pith. sign in

arxiv: 2604.18057 · v1 · submitted 2026-04-20 · 📊 stat.ME · stat.AP· stat.CO

Efficient Bayesian inference for non-linear association structures in joint models: A hierarchical approach via INLA

Denis Rustand , H{\aa}vard Rue , Lisa Le Gall , Karen Leffondre This is my paper

Pith reviewed 2026-05-10 04:12 UTC · model grok-4.3

classification 📊 stat.ME stat.APstat.CO
keywords joint modelsnon-linear associationINLABayesian inferencelongitudinal and survival dataBMImortality riskrandom walk prior
0
0 comments X

The pith

A hierarchical INLA framework lets joint models capture non-linear biomarker effects on survival risks by separating baseline from smooth deviations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that joint models linking longitudinal biomarkers to event times can move beyond the usual linear assumption to handle realistic non-linear patterns such as U-shapes or plateaus. It does so by building a hierarchy inside the INLA framework: the marker's effect on the log-hazard is split into a simple parametric baseline plus a flexible smooth deviation expressed through an orthogonal basis taken from a second-order random walk. This structure lets users add or remove flexibility with standard model-choice tools and runs quickly enough for routine use. Simulations confirm the approach recovers known complex trajectories, while the real-data example on BMI and mortality uncovers a U-shaped risk curve for current BMI value and a non-linear link with the rate of weight change.

Core claim

The scaling effect of the longitudinal marker on the event hazard can be decomposed into a parametric baseline (constant and linear terms) plus a data-driven smooth deviation modeled by an orthogonal basis derived from a second-order random walk prior; this hierarchy, implemented via INLA, yields fast stable inference, recovers non-linear trajectories in simulation, and identifies U-shaped BMI mortality risk plus non-linear slope effects in the Health and Retirement Study.

What carries the argument

The hierarchical decomposition of the marker scaling effect into a parametric baseline and an orthogonal-basis smooth deviation drawn from a second-order random walk prior.

If this is right

  • Users can directly control model flexibility and test the linearity assumption with ordinary information criteria.
  • Simulation studies show the method recovers complex non-linear trajectories with good accuracy.
  • The BMI-mortality analysis reveals a U-shaped risk for current BMI value and a non-linear effect of the rate of weight change.
  • The framework runs stably and quickly enough to be used on large health datasets via the INLAjoint package.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same hierarchy could be applied to other longitudinal markers or to multiple markers jointly without major changes to the computational setup.
  • If the random-walk basis approximates many biological curves well, the approach may serve as a default flexible option in routine joint-model analyses.
  • Clinical risk calculators could incorporate the resulting non-linear surfaces once they are validated on additional cohorts.

Load-bearing premise

The non-linear deviation from the linear baseline can be adequately captured by the chosen orthogonal basis from the second-order random walk prior.

What would settle it

A controlled simulation in which the true association is a known non-linear function poorly approximated by the random-walk basis, yet the fitted model still passes information-criterion checks or recovers the trajectory inaccurately.

Figures

Figures reproduced from arXiv: 2604.18057 by Denis Rustand, H{\aa}vard Rue, Karen Leffondre, Lisa Le Gall.

Figure 1
Figure 1. Figure 1: Estimated log hazard ratios across the three simulation scenarios com [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Estimated log-hazard ratios for all-cause mortality associated with the [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
read the original abstract

