL) framework for hierarchical variable selection in longitudinal settings, enforcing an "earlier-before-later" structure so that later measurements enter the model only if earlier information is retained. The method simultaneously enables group-level selection of time windows and within-window sparsity among predictors, yielding compact and temporally consistent models. Optional time-dependent penalty weights further prioritize early, clinically actionable predictors. We extend the framework to generalized linear models to accommodate non-Gaussian ICU outcomes such as mortality. Simulation studies and applications to multi-center ICU data demonstrate that TNL achieves a favorable balance between predictive accuracy, sparsity, and timeliness compared with existing nested and group-structured penalties, supporting earlier and more interpretable clinical prediction.
Keywords
Nested LASSO
Hierarchical variable selection
Timely prediction
ICU outcomes
GLM
We investigate statistical inference for treatment effects in two-stage designs under covariate-adaptive randomization. We select the treatment with the most promising estimated effect from multiple experimental arms in the initial stage and use a controlled study to confirm the efficacy of the selected treatment in the second stage. However, this setting faces two key challenges: (i) obtaining valid and efficient inference results under covariate-adaptive randomization, and (ii) correcting for the selection bias introduced during the selection in the first stage. To address these, we develop an inference procedure under covariate-adaptive randomization without discretizing continuous covariates, enabling more efficient inference results for two-stage designs. Additionally, we employ a resampling method to mitigate the selection bias. Numerical studies show that the proposed approach often improves performance over the existing approaches for inference under a two-stage design employing covariate-adaptive randomization.
Keywords
Clinical trial
Covariate adaptive-randomization
Two-stage design
Longitudinal observational studies are increasingly used to evaluate treatment effectiveness in real-world settings, yet valid causal inference remains challenging when covariates evolve over time. Treatment groups in longitudinal studies may diverge over time due to differences in treatment intensity, adherence, disease progression, or behavioral changes, resulting in longitudinal trajectory divergence that can increase bias and variance in treatment effect estimation. In this paper, we focus on a hybrid design representing a naturalistic follow-up study after initial randomization, with fixed treatment and time-varying outcomes and covariates. We propose an extreme value theory-based stable balancing weights method (EVT-SBW) that directly targets covariate balance through convex optimization and penalizes individual weights using EVT when overlap is limited due to extreme trajectory imbalance. Unlike ad hoc trimming or approaches that redefine the target estimand, EVT-SBW aims to improve efficiency and stability without discarding observations. Simulation studies demonstrate that EVT-SBW reduces bias and RMSE across a wide range of longitudinal scenarios with limited overlap.
Keywords
longitudinal studies
convex optimization
trajectory imbalance
extreme value theory
limited overlap
Speaker
Jooyeon Lee, The University of Texas Health Science Center at Houston
Co-Author
Evan Kwiatkowski, The University of Texas Health Science Center at Houston
Endpoint selection for progressive diseases like Alzheimer's disease is often hindered by rigorous regulatory validation requirements, including construct validity, internal consistency, and external responsiveness. Extensive endpoint validation trials are costly and can delay the delivery of effective treatments to patients by leaving sponsors with a choice of sub-optimal outcomes which require larger sample sizes and lead to higher risk of study failure. While traditional metrics are valuable, the mean-to-standard-deviation ratio (MSDR), a standardized ratio of control arm progression, offers a universal evaluation for progressive diseases. It is mathematically demonstrated how a high MSDR inherently indicates that an endpoint fulfills traditional validation criteria such as those mentioned previously. By prioritizing MSDR, regulators and sponsors can identify sensitive outcomes and evaluate investigational products concurrently with the formal validation process. This approach streamlines drug development without compromising statistical rigor, ensuring that the most responsive tools are used for measuring treatment effects.
Keywords
Alzheimer's Disease
Progressive Diseases
Clinical Endpoints
Endpoint Validation
The area under the ROC curve (AUC) is the standard measure of a biomarker's discriminatory accuracy; however, AUC is rarely treated as a population-specific estimand. When validation cohorts differ from the intended target population in case mix, na\"ive AUC estimates can mislead both generalization and cross-study comparison. We develop an estimand-focused framework that anchors biomarker AUC inference to a prespecified target population, aligning with the ICH E9(R1) estimand perspective adapted to discrimination rather than treatment effect. The framework supports two scientific goals: generalizing a study-specific AUC to a clinically relevant target population, and benchmarking AUCs across studies on a common population footing. Methodologically, we extend calibration weighting to the U-statistic formulation of AUC, proposing a family of estimators that accommodate either patient-level or summary-level target covariate information; augmented variants attain double robustness. We establish asymptotic properties and study their performances through comprehensive simulations. Furthermore, we demonstrate the proposed framework on the POWER trials, evaluating baseline stair-climb power (SCP) as a prognostic marker for 6-month survival in advanced non-small-cell lung cancer (NSCLC). Unlike prior work on transporting model-based predictive accuracy, our framework targets the biomarker-level estimand directly and addresses cross-study comparability — an issue not resolved by current methods.
Keywords
Biomarker evaluation
Calibration weighting
Covariate shift
Diagnosis accuracy
Prediction medicine
U-statistics
Speaker
Jiajun Liu, Duke University School of Medicine
Co-Author(s)
Guangcai Mao, School of Mathematics and Statistics, Central China Normal University
Xiaofei Wang, Duke University Medical Center
The restricted mean survival time (RMST), defined as the mean event-free survival time within a prespecified horizon [0, tau], provides an interpretable estimand for comparing event times that remains valid under nonproportional hazards. Group sequential (GS) designs enable early stopping for efficacy and futility, while covariate adjustment can improve efficiency when baseline prognostic factors are available. We develop a covariate-adjusted GS testing framework for comparing RMSTs in randomized trials. RMST estimation is conducted by a generalized linear model fitted using inverse probability of censoring weighted estimating equations, allowing efficient adjustment without assuming proportional hazards. The covariate-adjusted GS test statistics are shown to possess an independent increments covariance structure asymptotically, permitting the use of standard methods to control the overall type I error rate. Simulation studies across various choices of link functions, censoring patterns and sample sizes show that good type I error control and nominal power levels are achieved. An application to the BMT CTN 1101 trial illustrates the proposed methodology in a real-world setting.
