ABSTRACT Accurate estimation of incubation period is fundamental to surveillance for infectious diseases. This task faces two major challenges: (1) surveillance data are often coarsely observed, comprising interval- and right-censored observations, and (2) patient-level data are frequently fragmented across multiple facilities and cannot be pooled due to privacy constraints. While traditional meta-analysis (MA) aggregates summary statistics, its one-shot nature limits the ability to fit complex hierarchical models. We propose a novel federated Bayesian framework, Federated Stochastic Variational Inference (Fed-SVI), which extends the summary-statistic philosophy of MA into an iterative, model-based federated approach. Our method fits a hierarchical lognormal model, explicitly modeling heterogeneity while estimating a global distribution. To fit this model privately, Fed-SVI iteratively exchanges summary statistics (local variational parameters). Each facility analytically computes these local parameters, rigorously incorporating the likelihood contributions of coarsely-observed data via expectations of the truncated normal distribution. A central server aggregates these parameters using stochastic natural-gradient ascent, enabling efficient information borrowing and mini-batch updates. Through extensive simulations, we demonstrate that Fed-SVI achieves accuracy nearly identical to a fully centralized analysis, while substantially outperforming naive federated or MA methods that discard censored information. We apply Fed-SVI to a distributed COVID-19 line list from 177 facilities, yielding results consistent with centralized estimates. Fed-SVI provides a scalable, accurate, and privacy-preserving solution for robust surveillance in modern data networks.
ABSTRACT In meta-analyses of binary outcomes, two-stage methods based on the DerSimonian–Laird-type random-effects model have been widely used, but these methods fix the within-study variances to their asymptotic estimates and rely on normality assumptions of the outcome measures. Such strict and often unrealistic assumptions can yield biased or inefficient inference. Likelihood-based approaches such as the binomial–normal generalized linear mixed model and the hypergeometric-normal (HGN) model alleviate some of these issues but are intrinsically restricted to the odds-ratio scale, limiting interpretability and precluding direct estimation of the risk-ratio—a more clinically meaningful measure. To overcome these limitations, we propose a unified likelihood-based framework that enables direct estimation of the risk-ratio under the HGN formulation. The proposed method introduces a pseudo–data augmentation approach that preserves the conditional likelihood structure of the HGN model while yielding an unbiased estimating function for the log-risk-ratio. Higher-order inference is achieved via jackknife bias and variance corrections, ensuring accurate coverage even in small or sparse datasets. Simulation studies confirm nearly unbiased estimation and well-calibrated confidence intervals across a wide range of conditions. Applications to published meta-analyses demonstrate that this approach yields interpretable, coherent, and computationally tractable inference.
ABSTRACT This study proposes a nonparametric decomposition framework to quantify and prioritize the impact of multiple intervenable mediators contributing to disparities in health outcomes. Identifying modifiable components of health disparities is essential for informing policy and guiding resource allocation, but several methodological challenges remain. These include formulating hypothetical interventions for immutable characteristics such as race or ethnicity, mitigating bias from potentially incompatible statistical models used to represent causal pathways, and evaluating multiple mediators for their joint and individual effects. To address these challenges, we integrate recent advances in health disparities research with modern machine learning techniques. A key assumption, common to causal disparity frameworks, is the absence of unmeasured confounding in the mediator-outcome relationship. We extend existing methods in mediation analysis and introduce a nonparametric sensitivity analysis that provides sharp bounds on modifiable disparity estimates in the presence of unmeasured confounding. Together, these methods offer a robust and flexible approach to causal disparity analysis, enabling the identification of high-impact, actionable interventions. We demonstrate the utility of the methods in the context of colorectal cancer survival disparities, a setting where persistent differences across racial, ethnic, and underserved populations in the United States underscore the urgent need for analytic strategies that uncover effective intervention points.
ABSTRACT Recent neuroimaging studies have revealed that brain functional networks exhibit temporal fluctuations during cognitive tasks, sparking growing interest in dynamic functional connectivity (FC). However, few existing methods effectively incorporate temporal dynamics when examining the relationship between brain connectivity and cognition. In this work, we propose a Bayesian deep learning framework that decomposes FC into dynamic and stationary components, thereby disentangling transient fluctuations from stable network structures. By leveraging deep kernel learning, the model flexibly represents transient functions as continuous processes, allowing the data to guide the discovery of meaningful temporal patterns beyond conventional assumptions. Building on this representation, the dynamic component is modeled via a thresholded Gaussian process, which captures smooth temporal variations while inducing sparsity to highlight significant temporal signals. Simultaneously, the stationary component is modeled using a deep neural network with an integrated thresholding layer, enabling the learning of complex nonlinear relationships and enhancing interpretability by emphasizing key connections. This integrative approach balances model flexibility, temporal resolution, and interpretability, providing a powerful tool for predicting cognitive outcomes from functional neuroimaging data. We demonstrate that our method outperforms existing approaches in both prediction accuracy and feature selection through extensive simulations. Applied to data from the Adolescent Brain Cognitive Development study, our framework uncovers dynamic brain signals during cognitive tasks that significantly contribute to variations in cognitive performance.
ABSTRACT We propose a deep generative approach to nonparametric estimation of conditional survival and hazard functions with censored data. The key idea of the proposed method is to first learn a conditional generator for the joint conditional distribution of the observed time and censoring indicator given covariates, and then construct the Kaplan–Meier and Nelson–Aalen estimators based on this conditional generator for conditional hazard and survival functions. Our method combines ideas from the recently developed deep generative learning and classical nonparametric estimation in survival analysis. We establish the convergence properties of the generative nonparametric estimators. Our numerical studies validate the proposed method.