Joint models for longitudinal and time-to-event data are increasingly used in health research to characterize the association between biomarker trajectories and the risk of clinical events. However, these models usually assume a linear relationship between the longitudinal marker and the log-hazard of the event. This assumption is rarely verified and often fails to capture complex biological mechanisms, such as U-shaped risk profiles or plateau effects. In this paper, we propose a fast and stable hierarchical framework for non-linear association structures in joint models using Integrated Nested Laplace Approximations (INLA), implemented in the INLAjoint R package. Our approach builds upon a unified framework where the scaling effect of the marker is decomposed into a parametric baseline (constant and linear components) and a data-driven smooth deviation modeled via an orthogonal basis derived from a second-order random walk. This natural hierarchy allows researchers to adapt model flexibility directly and verify the linearity assumption using standard information criteria. Through simulation studies, we demonstrate that the proposed method accurately recovers complex non-linear trajectories. We illustrate the practical utility of our framework by analyzing the joint association of the current value and current slope of body mass index (BMI) with all-cause mortality in the Health and Retirement Study. This analysis reveals a U-shaped mortality risk for the BMI value, and a non-linear effect for the rate of weight change, where a declining weight trajectory is associated with higher mortality risk.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes a hierarchical INLA-based framework for joint longitudinal-survival models that decomposes the marker-to-hazard association into a parametric baseline (constant plus linear) plus a smooth non-linear deviation expressed on an orthogonal basis derived from a second-order random walk prior. This structure is claimed to enable efficient Bayesian inference, direct verification of the linearity assumption via standard information criteria, accurate recovery of complex trajectories in simulations, and improved modeling of real associations such as the U-shaped BMI-mortality link and non-linear slope effect in the Health and Retirement Study.

Significance. If the central claims hold, the work supplies a computationally tractable and practically usable extension of joint models that relaxes the ubiquitous linear-association assumption while retaining the ability to test that assumption. The INLA implementation and the explicit hierarchy for model selection address a genuine methodological gap in health-research applications where non-linear biomarker effects are biologically plausible.

major comments (2)
  1. [Simulation studies] Simulation studies: the abstract asserts that the method 'accurately recovers complex non-linear trajectories,' yet no quantitative recovery metrics (bias, RMSE, coverage probabilities, or interval scores for the estimated non-linear function), no error bars, and no sensitivity checks under realistic measurement-error or irregular-observation regimes are reported. These diagnostics are load-bearing for the headline claim of accurate recovery.
  2. [Methods] Methods (hierarchical decomposition and orthogonality): the claim that the RW2-derived basis remains orthogonal to the parametric baseline under the joint likelihood (with error-prone, irregularly timed observations) is central to both the recovery guarantee and the use of information criteria for linearity verification. No explicit diagnostics for leakage, effective support of the basis, or identifiability under the coupled longitudinal-survival likelihood are provided; without them the separation between baseline and deviation cannot be taken as given.
minor comments (1)
  1. [Abstract] The abstract would be strengthened by stating the number of simulation replicates, the specific performance metrics used, and the range of non-linear functions tested.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments identify key areas where additional rigor will strengthen the presentation of our simulation results and the supporting evidence for the hierarchical decomposition. We respond to each major comment below and commit to the corresponding revisions.

read point-by-point responses

  1. Referee: Simulation studies: the abstract asserts that the method 'accurately recovers complex non-linear trajectories,' yet no quantitative recovery metrics (bias, RMSE, coverage probabilities, or interval scores for the estimated non-linear function), no error bars, and no sensitivity checks under realistic measurement-error or irregular-observation regimes are reported. These diagnostics are load-bearing for the headline claim of accurate recovery.

    Authors: We agree that the current simulation section relies primarily on visual inspection and would benefit from quantitative support. In the revised manuscript we will add bias, RMSE, coverage probabilities, and interval scores for the recovered non-linear functions, include error bars on all simulation plots, and report sensitivity results under added measurement error and irregular observation schedules. These metrics will be presented in a new table and expanded figure set within the simulation studies section. revision: yes

  2. Referee: Methods (hierarchical decomposition and orthogonality): the claim that the RW2-derived basis remains orthogonal to the parametric baseline under the joint likelihood (with error-prone, irregularly timed observations) is central to both the recovery guarantee and the use of information criteria for linearity verification. No explicit diagnostics for leakage, effective support of the basis, or identifiability under the coupled longitudinal-survival likelihood are provided; without them the separation between baseline and deviation cannot be taken as given.