Keywords
Restricted mean survival time (RMST)
Group sequential design
Covariate adjustment
Independent increments
Blood and marrow transplantation
Inverse probability of censoring weighting
Spatial transcriptomics provides high-resolution gene expression maps across tissues. These biological advances necessitate statistical frameworks for characterizing spatially heterogeneous gene co-expression patterns. We introduce the novel concept of Spatially Co-expressed Genes (SCGs), which provide a spatial map of gene co-expression and enable the study of spot-specific gene networks, offering insights into disease etiology. To identify SCGs, we propose a spatial covariance regression (SCR) model that learns spatially varying marginal dependence while exploiting low-rank structure. SCR models location-specific correlation matrices using a spatially varying factor model with loadings decomposed into global loadings and spatial basis functions. We impose a multiplicative gamma shrinkage prior on the global loadings and Gaussian process on the spatial bases. SCGs are identified via Otsu's thresholding applied to spatial variation of estimated edges. We develop a GPU-accelerated variational Bayes algorithm to ensure scalability and assess the performance through comprehensive simulation study. We apply SCR on Xenium Alzheimer's mouse brain data to discover spatial biomarkers.
Keywords
Spatial transcriptomics
Spatial covariance modeling
Gene co-expression networks
Factor models
Gaussian processes
Variational Bayes inference
Neurodegenerative diseases such as Alzheimer’s disease progress gradually, with structural brain changes accumulating over years. Standard longitudinal imaging analyses based on voxelwise comparisons or region-level summaries are often fail to capture how structural patterns evolve across spatial scales. In this work, we study whether Alzheimer’s brain dynamics can be identified through longitudinal topological summaries of brain structure. Using longitudinal structural MRI from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), we extract persistent homology features from preprocessed brain volumes and represent them as betti curves across filtration thresholds. Treating betti curves as functional data, we align trajectories across visits using the square-root velocity function framework to reduce phase variability and then apply longitudinal functional principal component analysis to separate subject-specific baseline patterns from patterns of change over time. The analysis reveals stable topological signatures across visits and systematic time-varying components that capture longitudinal brain dynamics, reflect disease progression, and enable score-based sex comparisons.
Keywords
Longitudinal functional data analysis
Functional principal component analysis
Persistent homology
Betti curves
Square-root velocity function
We propose a scalable framework for conditional density regression based on logistic Gaussian processes. A barrier to scalable computation for these models has traditionally been the need to calculate observation dependent normalizing constants by numerical integration. We avoid this by using generalized Bayesian inference, replacing the negative log-likelihood with a Hyvärinen score. This score-based approach depends only on derivatives of the response's log density, eliminating the need to compute any normalizing constants. However, the required Gaussian process computations can still be computationally expensive. We address this using sparse inducing point variational approximations, making our method scalable to large datasets. Our nonparametric prior can be centered on an existing parametric model. The nonparametric corrections provide interpretable diagnostics that reveal inadequacies in the base model, such as missing covariate-dependent heteroscedasticity and skewness in the response distribution. We demonstrate good predictive performance and scalability through simulated and real examples, including a large spatio-temporal temperature dataset.
Keywords
Density regression
Generalized Bayesian inference
Variational inference
Logistic Gaussian pro-
cess
Misspecification
Population forecasts inform critical decisions in public policy and economic planning, yet existing state-of-the-art methods often underestimate predictive uncertainty by modeling key demographic variables, such as fertility rates and life expectancy, as independent stochastic processes. This work proposes a hierarchical Bayesian framework for probabilistic population forecasting that models the joint dynamics of fertility and mortality across countries, regions, and time. Dependence structure will be represented using copulas, which allow flexible joint modeling while preserving interpretable marginal structures. The hierarchical design enables partial pooling across countries and regions, allowing countries with sparse data to be partially informed by broader regional patterns of demographic change while retaining country-specific trajectories. By producing more faithful joint predictive uncertainty quantification, this work delivers uncertainty-aware population projections that better support evidence-based policy and equitable decision-making. The proposed methods will be illustrated using simulations and a data analysis.
Keywords
Bayesian Hierarchical Model
Dependence Modeling
Probabilistic Forecasting
Copula
Bayesian Demography
Applied Statistics
Romosozumab (Romo) is a monthly injectable osteoporosis therapy intended for 12 doses. Reliably predicting completion can inform targeted adherence support but is challenging in populations with lower fracture risk due to small samples and sparse covariates. Using U.S. Medicare claims (2019–2022), we identified women ≥65 initiating Romo with ≥15 months of follow-up. Completion was defined as receiving 12 doses of Romo within 15 months. We fit Bayesian hierarchical logistic models (BHM) to predict completion in Asian and Black users, borrowing information from White users. A heterogeneity parameter (τ²) quantified the degree of borrowing. Performance was evaluated via AUC and calibration plots. We identified 13,373 romo users (White: 12,094, Asian: 457, Black: 142). Predictors were selected by clinical experts and LASSO in White group, then incorporated as commensurate priors in BHM. AUCs were similar for Frequentist vs. Bayesian models. However, partial borrowing (τ²≈0.17–0.21) stabilized rare-predictor coefficients and improved calibration across risk deciles. Bayesian borrowing can produce more reliable subgroup predictions than frequentist models in small sample settings.
Keywords
Bayesian Hierarchical Modeling
Commensurate Prior
Treatment Adherence
Speaker
Hongke Wu, The University of Alabama at Birmingham
Co-Author(s)
Ye Liu, University of Alabama at Birmingham
Tarun Arora, University of Alabama at Birmingham
Jeffrey Curtis, The University of Alabama at Birmingham
Stroke remains a leading cause of death and disability in the USA, with timely treatment being crucial for optimal recovery. Endovascular therapy (EVT) is associated with a 67% reduction in stroke-related disability for eligible patients; however, its benefit is highly time sensitive. Conventional prehospital stroke protocols often direct patients to the nearest hospital, which may not be the optimal destination when specialized care is required.
MAP-STROKE is a prehospital destination-selection algorithm that predicts stroke type and treatment outcomes under alternative triage strategies. It uses a joint Bayesian model for diagnosis and outcomes, along with model-based imputation to harmonize data across clinical trials, and recommends the optimal initial transport destination to maximize recovery. Using a large-scale geospatial posterior-predictive simulation, representing over 400 million stroke alerts across the USA, MAP-STROKE improved outcomes for patients with large vessel occlusion (LVO), with particularly notable benefits in rural settings. These gains were accompanied by modest trade-offs for some non–LVO patients due to time-sensitive competing therapies.
Keywords
Bayesian Models
Machine Learning
Simulation
Stroke
Emergency medical services
Geospatial
We present a Bayesian vector autoregressive (BVAR) model designed for panel data. In small-area applications, traditional vector autoregressive (VAR) models quickly become overparameterized, and standard BVARs often rely on aggressive global shrinkage, risking over-shrinking meaningful regional dynamics. To address these challenges, we develop a spatial BVAR-CAR framework that combines global regularization with flexible local shrinkage. We introduce a conditional autoregressive prior on region-specific intercepts to capture spatial dependence and a hierarchical shrinkage (horseshoe-type) prior on autoregressive coefficients to borrow strength across regions, stabilizing estimation in high-dimensional settings. This setup eliminates redundant parameters while retaining heterogeneous dynamics across local areas. We evaluate forecasting performance using two annual panel datasets: average hourly earnings of production employees in California Metropolitan Statistical Areas (small areas) and gender unemployment gaps in African countries. Our BVAR-CAR model outperforms three benchmarks - a univariate AR(1) model, a restricted VAR model with shared hyperparameters, and an unrestricted BVAR without spatial and global-local shrinkage priors - in both panels. The results highlight the benefits of spatial pooling in Bayesian models when time series are short. By integrating cross-sectional structure with temporal dependence, our approach provides a flexible and interpretable solution for forecasting and small-area estimation in regional economic analysis.
