pith. sign in

arxiv: 2604.09012 · v1 · submitted 2026-04-10 · 📊 stat.ME

Spatially varying distributed lag non-linear models using Laplacian P-splines

Sara Rutten , Thomas Neyens , Elisa Duarte , Antonio Gasparrini , Christel Faes This is my paper

Pith reviewed 2026-05-10 18:13 UTC · model grok-4.3

classification 📊 stat.ME
keywords spatially varying DLNMLaplacian P-splinesBayesian modelingdistributed lag modelscount dataspatial heterogeneitytemperature-mortalityLaplace approximation
0
0 comments X

The pith

A Bayesian one-stage method enables spatially varying distributed lag non-linear models for sparse count data.

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

The paper develops a Bayesian approach to distributed lag non-linear models that can vary across locations. This matters because many environmental studies have small counts at each site, making standard two-stage meta-analysis unreliable. The method defines four variants that differ in how they model spatial dependence and how flexible the lag and exposure splines are. Laplace approximations replace full MCMC sampling to keep computation feasible. The approach is tested in simulations and applied to municipality-level temperature-mortality data across Sicily.

Core claim

We introduce a computationally efficient Bayesian one-stage spatially-varying DLNM for count data. We define four model variants, differing in the assumed spatial dependence structure and the flexibility of the DLNM spline specification. To address the computational burden, we use Laplace approximations as an efficient alternative to MCMC.

What carries the argument

Laplacian P-splines that jointly model the spatially varying exposure-response and lag-response surfaces inside a Bayesian hierarchical framework.

If this is right

  • Small-area count data can be analyzed in one stage without needing large local samples at every location.
  • Researchers can compare models that assume different strengths of spatial smoothing and different degrees of spline flexibility.
  • Environmental health applications can now capture local differences in how exposure affects outcomes over time.
  • Model selection tools are supplied so users can choose the variant that fits their data best.

Where Pith is reading between the lines

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

  • The same Laplace-P-spline construction could be applied to continuous or binary outcomes with only minor changes to the likelihood.
  • Adding time-varying spatial covariates might further improve how well the model captures changing local effects.
  • Direct comparison of run times and accuracy against existing two-stage DLNM packages on the same datasets would clarify practical gains.

Load-bearing premise

Laplace approximations remain accurate enough for reliable posterior inference in these flexible models even when the count data are sparse.

What would settle it

A simulation in which the Laplace-based posteriors for the spatially varying coefficients differ substantially from those obtained by running full MCMC on identical data would falsify the efficiency claim.

Figures

Figures reproduced from arXiv: 2604.09012 by Antonio Gasparrini, Christel Faes, Elisa Duarte, Sara Rutten, Thomas Neyens.

Figure 1
Figure 1. Figure 1: Assumed exposure-response relationship (overall cumulative RR) for dif [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Assumed overall cumulative RR at exposure level 8.5 on a map of [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Estimated overall cumulative RR for Setting 2 (averaged over the 250 [PITH_FULL_IMAGE:figures/full_fig_p021_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Estimated overall cumulative RR for different temperature values, com [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Estimated overall cumulative RR at 28 degrees, compared to a reference [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Estimated overall cumulative RR for different temperature values, com [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Posterior probability that a certain region belongs to the top 10% and top [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
read the original abstract

Although distributed lag non-linear models (DLNMs) are commonly used to quantify delayed and non-linear exposure-response relationships, most existing applications assume that these relationships are constant across space. However, in many geographical and environmental studies, local characteristics vary substantially across areas, making a spatially varying effect more realistic. Extending DLNMs to allow for spatial heterogeneity remains challenging, and only a limited number of modelling strategies have been proposed in literature. The most popular extension is a two-stage meta-analysis approach, which requires sufficiently large sample sizes at each location. Therefore, its usefulness is limited when working with sparse count data in small area data analyses. Although a number of alternative one-stage approaches have been introduced, their computational burden restricts their applicability in real-life data applications. In this paper, we introduce a computationally efficient Bayesian one-stage spatially-varying DLNM for count data. We define four model variants, differing in the assumed spatial dependence structure and the flexibility of the DLNM spline specification. To address the computational burden typically associated with these flexible models, we use Laplace approximations, offering an efficient alternative to classically used Markov Chain Monte Carlo (MCMC) approaches. Model comparison criteria are provided to facilitate the selection of a suitable model in a real-life data application. The proposed methods are evaluated through simulation studies, and their practical usefulness is illustrated through a real-life data application, investigating the temperature-mortality relationship in every municipality of Sicily, Italy.

