Approximate Bayesian inference for joint partially linear modeling of longitudinal measurements and spatial time-to-event data

T. Baghfalaki, M. Ganjali, R. Martins · 2025

Read Approximate Bayesian inference for joint partially linear modeling of longitudinal measurements and spatial time-to-event data by T. Baghfalaki, M. Ganjali, R. Martins (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-05880.txt.

Introduction

Taban Baghfalaki, Mojtaba Ganjali, and Rui Martins present a statistical model that follows two health outcomes at once in this 2025 paper. Many studies track a repeated measure, such as a blood marker over time, alongside the time until an event such as death. When patients also cluster by place, standard joint models can miss the spatial pattern.

Their model links the two outcomes in one Bayesian framework. A spatial term lets nearby regions share information, while a flexible curve captures how the repeated measure bends over time. Fast approximate methods keep the computing practical. Simulations showed the method recovers the truth well. The team then applied it to HIV and AIDS data from Brazilian states. The work gives health researchers a tool for map-aware studies where both a marker's path and survival matter.

Catalog record

CREATORT. Baghfalaki, M. Ganjali, R. Martins
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-05880

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Approximate Bayesian inference for joint partially linear modeling of longitudinal measurements and spatial time-to-event data T. Baghfalaki, M. Ganjali, R. Martins Abstract The integration of longitudinal measurements and survival time in statistical modeling offers a powerful framework for capturing the interplay between these two essential outcomes, particularly when they exhibit associations. However, in scenarios where spatial dependencies among entities are present due to geographic regions, traditional approaches may fall short. In response, this paper introduces a novel approximate Bayesian hierarchical model tailored for jointly analyzing longitudinal and spatial survival outcomes. The model leverages a conditional autoregressive structure to incorporate spatial effects, while simultaneously employing a joint partially linear model to capture the nonlinear influence of time on longitudinal responses. Through extensive simulation studies, the efficacy of the proposed method is rigorously evaluated. Furthermore, its practical utility is demonstrated through an application to real-world HIV/AIDS data sourced from various Brazilian states, showcasing its adaptability and relevance in epidemiological research. Keywords: Approximate Bayesian inference; Integrated Laplace approximation (INLA); Joint modeling; Latent Gaussian model; Spatial effects; Spline functions. 1 Introduction In numerous clinical longitudinal studies, data on longitudinal markers are gathered until an event occurs [ 44 , 18 ] . For instance, in HIV studies, CD4 counts are typically monitored until a patient’s demise or dropout from the study, highlighting the interplay between longitudinal and survival processes. However, these processes are often analyzed separately, leading to biased parameter estimates [ 26 ] . The solution lies in joint modeling, a burgeoning field with extensive medical applications [ 12 , 56 , 35 , 13 , 22 , 57 ] . Joint modeling typically involves two submodels: the longitudinal submodel, often a linear mixed effects model [ 62 ] , and the survival submodel, ranging from the traditional Cox model to accelerated failure time models [ 33 ] . Central to joint modeling is the establishment of a linking structure, often achieved by sharing random effects between the longitudinal and survival processes [ 44 , 18 ] . Recent years have seen significant advancements in research within this domain, extending beyond the conventional framework to analyze diverse data scales such as counts or ordinal longitudinal outcomes [ 34 , 2 , 5 , 64 , 3 ] , as well as multivariate longitudinal outcomes [ 14 , 4 , 36 , 29 , 32 ] . Additional extensions include addressing competing risks, multiple causes of death [ 60 , 34 , 30 , 55 , 5 ] , and cure fraction models [ 11 , 63 ] . For a comprehensive review of joint modeling, readers may refer to [ 53 , 43 , 20 , 1 ] . A promising direction for further research is the incorporation of spatial effects based on geographical regions, an area that has been relatively underexplored [ 37 , 38 , 40 ] . Although joint modeling is highly effective, challenges remain in parameter estimation, which frequently involve complex and time-consuming computations. In the realm of joint modeling, various strategies for parameter estimation