Keywords
Bayesian predictive inference
Global-Local shrinkage prior
Hierarchical Bayesian model
Markov-chain Monte Carlo algorithm
Panel forecasting
Spatial CAR prior
In clinical studies, it is challenging to evaluate treatment efficacy when the outcome of interest is measured long-term or is otherwise difficult to measure. One solution is to replace the primary outcome with a more accessible surrogate endpoint (e.g. biomarker). Once validated, surrogate endpoints can accelerate clinical decision-making in future trials. In this work, we approach the validation of surrogates through a causal principal stratification framework with joint Gaussian outcomes. Assessment of surrogate endpoints is done using two causal quantities, average causal necessity (ACN) and average causal sufficiency (ACS), which we estimate using a Bayesian imputation procedure. However, estimating ACN and ACS via principal stratification relies on a correlation matrix that is not fully identifiable without making untestable assumptions. To improve identifiability, we propose incorporating conditional independence assumptions, such as through Bayesian priors on individual correlation parameters or by imposing them within Metropolis Hastings steps. We also condition on baseline covariates which may make the assumptions more plausible. Impacts on estimation are assessed.
Keywords
surrogate endpoints
principal stratification
Bayesian methods
Physical activity (PA) provides protective health benefits and helps inform the intervention design. Commonly used time-course PA summary metrics, such as the averaged step count, may sacrifice some key nuances of micro PA patterns that could serve as digital markers. In this project, we are interested in temporal associations between nutrient intake and daily PA distributions expressed in step counts. We extend Fréchet regression to longitudinal object data using working correlation matrices via the quadratic inference function method in Python, which enables correlation adjustment and bootstrap inference. Daily PA distributions are characterized by quantiles derived from empirical cumulative distribution functions. The method is illustrated using seven-day accelerometer data from 352 adolescents in the Early Life Exposure in Mexico to ENvironmental Toxicants (ELEMENT) Study and usual micronutrient intake from a validated food frequency questionnaire. Analyses were adjusted for the demographic and lifestyle variables. This presentation will report the effects of mean daily magnesium, vitamin C, and vitamin D intake on PA pattern distributions under different working correlations.
Keywords
Frechet regression
Quadratic inference function
Generalized estimating equations
Longitudinal data analysis
Metric-spaced objects
Accurate event prediction lies at the center of critical decision making in domains such as wildfire, earthquake and tropical cyclone prediction, where events are commonly modeled as point processes on a sphere. Recent advances in machine learning
have moved from point predictions to more complex structures including sets, distributions, and clouds of possible events, however, uncertainty quantification for these distributional structures remains a challenge.
We introduce a conformal prediction framework for spherical point processes based on optimal transport, utilizing spherical sliced Wasserstein as a conformity score along with a gradient based velocity field that generates ensembles of point clouds lying at
the boundary of our conformal region. By comparing predicted and observed event patterns in a geometrically faithful way, our method yields distribution-free prediction regions with finite sample validity. We demonstrate our approach on synthetic and real datasets with differing underlying manifolds, providing visual and quantitative uncertainty summaries over the space of point-process realizations.
Keywords
Conformal Prediction
Optimal Transport
Spherical Sliced Wasserstein
Uncertainty Quantification
Point Processes
Natural Disaster Forecasting
Understanding signal behavior across scales is important in applications such as natural phenomenon analysis and financial modeling. Self-similarity, measured by the Hurst exponent (H), reflects long-term dependence in signals. Wavelet-based methods are effective for estimating H due to their multi-scale nature, but additive noise in real-world data often reduces accuracy. We propose Noise-Controlled ALPHEE (NC-ALPHEE), an enhanced version of the Average Level-Pairwise Hurst Exponent Estimator that incorporates noise mitigation and generates multiple level-pairwise estimates from signal energy pairs. A neural network (NN) replaces traditional averaging to combine these estimates, preserving ALPHEE's behavior in noise-free cases while improving robustness under noise. Simulations show that NC-ALPHEE matches ALPHEE's accuracy for noise-free data. In noisy conditions, conventional averaging degrades and requires restrictive level selection, whereas NC-ALPHEE consistently outperforms existing methods without such constraints.
Keywords
Self-similarity
Hurst Exponent
Processing Noisy Signals
Wavelet Transform
Local governments are increasingly using machine learning to manage community resources, from prioritizing road repairs to mapping out public transit routes. However, these algorithms risk reinforcing or even worsening existing inequalities between neighborhoods of different backgrounds. This study introduces a practical auditing method designed to catch and measure potential bias in these public-sector systems by combining standard statistical fairness metrics with spatial analysis across census tracts.
To test this, I analyzed city datasets-specifically infrastructure service requests and transit operations-using gradient boosting and random forest models. I evaluated predictive accuracy and service outcomes across neighborhoods grouped by income and demographics. My findings show significant disparities; lower-resourced areas often face different treatment from models, even when their needs are similar to wealthier areas.
Finally, I explored how to bridge these gaps using post-hoc calibration strategies, demonstrating meaningful reductions in fairness disparities. This research offers a practical roadmap and provides evidence needed to build more equitable community services.
Keywords
algorithmic fairness
community equity
machine learning auditing
municipal data
bias detection
spatial statistics
Experimental design on networks is complicated by interference, where outcomes may depend on the treatment assignments of neighboring units. While existing methods account for network structure, they are typically evaluated on small or simplified networks, limiting their relevance for complex real-world settings. We propose a network-aware design framework for treatment allocation that integrates allocation balance with network topology through an optimality criterion based on the Fisher information matrix. An efficient local search algorithm enables scalable optimization over large combinatorial design spaces. We examine the causal properties of the resulting designs by evaluating total, direct, and indirect treatment effects under interference. Simulation studies across multiple random graph models, including Erdös-Rényi, geometric random graphs, preferential attachment models, and stochastic block models, demonstrate how network structure shapes optimal allocations. Applications to real-world networks, e.g., college housing and ego-Facebook networks, demonstrate topology-aligned treatment allocations.