    Authors: The RW2 basis is constructed to be orthogonal to the constant and linear terms by design, using the standard orthogonalization step prior to model fitting. We recognize, however, that explicit verification under the joint likelihood is necessary. In the revision we will add correlation diagnostics between the basis functions and the parametric baseline evaluated on simulated joint data, report effective degrees of freedom, and include a brief identifiability analysis in the methods section (with supporting results moved to an appendix). These additions will directly address potential leakage and confirm the separation holds in practice. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the INLA hierarchical decomposition for non-linear joint models

full rationale

The paper's central construction decomposes the marker scaling into a parametric baseline plus an orthogonal RW2-derived smooth deviation, then uses standard INLA and information criteria for flexibility and linearity checks. This is a modeling choice with external simulation validation for recovery of trajectories; no derivation step reduces a claimed prediction or result to an input by construction, and any self-citations are not load-bearing for the core claims. The approach remains self-contained against established INLA and random-walk machinery.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility into exact parameter counts; the central modeling choice rests on one key domain assumption about the random-walk basis.

axioms (1)
  • domain assumption The non-linear component of the marker effect can be represented by an orthogonal basis derived from a second-order random walk prior.
    Explicitly stated in the abstract as the unified framework that decomposes scaling into baseline and smooth deviation.

pith-pipeline@v0.9.0 · 5562 in / 1316 out tokens · 25675 ms · 2026-05-10T04:12:31.634128+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

  1. [1]

    Biometrics , pages=

    A joint model for survival and longitudinal data measured with error , author=. Biometrics , pages=. 1997 , publisher=

    work page 1997

  2. [2]

    2012 , publisher=

    Joint models for longitudinal and time-to-event data: With applications in R , author=. 2012 , publisher=

    work page 2012

  3. [3]

    Scientific Reports , volume=

    Longitudinal uric acid has nonlinear association with kidney failure and mortality in chronic kidney disease , author=. Scientific Reports , volume=. 2023 , publisher=

    work page 2023

  4. [4]

    The lancet , volume=

    Body-mass index and all-cause mortality: individual-participant-data meta-analysis of 239 prospective studies in four continents , author=. The lancet , volume=. 2016 , publisher=

    work page 2016

  5. [5]

    Statistics in medicine , volume=

    Nonlinear association structures in flexible Bayesian additive joint models , author=. Statistics in medicine , volume=. 2018 , publisher=

    work page 2018

  6. [6]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=

    Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations , author=. Journal of the Royal Statistical Society Series B: Statistical Methodology , volume=. 2009 , publisher=

    work page 2009

  7. [7]

    Computational Statistics & Data Analysis , volume=

    A new avenue for Bayesian inference with INLA , author=. Computational Statistics & Data Analysis , volume=. 2023 , publisher=

    work page 2023

  8. [8]

    Statistical Science , volume=

    Penalising Model Component Complexity: A Principled, Practical Approach to Constructing Priors , author=. Statistical Science , volume=. 2017 , doi=

    work page 2017

  9. [9]

    , author=

    Asymptotic equivalence of Bayes cross validation and widely applicable information criterion in singular learning theory. , author=. Journal of machine learning research , volume=. 2010 , url=

    work page 2010

  10. [10]

    Journal of the royal statistical society: Series b (statistical methodology) , volume=

    Bayesian measures of model complexity and fit , author=. Journal of the royal statistical society: Series b (statistical methodology) , volume=. 2002 , publisher=. doi:doi=

    work page 2002

  11. [11]

    Statistica Sinica , pages=

    Joint modeling of longitudinal and time-to-event data: an overview , author=. Statistica Sinica , pages=. 2004 , publisher=

    work page 2004

  12. [12]

    Biometrika , volume=

    Covariate measurement errors and parameter estimation in a failure time regression model , author=. Biometrika , volume=. 1982 , publisher=

    work page 1982

  13. [13]