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

1 major / 1 minor

Summary. The paper introduces four variants of a Bayesian one-stage spatially-varying distributed lag non-linear model (DLNM) for count data, using Laplacian P-splines and Laplace approximations for computational efficiency instead of MCMC. The variants differ in spatial dependence structure and DLNM spline flexibility; model comparison criteria are provided, with evaluation via simulation studies and an application to temperature-mortality relationships across all municipalities in Sicily, Italy.

Significance. If the Laplace approximations deliver accurate point estimates, credible intervals, and model selection for these flexible models on sparse count data, the work would offer a practical one-stage alternative to two-stage meta-analysis approaches that require large per-location samples, enabling more realistic spatial heterogeneity modeling in environmental epidemiology.

major comments (1)
  1. [Simulation studies] Simulation studies section: the claim that Laplace approximations enable reliable inference rests on unverified accuracy for sparse counts. Direct comparisons of Laplace-derived coverage, bias, and interval properties against MCMC (or INLA) on low-count replicates are needed, as the Poisson likelihood combined with non-linear lag surfaces and spatial penalties can induce posterior skewness that violates the local normality assumption.
minor comments (1)
  1. [Abstract] Abstract: the four model variants are mentioned but their precise differences in spatial dependence (e.g., intrinsic CAR) and spline flexibility are not summarized; a brief clause would improve clarity for readers.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive comments, which help strengthen the manuscript. We address the major comment point by point below.

read point-by-point responses

  1. Referee: Simulation studies section: the claim that Laplace approximations enable reliable inference rests on unverified accuracy for sparse counts. Direct comparisons of Laplace-derived coverage, bias, and interval properties against MCMC (or INLA) on low-count replicates are needed, as the Poisson likelihood combined with non-linear lag surfaces and spatial penalties can induce posterior skewness that violates the local normality assumption.

    Authors: We agree that direct comparisons of the Laplace approximation against MCMC (or INLA) on low-count data would provide stronger validation of the method's reliability, especially given the potential for posterior skewness in these models. The current simulation studies evaluate performance across varying count levels and spatial settings but do not include head-to-head MCMC benchmarks for coverage and bias. In the revised manuscript we will add a dedicated set of low-count simulation replicates that directly compare Laplace-derived point estimates, credible intervals, and frequentist coverage properties to MCMC results (and, where feasible, INLA). This addition will explicitly address concerns about the local normality assumption. revision: yes

Circularity Check

0 steps flagged

New model variants defined and evaluated independently via simulations

full rationale

The paper defines four explicit model variants for spatially varying DLNMs (differing in spatial dependence structure and P-spline flexibility on lags/space), applies standard Laplace approximations as a computational tool, and evaluates them through separate simulation studies plus a real-data application on Sicilian temperature-mortality counts. No equation or claim reduces by construction to a fitted input renamed as a prediction, no self-citation chain justifies the core construction, and the derivation chain remains self-contained against external simulation benchmarks rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the accuracy of Laplace approximations for posterior inference and on the suitability of the chosen spatial dependence structures for capturing heterogeneity in sparse count data; these are standard domain assumptions rather than new postulates.

axioms (1)
  • domain assumption Laplace approximations provide sufficiently accurate inference for the posterior distributions of the proposed spatially varying DLNM parameters.
    The paper explicitly adopts Laplace approximations as an efficient substitute for MCMC to address computational burden.

pith-pipeline@v0.9.0 · 5568 in / 1246 out tokens · 79158 ms · 2026-05-10T18:13:25.380131+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

2 extracted references · 2 canonical work pages

  1. [1]

    Economou, T., Parliari, D., Tobias, A., Dawkins, L., Steptoe, H., Sarran, C., Stoner, O., Lowe, R., and Lelieveld, J. (2025). Flexible distributed lag models for count data using mgcv. American Statistician, 79:371–382. Quijal-Zamorano, M., Martinez-Beneito, M. A., Ballester, J., and Mar´ ı-Dell’olmo, M. (2025). Spatial bayesian distributed lag non-linear...

    work page 2025

  2. [2]

    Rue, H., Martino, S., and Chopin, N. (2009). Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations.Journal of the Royal Statistical Society: Series B (Statistical Methodology), 71(2):319–392. Simpson, D., Rue, H., Riebler, A., Martins, T. G., and Sørbye, S. H. (2017). Penalising model component complex...

    work page 2009