Keywords
Complex networks
Experimental design
Interference
Heterogeneous causal treatment effect
We study offline imitation learning (IL) when part of the decision-relevant state is observed only through noisy measurements and the distribution may change between training and deployment. Such settings induce spurious state--action correlations, so standard behavioral cloning (BC)---whether conditioning on raw measurements or ignoring them---can converge to systematically biased policies under distribution shift. We propose a general framework for IL under measurement error, inspired by explicitly modeling the causal relationships among the variables, yielding a target that retains a causal interpretation and is robust to distribution shift. Building on ideas from proximal causal inference, we introduce CausIL, which treats noisy state observations as proxy variables, and we provide identification conditions under which the target policy is recoverable from demonstrations without rewards or interactive expert queries. We develop estimators for both discrete and continuous state spaces; for continuous settings, we use an adversarial procedure over RKHS function classes to learn the required parameters. We evaluate CausIL on semi-simulated longitudinal data from
Keywords
Causal Inference
Imitation Learning
Distribution Shift
Measurement Error
Admissions decisions in graduate programs involve considering academic excellence and professional readiness across large and heterogeneous applicant pools. While logistic regression is a standard analytical approach, it may fail to capture nonlinear relationships and interaction effects. This study applies a multi-method framework to examine MBA admissions for 6,000 applicants. The analysis combines logistic regression, Random Forests, Gradient Boosting Decision Trees with SHAP value plots, and Chi-square Automatic Interaction Detection segmentation. Logistic regression offers interpretable odds ratios and statistical inference, while tree-based models capture nonlinearities. SHAP values are used to explain individual predictions, and CHAID identifies distinct subgroups.
Across all methods, GMAT score emerges as the dominant predictor, with GPA playing secondary role. Demographic variables exhibit detectable effects. Notably, results are highly concordant across models. CHAID segmentation reveals segment structures. This work demonstrates how integrating classical models with explainable machine learning enhances transparency and is broadly applicable to selection processes.
Keywords
Logistic regression
GBDT
CHAID segmentation
Positional encoding (PE) is a core architectural component of Transformers, yet its impact on the Transformer's generalization and robustness remains unclear. In this work, we provide the first generalization analysis for a single-layer Transformer under in-context regression that explicitly accounts for a completely trainable PE module. Our result shows that PE systematically enlarges the generalization gap. Extending to the adversarial setting, we derive the adversarial Rademacher generalization bound. We find that the gap between models with and without PE is magnified under attack, demonstrating that PE amplifies the vulnerability of models. Our bounds are empirically validated by a simulation study. Together, this work establishes a new framework for understanding the clean and adversarial generalization in ICL with PE.
Keywords
Positional Encoding (PE)
In-Context Learning (ICL)
Rademacher Complexity
Generalization Gap
Adversarial Rademacher Complexity (ARC)
Incomplete time series data pose challenges for accurate analysis, especially when periodic structures like seasonal trends are present. Traditional imputation methods often fail to preserve these temporal dynamics, leading to biased estimates. This study introduces a structure-preserving imputation framework that integrates periodic components into the multiple imputation process using the Variable Bandpass Periodic Block Bootstrap (VBPBB). We simulate time series data with annual and monthly periodicities, varying noise levels (low, moderate, high), and missingness under Missing Completely at Random (MCAR) across missingness proportions (5%-70%). VBPBB extracts dominant periodic components, which are bootstrapped and incorporated as covariates in the Amelia II imputation model. Results show that VBPBB-enhanced imputation consistently outperforms standard methods, with the most significant gains observed in high-noise settings and when multiple periodic components are retained. This framework provides a flexible solution that preserves temporal structure, offering potential for improving imputation in temporally correlated data environments.
Keywords
Missing data; Time Series; Periodicity; periodic component; Variable Bandpass Periodic Block Bootstrap;Multiple Imputation; Amelia II; Bootstrap
MCAR: Missing Completely at Random; MAE: Mean Absolute Error; RMSE: Root Mean Square Error; VBPBB: Variable Bandpass Periodic Block Bootstrap; PC: Periodic Component;
kzft:Kolmogorov-Zurbenko Fourier Transform
In a world of big data, machine learning (ML) practice tends to rely on asymptotic assumptions to achieve low bias when estimating generalizable test error. However, for small-sample problems, ML is being used even though these assumptions no longer hold. This is problematic, as the common practice of using the combination of k-fold cross-validation and grid search to tune hyperparameters results in high bias between the validated and true generalization errors. We propose using optimal design of experiments principles to fit a response surface to a space-filling design that can be used to generate an optimal hyperparameter set. By performing Monte Carlo simulations on real datasets, we show that this approach generates hyperparameter sets with similar performance to grid search while also drastically decreasing the discrepancy between validation and generalization error. This decreased bias will aid practitioners in accurately assessing model performance without a significant reduction in predictive accuracy.
Keywords
Optimal Experimental Design
Small Sample Machine Learning
Hyperparameter Tuning
Clustering is widely used in biomedical research to identify heterogeneous patient subpopulations with diffuse boundaries. While fuzzy c-means (FCM) allows mixed memberships, statistical inference for fuzzy clustering remains limited.
We propose weighted fuzzy c-means (WFCM) framework for settings with cluster-size imbalance. Cluster-specific weights prevent small clusters from being dominated and induce a likelihood-based model. Estimation is performed via a majorize–minimize algorithm, enabling likelihood-ratio tests and bootstrap confidence intervals. We establish consistency and asymptotic normality of the estimator. Simulation studies demonstrate improved accuracy and uncertainty quantification under imbalance. The method is robust to tuning choices and scales well to moderate dimensions. It provides interpretable soft memberships that reflect continuous disease or cellular states. Applications to RNA-seq and ADNI data show stable uncertainty quantification and biologically meaningful soft memberships, ranging from imbalanced cell populations to a graded Alzheimer's disease progression.
Keywords
Clustering
Inference
FCM
weighted Fuzzy c-means
Learning predictive models in heterogeneous, resource-limited settings is challenging: local models lack statistical power, while standard transfer learning methods often fail to adapt to context-specific structure. We introduce Contextualized Transfer Learning (CTL), a framework that models context-dependent prediction through shared latent representations, enabling information sharing across related tasks while preserving individualized adaptation. We derive learning bounds for CTL under stability and learnability conditions, characterizing its generalization behavior. Empirically, CTL achieves predictive performance comparable to state-of-the-art black-box models while providing individual-level interpretability through its structured parameterization, making it a principled and interpretable approach for learning under heterogeneity.