    Biometrical Journal , volume=

    Joint modelling of longitudinal and time-to-event data with application to predicting abdominal aortic aneurysm growth and rupture , author=. Biometrical Journal , volume=. 2011 , publisher=

    work page 2011

  14. [14]

    Biometrics , volume=

    Semiparametric modeling of longitudinal measurements and time-to-event data--a two-stage regression calibration approach , author=. Biometrics , volume=. 2008 , publisher=

    work page 2008

  15. [15]

    Scandinavian Journal of Statistics , volume=

    Dynamic prediction by landmarking in event history analysis , author=. Scandinavian Journal of Statistics , volume=. 2007 , publisher=

    work page 2007

  16. [16]

    Statistics in Medicine , volume=

    Landmarking 2.0: bridging the gap between joint models and landmarking , author=. Statistics in Medicine , volume=. 2022 , publisher=

    work page 2022

  17. [17]

    Frontiers in Medicine , volume=

    Non-linear association between CD4+ T-cell counts and mortality risk in people living with HIV: evidence from a 10-year cohort study , author=. Frontiers in Medicine , volume=. 2025 , publisher=

    work page 2025

  18. [18]

    Biostatistics , volume=

    Fast and flexible inference for joint models of multivariate longitudinal and survival data using integrated nested Laplace approximations , author=. Biostatistics , volume=. 2024 , publisher=

    work page 2024

  19. [19]

    Statistics in medicine , volume=

    Comparison of algorithms to generate event times conditional on time-dependent covariates , author=. Statistics in medicine , volume=. 2008 , publisher=

    work page 2008

  20. [20]

    American journal of epidemiology , volume=

    Obesity and mortality risk: new findings from body mass index trajectories , author=. American journal of epidemiology , volume=. 2013 , publisher=

    work page 2013

  21. [21]

    The Gerontologist , volume=

    Body mass trajectories and mortality among older adults: A joint growth mixture--discrete-time survival analysis , author=. The Gerontologist , volume=. 2014 , publisher=

    work page 2014

  22. [22]

    Annals of epidemiology , volume=

    Life-long body mass index trajectories and mortality in two generations , author=. Annals of epidemiology , volume=. 2021 , publisher=

    work page 2021

  23. [23]

    Biometrics , pages=

    Life tables with concomitant information , author=. Biometrics , pages=. 1976 , publisher=

    work page 1976

  24. [24]

    Journal of the Royal Statistical Society Series C: Applied Statistics , volume=

    Fitting Cox’s regression model to survival data using GLIM , author=. Journal of the Royal Statistical Society Series C: Applied Statistics , volume=. 1980 , publisher=

    work page 1980

  25. [25]

    Modeling survival data: extending the Cox model , pages=

    The cox model , author=. Modeling survival data: extending the Cox model , pages=. 2000 , publisher=

    work page 2000

  26. [26]

    Biometrical Journal , volume=

    Bayesian estimation of two-part joint models for a longitudinal semicontinuous biomarker and a terminal event with INLA: Interests for cancer clinical trial evaluation , author=. Biometrical Journal , volume=. 2023 , publisher=

    work page 2023

  27. [27]

    2026 , publisher=

    Bayesian Survival, Longitudinal, and Joint Models with INLA , author=. 2026 , publisher=

    work page 2026

  28. [28]

    Scandinavian Journal of Statistics , volume=

    Approximate Bayesian inference for survival models , author=. Scandinavian Journal of Statistics , volume=. 2011 , publisher=

    work page 2011

  29. [29]

    2003 , publisher=

    Semiparametric regression , author=. 2003 , publisher=

    work page 2003

  30. [30]

    Statistics and computing , volume=

    Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC , author=. Statistics and computing , volume=. 2017 , publisher=

    work page 2017

  31. [31]

    bmj , volume=

    Dealing with continuous variables and modelling non-linear associations in healthcare data: practical guide , author=. bmj , volume=. 2025 , publisher=

    work page 2025