are employed. Classical approaches, such as the EM algorithm [ 31 , 21 , 27 , 39 , 41 ] , have been widely utilized alongside Bayesian methods [ 1 ] . However, these methods often involve intricate and time-consuming computations, prompting the exploration of alternative strategies. To address this, researchers have explored approximate Bayesian methods as an efficient alternative to traditional approaches [ 58 , 3 , 42 , 51 ] . Approximate Bayesian methods are often employed when exact Bayesian inference becomes computationally prohibitive due to the complexity or dimensionality of the model. Several approaches exist to address these challenges. Markov Chain Monte Carlo (MCMC) methods are a widely used class of algorithms that generate samples from the posterior distribution. However, while flexible, they can be computationally intensive, especially in high-dimensional settings [ 25 ] . Approximate Bayesian Computation (ABC) is another approach designed for models with intractable likelihoods, where inference is performed by comparing observed data to data simulated under the model [ 6 , 52 ] . ABC is often applied in fields such as genetics and ecology, where exact likelihood functions are unavailable. Alternatively, methods such as Variational Inference (VI) and Integrated Nested Laplace Approximation ( INLA ) offer more computationally efficient solutions. VI approximates the posterior by finding a simpler distribution that minimizes the divergence from the true posterior, often at the cost of some accuracy [ 9 ] . INLA , on the other hand, is particularly effective for latent Gaussian models, using Laplace approximations to provide fast and accurate approximations of marginal posteriors [ 47 , 48 ] . In this paper, we propose an approximate Bayesian method for joint modeling of longitudinal and spatial survival data. Our approach builds upon existing models while introducing innovations to enhance computational efficiency and model flexibility. Performance evaluation through simulation studies and application to real-world HIV/AIDS data from Brazil underscores the utility of our method. For implementation, we leverage the INLA package, which allows for efficient Bayesian inference using Integrated Nested Laplace Approximations. The INLA framework is particularly well-suited for high-dimensional latent models like those involving spatial and longitudinal data. Additionally, we provide accompanying R code to ensure reproducibility and facilitate the application of our method in other research settings. By combining the strengths of INLA with our methodological enhancements, our approach offers a robust and computationally feasible solution to the challenges of joint modeling in complex data scenarios. We employ an approximate Bayesian approach to implement joint modeling of longitudinal and spatial survival data. While the core model resembles that of [ 38 ] , we introduce two key innovations: (1) leveraging approximate Bayesian inference instead of the traditional MCMC method, resulting in reduced computational time compared to Gibbs sampling, and (2) incorporating a partial linear joint model to account for nonlinear effects of time, thereby enhancing model versatility. To evaluate the performance of our proposed model, we conduct simulation studies and apply it to real-world Brazilian HIV/AIDS data. We utilize the INLA package for implementation, and the accompanying R code is publicly available at https://github.com/tbaghfalaki/ASJM . The structure of this paper is organized as follows: Section 2 details the specification of the joint modeling framework, including longitudinal submodels, spatial survival submodel, and the approximate Bayesian approach for model implementation. Section 3 focuses on the analysis of real data, while Section 4 presents the results of simulation studies. Finally, Section 5 provides concluding remarks and discusses avenues for future research. 2 Joint modeling specification 2.1 Notation The spatial joint model has two separate submodels including a linear mixed effects model and a spatial survival model. The association of these two separate models are taken into account by a shared random effect. Let N individuals be collected from K regions such that the number of individuals in each region is n_{k} and \sum_{k=1}^{K}n_{k}=N . Let T_{ik}^{*} be the true value of survival time for the i^{th} individual in the k^{th} region, such that the observed survival time T_{ik} is the minimum value of T_{ik}^{*} and the censoring time C_{ik} , that is, T_{ik}=min(T_{ik}^{*},C_{ik}) . Consider an indicator \delta_{ik} such that \delta_{ik}=0 indicates a right censored observation. Thus, the survival

Read the complete text