Keywords
Learning Theory
Transfer Learning
Contextualized Learning
Bulk multi-omics clustering methods either fuse modalities into a single similarity graph and lose modality-specific signal, or decompose variation into joint and individual factors under linear and distributional assumptions that rarely hold across layers with different marginal characteristics. No existing framework pairs a flexible, distribution-free similarity with an explicit shared and specific decomposition that can be clustered independently. We introduce multiRF, which addresses this gap using directed multivariate random forests fitted between all pairs of omics blocks. For each directed connection, a multivariate random forest is trained with one omics block as predictor and another as response, producing a sample-level weight matrix whose rows encode cross-modal predictive neighbourhoods. Because splits are axis-aligned and operate feature by feature, the forests accommodate mixed data types, nonlinear inter-omics relationships, and the high-dimensional, low-sample-size setting common to molecular profiling, all without requiring a parametric functional form or matched feature scales across modalities. A response subsampling mechanism further regulates the multivariate split criterion when the response block is high-dimensional, ensuring that the weight matrix reflects genuine cross-modal structure rather than spurious high-dimensional correlations. The per-connection weight matrices are then fused into a single global weight matrix that captures shared structure across all modalities. This matrix partitions every omics block into a shared reconstruction driven by cross-modal consensus and a residual that isolates modality-specific variation, without explicit latent-factor estimation or rank assumptions. Shared and specific similarity matrices are constructed from these two components and clustered independently, so that cross-omics disease subtypes and modality-private biological axes both emerge from a single pipeline with full traceability to the underlying inter-omics predictive relationships. multiRF is available as an open-source R package at https://github.com/novawz/multiRF.
Keywords
Multivariate Random Forest
Multi-Omics Integration
High-Dimensional Data
The eruption time of Mount Vesuvius in 79 AD has been debated for decades. The traditional view is August 24th, while recent archaeological evidence indicates that the eruption date was in autumn. Gaussian Matrix Mixture Model (GMMM) and bootstrap trajectory simulation were implemented to solve this problem. Based on the analysis of 50 years of ERA5 reanalysis wind data (1950-1999) in the Vesuvius volcanic area, we fitted the model with the AR(1) structure to characterize daily wind patterns. Using bootstrap to generate two-day wind trajectory, we estimated the probability of volcanic ash deposition in the target area, which matched the known 79 AD volcanic ash distribution. Our results show a clear seasonal pattern, with higher deposition probability in summer. Comparison of the two competing hypotheses shows that the August hypothesis (August 17-24) yields a relative probability of 65.78%, almost twice that of the October hypothesis (October 17-31) at 34.22%. These findings provide quantitative support for the traditional August eruption date documented in Pliny the Younger's letters.
Keywords
Mount Vesuvius
Eruption timing
ERA5 reanalysis
Matrix-normal mixture model
EM algorithm
Bootstrap
Automated annotation of clinical narratives with Human Phenotype Ontology (HPO) terms enables downstream statistical analysis in precision medicine and rare disease research. Advances in large language models (LLMs) have expanded phenotype extraction capabilities, yet their comparative behavior under ontology-aware evaluation remains poorly characterized.
We present a systematic comparison of HPO annotation methods spanning rule-based systems, neural encoder models, and generative LLM-based approaches. Experiments use multiple publicly available and expert-annotated datasets varying in text length and annotation density.
Beyond exact-match metrics, we incorporate HPO hierarchy-aware evaluation and examine the effects of text segmentation, annotation filtering, and hierarchy-based tuning. Results indicate that performance is sensitive to both data characteristics and evaluation design, with methods exhibiting distinct trade-offs in coverage, specificity, and robustness. These findings highlight the importance of ontology-aware evaluation and motivate further methodological refinement.
Keywords
Human Phenotype Ontology
Clinical text analysis
Ontology-aware evaluation
Large language models
Biomedical NLP
Statistical comparison
Speaker
Yuyan Yi, National Institute of Allergy and Infectious Diseases
Co-Author(s)
Rachel Waymack, National Institute of Allergy and Infectious Diseases
Daniel Veltri, National Institute of Allergy and Infectious Diseases
Wildland fire behavior is determined by fuels and environmental variables. We analyze fire intensity data gathered during managed burns at the Fort Stewart-Hunter Army Airfield in Georgia in 2022. Because fire intensity changes continually throughout the burn, we treat intensity as a smooth functional response measured across time, capturing the full temporal trajectory of fire intensity. Response functions were expressed using a B-spline basis expansion to provide a flexible functional representation. The independent variables are scalar-valued vegetation characteristics measured at each site, e.g., live and dead biomass components including live biomass and litter weights; 1-hour and 10-hour fuels; pine needles; pine cones; conifer, fragmented, and cypress litter; and coarse and fine chars. These provided the fuel for prescribed fires. A functional regression approach is used to estimate the time-varying effects of fuel variables on fire intensity, with covariate effects changing over time. This approach detects early-stage impacts that scalar regression models may not capture and shows how fuel variables influence the temporal dynamics of fire intensity.
Keywords
Functional data analysis
Basis expansion
Prescribed fire behavior
Wildland fire intensity
Temporal fire dynamics
Functional response
Deep kriging combines classical kriging ideas with neural network–based representations, providing a flexible framework for spatial prediction beyond parametric covariance models. For global spatial data defined on the sphere, recently developed intrinsic random functions offer a principled approach to modeling non-stationarity and large-scale trends within spherical geometry.
In this work, we investigate deep kriging models for spatial processes on the sphere. Using a simulation framework based on spherical harmonic representations, we compare Euclidean deep kriging with geometry-aware formulations that incorporate great-circle structure and intrinsic covariance constructions. Our study highlights how classical intrinsic kriging ideas on the sphere can be integrated with deep learning–based spatial prediction, offering stable and interpretable modeling of global spatial processes.
Keywords
kriging on the sphere
intrinsic random functions
deep neural networks
spatial statistics
nonstationary spatial processes
Liquid association measures how the association between two variables varies with a third variable, yet its classical formulation collapses this modulation into a single scalar, potentially obscuring heterogeneous, non-monotonic, or localized effects. We introduce Functional Liquid Association (FLA), which extends liquid association to a functional representation that characterizes how association evolves across the full range of the moderating variable. Within a unified projection-based framework, FLA supports both a local velocity-centric analysis that highlights rapid changes or threshold effects in association strength, and a global mode-centric analysis that summarizes dominant regimes. To enable inference without restrictive distributional assumptions, we develop permutation-based tests for detecting overall association modulation and identifying localized regions of significant change. Simulations and applications to gene expression analyses demonstrate that FLA reveals structured, condition-specific association patterns missed by classical liquid association. These results establish FLA as a practical and interpretable framework for studying context-dependent associations
Keywords
Context-dependent association
Liquid association
Nonparametric inference
Permutation testing
Food insecurity is a major public health concern in the United States, with profound implications for physical health as well as cognitive and social development in children. It is known that food insecurity is caused by a combination of availability, affordability, access and utilization. While measures of cost of a balanced and affordable meals and physical access to retailers exist, they often do not account for the joint impact. For example, a store within access may be unaffordable and vice versa. This study aims to evaluate the local food environment by estimating the cost of a balanced meal in a given store type in Massachusetts counties. We analyze products and prices reported by a sample of retailers, mapping retail food items to nutritional value and cost, and calculate cost of a balanced meal according to USDA My Plate dietary guideline. In the future, we plan to combine transportation analysis to retailers and build a framework to assess Massachusetts local food environment. This would help inform food assistance programs and community nutrition initiatives in Massachusetts.
Keywords
Food access equity, food insecurity
USDA My Plate
Healthy diet affordability
Local food environment
Retail food analysis
Massachusetts, community nutrition initiatives
Differences in health outcomes such as uncontrolled hypertension persist between social groups (class, race, etc.). Causal decomposition analysis models a hypothetical intervention while classifying confounders as just or unjust strata for differential outcomes or treatment. However, causal decomposition assumes that all variables are measured, which may not be true in observational data. Here, we generalize a causal decomposition from a study population (where all variables are measured) to a target population (with unmeasured variables). We define and identify a causal estimand and propose g-computation and weighting estimators. We apply our method to a dataset of over 30 local clinics in Maryland and Pennsylvania, some of which participated in the RICH LIFE study to reduce racial disparities in hypertension control. Our contributions are to generalize a non-randomized, stochastic intervention (a realistic yet under-explored setting) and to consider unmeasured variables and confounding in a health equity context.
Keywords
Causality
Generalizability
Stochastic Intervention
Allowability
Disparity
A Sequential, Multiple Assignment, Randomized Trial (SMART) is a clinical trial design for developing and comparing dynamic treatment regimens. Standard SMART analyses often use weighted and replicated regression. In non-SMART studies with count outcomes, especially in settings of substance use, zero-inflation is a common issue.
One solution for zero-inflated longitudinal outcomes is the two-part hurdle model (HM). This model handles random zero cases where subjects have zero outcomes but remain at risk. HMs have not yet been applied to SMART data.
We propose a two-part HM for zero-inflated, longitudinal count outcomes in SMARTs. The model combines logistic regression for zero/nonzero outcomes with a truncated Poisson model for nonzero counts. Our approach is motivated by and applied to the SafERteens M-Coach SMART, which studied investigated the effects of brief interventions and text messaging on young adults on alcohol consumption outcomes.
Keywords
Sequential Multiple Assignment Randomized Trial
Zero-Inflated Outcomes
Clinical Trial Methods
Poisson Regression
Longitudinal Data
Speaker
Hanna Venera, University of Michigan
Co-Author(s)
Kelley Kidwell, Department of Biostatistics, School of Public Health, University of Michigan
Bingkai Wang, Department of Biostatistics, School of Public Health, University of Michigan
Maureen Walton, University of Michigan
Recent research indicates that difficulty reducing alcohol consumption may serve as a risk marker for suicidal ideation, but whether this relationship differs based on alcohol use disorder (AUD) status remains unclear. Using data from the 2021-2023 National Survey on Drug Use and Health (NSDUH), we examined whether meeting the criteria for AUD moderated the association between past-year inability to reduce drinking and past-year suicidal ideation. The sample included 22,872 adults aged 18+ with past-year alcohol use. Logistic regression models tested the interaction between inability to reduce in the past year and AUD status, adjusting for age, sex, and race/ethnicity. AUD status significantly moderated this association (interaction OR=2.77, 95% CI[1.11,6.93], p=.03). Among individuals with AUD, inability to reduce drinking was associated with elevated suicidal ideation (OR=1.59, p<.001), whereas no significant association was observed among those without AUD (OR=0.62, p=.298). Results demonstrate that inability to reduce drinking may serve as a suicide risk marker specifically among individuals with AUD, with potential implications for targeted screening and intervention efforts.
Keywords
Survey-weighted Logistic Regression
Suicidal Ideation
Alcohol Use Disorder
Behavior Change
Speaker
Colin Jarratt, Massachusetts General Hospital/Harvard Medical School
Co-Author
Rebecca Fortgang, Massachusetts General Hospital/Harvard Medical School
Heart disease remains a leading cause of mortality in the United States with substantial implications for public health and healthcare expenditures. This study applied logistic regression to assess associations between various predictors, including chronic conditions, lifestyle habits, social demographics, and healthcare access, and the odds of self-reported coronary disease or heart attack. The data were a subset of the 2015 Behavioral Risk Factor Surveillance System (BRFSS) with 22 features from 253,680 individuals. Variable selection was performed using forward selection and backward elimination based on AIC and BIC criteria, and model performance was evaluated using standard predictive accuracy measures. Key predictors included history of cholesterol screening, prior stroke, inability to afford medical care, sex, and income level. These results may inform healthcare providers and payors about risk stratification and the promotion of preventive care. Notably, the positive association between cost-related barriers to healthcare access and heart disease underscores the potential value of policies aimed at reducing financial obstacles to care in mitigating cardiovascular risk.
Keywords
logistic regression
odds ratios
coronary heart disease
risk factors
Behavioral Risk Factors Surveillance System (BRFSS)
healthcare access
Causal mediation analysis, based on the counterfactual approach, decomposes the total effect of an exposure into natural direct and indirect effects, with the mediation proportion (MP) quantifying the relative contribution of the mediator. However, the MP, along with its components, the natural indirect effect (NIE) and the natural direct effect (NDE), are functional estimators. In finite samples, their ratio and exponential forms can make them unstable and biased. This paper introduces the problem of transformation-induced bias in regression-based causal mediation analysis and proposes two likelihood-based bias-correction methods. These methods target exponential and ratio functionals, including the NDE, NIE, and MP, and provide closed-form corrections for continuous and binary outcomes and mediators, with or without exposure–mediator interaction. Simulation studies show that ordinary estimators exhibit notable finite-sample bias, especially for the MP and log-odds scale effects. In contrast, the proposed methods reduce relative bias and MSE while preserving large-sample properties. Real-data applications confirm more stable and precise inference, particularly for the MP.
Keywords
causal mediation analysis
likelihood thoery
mediation proportion
transformation-induced bias
bias-correction
Mediation analysis is a valuable method for investigating mechanisms through which treatments or exposures influence an outcome, with broad applications in biomedical and public health research. Methodological developments include, for example, regression-based approaches and weighting to enable causal interpretation. Although methodological advances have been substantial, the use of matching within causal mediation analysis remains largely unexplored. We introduce a matching-based causal mediation method for settings with a binary mediator and binary outcome. Within the counterfactual framework, our approach estimates natural direct and indirect effects without reliance on strong parametric assumptions. We apply the proposed method to the Population Assessment of Tobacco and Health (PATH) Study to evaluate whether a binary mediator explains the association between e-cigarette use and the outcome of long-term smoking cessation. We assess the performance of our approach through simulation studies under scenarios with common and rare outcomes and in the presence of an exposure-mediator interaction.
Keywords
Causal inference
Mediation analysis
Matching
Smoking cessation
E-cigarettes
Lyme disease (LD), transmitted by Ixodes ticks infected with Borrelia burgdorferi, is the most commonly reported vector-borne disease in the United States and continues to expand geographically. Transmission-blocking, reservoir-targeted vaccines (TBRTVs) have been proposed to reduce nymphal infection prevalence and disrupt transmission. Evaluating such interventions requires understanding infection and recovery dynamics over time. Yet, longitudinal infection data are often sparse, irregularly sampled and seasonal. Multi-year longitudinal data on canine Borrellia infection were collected from a cohort of hunting dogs, providing a view into the potential impact of coinfection. In particular, we investigate the relationship between Leishmania infection and Lyme Disease. We propose a Bayesian hierarchical hidden Markov model to characterize infection and recovery processes, allowing transition probabilities to vary across seasons and sites, and to respond to coinfection. Overall, this framework provides a flexible approach for disentangling seasonal and coinfection effects in sparse longitudinal disease data, with important implications for interventions targeting disease burden.
Keywords
Vaccine
Bayesian hierarchical model
Lyme disease
Leishmania Infection
Borrelia burgdorferi
High-dimensional mediation analyses of molecular data often ignore existing biological knowledge, despite the availability of databases describing molecular interactions, pathway memberships, and other functional relationships. We propose PALM (Prior-Aware Learning for Mediation), a graph learning framework that combines biological priors with data-driven empirical correlations to improve mediator effect estimation when any single source of prior information is incomplete or noisy. PALM fits a joint mediation model including all molecular candidates and uses a graph-based penalty to encourage related molecules to have similar effects. Rather than relying on one predefined molecular network, PALM learns a combined relationship graph from multiple complementary prior sources and incorporates this learned structure into the mediation model. By borrowing strength across connected mediators, PALM improves the stability and power of estimating indirect effects. Simulation studies show improved identification of true mediating signals across a range of heterogeneous prior-information settings compared with existing regularization methods. We apply PALM to metabolomics data from the Framingham Heart Study to illustrate its utility.
Keywords
Mediation analysis
High-dimensional data
Graph learning
Prior knowledge integration
Speaker
Yixin Zhang, Boston University School of Public Health
Co-Author
Ching-Ti Liu, Boston University School of Public Health
MELD 3.0 is a scoring system to prioritize patients for liver transplant (LT). We compared three methods of determining survival benefit (SB) of LT and effect modification of SB by MELD using time-dependent interaction terms: 1) Multivariable (MV) Cox model with time-dependent variables for LT and MELD - standard analysis, 2) MV sequential stratification (SS) model – SS matches time from listing to LT of LT recipients to waitlist candidates who have survived at least the same length of time, then uses stratified MV Cox model with robust standard errors, 3) Cox model using time-varying propensity score (tvPS)-matched data – uses hazard component from Cox model as tvPS, matches a treated patient to a not-yet-treated patient who has similar time-dependent covariates up to the moment when treatment occurs, and then uses stratified Cox model on matched data. Unlike the standard analysis, SS and tvPS ensure treatment groups are comparable at each time point when treatment eligibility changes. Overall survival benefit was the same across methods, but effect modification of SB by MELD varied. Further work is needed to investigate discrepancies and understand why there are differences.
Keywords
time-varying propensity score matching
sequential stratification
survival benefit
Cox with time-dependent exposure
Speaker
Amy Shui, University of California San Francisco
Co-Author
Catherine Lee, University of California San Francisco
E-cigarette use among middle and high school students and college students steadily increased in the last decade, raising significant public health concerns. It is argued that e-cigarettes contain a lower level of toxicants than combustible tobacco cigarettes. This perception has contributed to the growing popularity of e-cigarettes among youth. However, lower level of toxicants does not mean addiction is less likely. In this study, we examine trends in the usage of electronic cigarettes among middle and high school students in the U.S. from 2018 to 2023 and college students from Kent State University, compare these patterns with traditional cigarette use over the same period and population, and explore the underlying reasons driving e-cigarette use within this demographic. Understanding patterns and motivations are essential for developing effective prevention and intervention strategies. Our finding contradicts claims by e-cigarette manufacturers that their products facilitate smoking cessation. Further rising e-cigarette smoking exclusively is of public health concern, which warrants targeted intervention.
Keywords
National Youth Tobacco Survey
adolescent behavioral health
college students
e-cigarettes
nicotine use
tobacco trends
Real-world data (RWD) analysis often suffers from unobserved confounding. We integrated RWD with drug-related gene expression profiles to identify drug targets for Alzheimer's disease (AD) without generating new genetic data, while aiming to mitigate these confounding effects. We developed a meta-difference-in-differences (meta-DiD) framework to compare pooled effect sizes across cohort collections. We derived the mathematical properties of unobserved confounding in Cox models and meta-DiD frameworks, applying this to identify neuroinflammation-specific AD drug targets. A Shiny app was developed to calculate confounding effects in Cox models. Simulations demonstrated that: (i) model-based and empirical confounding effects are consistent; (ii) meta-DiD improves confounding control as cohort numbers increase or when confounding is large and homogeneous; and (iii) identified drug targets were consistently associated with AD across both RWD and differential gene expression analyses. Meta-DiD effectively controls for large, homogeneous confounding across cohorts. Integrating RWD with gene expression profiles provides a robust pipeline for drug target discovery.
Keywords
Real World Data
Causal Inference
Meta-Difference-in-Difference
Unobserved Confounding
Alzheimer's Disease
Drug Target Identification
We assessed the impact of retention in care on virological failure in people with HIV (PWH), ≥18 years old, and enrolled in North American AIDS Cohort Collaboration on Research and Design (NA-ACCORD). Previous work looked at time-fixed retention definitions (retained in both of the first 2 years). We enhanced this by implementing a nested trial approach with time-updated and lagged retention (attending ≥2 visits >90 days apart in the prior year) and multiple baselines per person. Additional time-updated and lagged covariates (diabetes diagnosis, hypertension diagnosis, AIDS diagnosis, age, ART medication receipt, and year in study) and time-fixed variables (HIV acquisition risk, NA-ACCORD location, sex, and race/ethnicity) were included. Using a Cox model with treatment and censoring weights, time-updated retention had a 25% decrease in hazards of virological failure (HR: 0.70-0.79) compared to the time-fixed result of a 35% decrease (HR: 0.62-0.68). The emulated trial method allows us to include time-updated data, leading to a more accurate depiction of the covariates' associations with the time-dependent outcome. This method enhanced our understanding of the HIV care continuum.
Keywords
emulated trial
nested trial
HIV
time-to-event
time varying covariates
weighted Cox regression
Speaker
Elisa Yazdani, Vanderbilt University Medical Center
Co-Author(s)
Kelly Gebo, George Washington University
Brenna Hogan, Johns Hopkins University
Alicia Rector, Vanderbilt University Medical Center
Amber Hackstadt, Vanderbilt University Medical Center
Peter Rebeiro, Vanderbilt University Medical Center
Michael Horberg, Kaiser Permanente Mid-Atlantic States
Jessica Edwards, University of North Carolina at Chapel Hill
Sonia Napravnik, University of North Carolina at Chapel Hill
John Gill, Alberta Health Services
Keri Althoff
Catherine Lesko, Johns Hopkins University
The detection of planets outside of our solar System, called exoplanets, is crucial for understanding the universe and learning about Earth's formation. Like Earth, exoplanets orbit a star, and require careful measurements to detect them. One measurement astronomers use for detection is the estimate of a star's velocity with respect to an observer, called its radial velocity (RV). RVs are naturally measured over time and across different wavelengths of light and can be analyzed as a panel using the line-by-line RV method. This method is used to both detect exoplanets and measure their mass but is hampered by outliers and missing observations. Indeed, anomalous behavior in data streams often indicates astrophysical or instrumental deviations unrelated to an exoplanet. To aid in predicting these anomalies, we propose the incomplete autoregressive conditional Fréchet model with time-varying parameters for regional extremes (I-ACRE). As model predictions follow a Fréchet distribution with a closed-form expression, this allows for confidence intervals, whose expected coverage levels (e.g., 90%) closely align with observed coverage levels in out-of-sample predictive tests. Unlike other models used to study extreme values, I-ACRE flexibly and quickly accommodates panel data with missing values by feeding forward predictions. By quantifying the probability of future extremes, I-ACRE provides a diagnostic framework for flagging anomalous observations that can limit the success of exoplanet detection algorithms.
Keywords
extreme value theory
tail index dynamics
exoplanet detection
panel data
missing data
Compositional data, where each component is a proportion of a whole, presents unique statistical challenges, particularly when data are incomplete. An acceptable missing data method for compositional data must maintain its characteristics, such as the inverse relationships between components and the constraint on each observation's sum. Multiple Imputation (MI) has become a standard method for imputing incomplete quantitative, ordinal, or categorical data, but there are not any proposed imputation methods for incomplete compositional data that are able to preserve the characteristics of the compositions. We propose methods for imputing compositional data by adapting MI, and use the imputed datasets to conduct analysis on exercise motivation survey data. The novel method will be used to impute missingness in the original dataset. The analysis results will be used to evaluate the performance of our proposal against standard methods.
Keywords
Applied Bayesian Statistics
Exercise Motivation
Missing Data
Multiple Imputation
Multivariate Statistics
Survey Methodology
Small-area estimates in official statistics must often be not only accurate but also coherent with known aggregate totals and area-specific logical bounds. Existing methods address benchmarking, inequality constraints, heteroscedastic variance modeling, and spatial dependence, but typically treat these separately rather than within one framework. We develop constrained heteroscedastic area-level models that simultaneously enforce benchmarking and lower-bound constraints, jointly model area-level means and sampling variances, and incorporate spatial dependence. Building on the conjugate framework of Parker et al. (2024), the models retain computational efficiency while allowing richer dependence and coherence than standard approaches. We consider a non-spatial model (CHALM) and a spatial extension (SCHALM) with ICAR structure on the mean and variance. In a simulation study using county-level Iowa corn data, both models substantially improve on direct estimates in RMSE and interval score, with the spatial model best overall. Applied to corn production across eight Midwestern states, both preserve the main geographic pattern while producing smoother, more stable estimates—SCHALM providing the strongest smoothing and the most coherent uncertainty surface.
Keywords
Small Area Estimation
Benchmarking
Inequality Constraints
Heteroscedastic area-level model
Bayesian Hierarchical Models
Spatial dependence
Technology tools have become increasingly influential in the sport of rowing; however, limited understanding remains regarding how technologies relate to team performance and how access varies across regions. This study examines geographic disparities in rowing access and the relationship between technology adoption and performance outcomes. Data from ~1,400 rowing clubs from USRowing.com were combined with state-level economic data and Google search trends measuring public interest. Coaches were surveyed regarding their teams' socioeconomic data, technology use, and performance, measured by race times. Regression analyses indicate that states with higher GDPs contain more clubs, and states with more clubs show greater public interest in rowing. Survey-based regression models suggest that coaches lack a clear understanding of the performance associations of individual rowing technologies and find that different tools can positively or negatively affect performance. Overall, findings suggest expanding infrastructure to enhance access and interest, improving coaches' understanding of each tool, and balancing resource allocation between accessibility and targeted technology use.
Keywords
Rowing
Sport technology
Geographical accessibility
Performance outcomes
This paper introduces the Depth-Separable Neural Network (DSNN), a novel neural-network-based framework for forecasting mixed-frequency data. Through a multi-depth architecture built from deep ReLU networks, the DSNN simultaneously performs adaptive frequency alignment and models complex nonlinear dependencies. Moreover, a parameter-sharing mechanism is adopted across the alignment networks, making the architecture separable by depth and computationally scalable even with a large set of higher-frequency predictors. We establish an approximation result for the DSNN over a broad class of hierarchical composition functions, and derive a non-asymptotic prediction error bound for its least squares estimator. Simulation studies demonstrate the finite-sample performance of the proposed method, and an empirical application to forecasting U.S. quarterly macroeconomic variables using monthly and daily indicators, highlights its superior predictive accuracy over existing mixed-frequency methods. The DSNN thus provides a theoretically grounded, scalable, and effective tool for complex mixed-frequency data analysis.
Keywords
frequency alignment
least squares estimation
mixed-frequency data
non-asymptotic properties
ReLU neural network
We assess whether large language model (LLM) assistants can build credible forecasting and risk‑flagging pipelines from multi‑year monthly insurance data without external sources or disclosure of proprietary levels. Using product–state–club–policy‑characteristic segments over ~3 years, we compare multiple LLM workflows under controlled conditions, strict time‑based splits, and leakage controls, with a naïve seasonal reference. Primary outcomes are 1‑month forecasts of loss ratio; 3‑month is secondary and 12‑month exploratory. Evaluation uses rolling‑origin cross‑validation and scale‑free metrics (MASE primary; sMAPE, 80/95% coverage secondary). Adverse selection is defined prospectively: segments whose next‑quarter loss ratio and PIF each increase ≥10% versus trailing‑12 baselines. We score probabilistic flags with a variety of statistical tests. Model differences are tested via paired comparisons. Results quantify accuracy, calibration, and reproducibility of LLM‑assisted analytics while preserving confidentiality through scale‑free reporting only.
Keywords
Large Language Models
Insurance
Forecasting
LLM