Improving Disease Risk Estimation in Small Areas by Accounting for Spatiotemporal Local Discontinuities

A. Adin; G. Santaf\'e; M.D. Ugarte

arxiv: 2509.19889 · v2 · submitted 2025-09-24 · 📊 stat.ME

Improving Disease Risk Estimation in Small Areas by Accounting for Spatiotemporal Local Discontinuities

G. Santaf\'e , A. Adin , M.D. Ugarte This is my paper

Pith reviewed 2026-05-18 14:25 UTC · model grok-4.3

classification 📊 stat.ME

keywords spatiotemporal clusteringscan statisticsBayesian disease mappingsmall area estimationrisk discontinuitiescancer mortalityrelative risk estimation

0 comments

The pith

A greedy scan-statistic method detects spatiotemporal clusters and improves Bayesian estimates of small-area disease risk.

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

The paper introduces a two-step procedure that first locates clusters of elevated or reduced disease risk across space and time using a greedy search inside scan windows, then inserts those clusters as adjustments inside a Bayesian hierarchical spatiotemporal model. Standard models assume smooth risk surfaces and therefore miss abrupt local changes; this method explicitly accounts for the discontinuities so that risk estimates become more accurate in small geographic units. When applied to municipality-level cancer mortality records across Spain, the detected clusters match known patterns better than those from SaTScan and the resulting Bayesian model shows improved fit and predictive scores compared with both SaTScan-adjusted and ordinary BYM2 models. A reader should care because more reliable maps of high- and low-risk zones can guide public-health resource allocation where data are sparse.

Core claim

The central claim is that a greedy scan-statistic algorithm can flexibly identify spatiotemporal clusters of disease risk over large domains, and that feeding these clusters as fixed adjustments into a Bayesian spatiotemporal model produces relative-risk estimates that account for local discontinuities and outperform both SaTScan-based and standard BYM2 models on DIC, WAIC, and logarithmic scoring rules.

What carries the argument

The greedy search within the scan window for cluster detection, followed by insertion of the resulting cluster indicators as fixed effects that break the smoothness assumption inside the Bayesian hierarchical model.

If this is right

Cluster detection accuracy exceeds that of SaTScan on the same mortality data.
Bayesian models that include the detected clusters achieve lower DIC and WAIC values than both SaTScan-adjusted and standard BYM2 models.
Logarithmic scores for out-of-sample risk predictions improve when cluster information is incorporated.
High- and low-risk municipalities are identified more precisely once local discontinuities are modeled explicitly.

Where Pith is reading between the lines

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

The same two-step logic could be applied to other count-based phenomena, such as crime incidents or traffic accidents, wherever sharp spatial boundaries are suspected.
If the greedy clusters prove stable under repeated sampling, they could serve as a data-driven way to define custom neighborhoods for other spatial models.
Extending the scan window to include temporal covariates might allow the method to flag emerging risk shifts before they become large enough for conventional surveillance systems.

Load-bearing premise

The clusters returned by the greedy scan algorithm correspond to genuine jumps in the underlying risk surface rather than being produced by the search procedure or by noise in the counts.

What would settle it

Generate synthetic count data on the same Spanish municipality grid in which the true locations and timings of risk discontinuities are known exactly, then check whether the greedy algorithm recovers those exact discontinuities at a high rate.

Figures

Figures reproduced from arXiv: 2509.19889 by A. Adin, G. Santaf\'e, M.D. Ugarte.

**Figure 2.** Figure 2: Clusters identified by the GscanStat algorithm in the analysis of overall cancer mortality in [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: Standardized Mortality Ratios (top) and posterior median estimates of relative risks (bottom) [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

**Figure 4.** Figure 4: Posterior exceedance probabilities obtained by noCluster and GscanStat models for high-risk [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

read the original abstract

This work proposes a two-step method to enhance disease risk estimation in small areas by integrating spatiotemporal cluster detection within a Bayesian hierarchical spatiotemporal model. First, we introduce an efficient scan-statistic-based clustering algorithm that employs a greedy search within the scan window, enabling flexible cluster detection across large spatial domains. We then integrate these detected clusters into a Bayesian spatiotemporal model to estimate relative risks, explicitly accounting for identified risk discontinuities. We apply this methodology to large-scale cancer mortality data at the municipality level across continental Spain. Our results show our method offers superior cluster detection accuracy compared to SaTScan. Furthermore, integrating cluster information into a Bayesian spatiotemporal model significantly improves model fit and risk estimate performance, as evidenced by better DIC, WAIC, and logarithmic scores than SaTScan-based or standard BYM2 models. This methodology provides a powerful tool for epidemiological analysis, offering a more precise identification of high- and low-risk areas and enhancing the accuracy of risk estimation models.

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.

Desk Editor's Note private letter to a colleague

This paper offers a two-step greedy scan plus BYM2 integration for handling abrupt risk jumps in small-area disease data, but the fit gains look vulnerable to post-selection effects since clusters are treated as fixed.

read the letter

The main takeaway is a concrete procedure that first runs a greedy-search scan statistic to locate spatiotemporal clusters in count data, then feeds those clusters as fixed adjustments into a BYM2-style Bayesian spatiotemporal model so that relative-risk estimates can respect local discontinuities instead of smoothing over them. The authors test it on municipality-level cancer mortality across continental Spain and report better cluster recovery than SaTScan plus lower DIC, WAIC, and log scores than either SaTScan hybrids or plain BYM2. That real-data scale and the direct comparison to existing tools are the parts that actually land. The greedy search step itself is presented as an efficiency improvement for large domains, which is a modest but usable refinement over standard scan-statistic implementations. The Spain application shows the combined pipeline can be run end-to-end on national registry data, which is the kind of evidence that matters for applied users. The soft spot is exactly the one the stress-test flags: once the greedy algorithm selects the clusters, they enter the second-stage model as known quantities with no uncertainty propagated and no mention of held-out space-time units or simulation checks for selection bias. Any reported improvement in model-fit metrics therefore risks being partly circular, and the abstract gives no indication that the authors ran cross-validation or sensitivity analyses on the cluster step. Minor additional questions would be how sensitive the final risk maps are to the scan-window size or the greedy stopping criterion, but those are secondary to the main validation gap. Readers already working with BYM2 or small-area health mapping will find the method description and the Spain results directly usable; the paper is not aimed at a broad methodological audience. It is coherent on its own terms and shows honest engagement with the relevant tools, so it deserves a serious referee who can press on the validation design and ask for explicit checks against post-selection overfitting. I would send it out rather than desk-reject.

Referee Report

3 major / 2 minor

Summary. The paper proposes a two-step method for small-area disease risk estimation: a greedy scan-statistic algorithm detects spatiotemporal clusters of local discontinuities, which are then inserted as fixed adjustments into a Bayesian hierarchical spatiotemporal model (extending BYM2) to estimate relative risks. Applied to municipality-level cancer mortality data across continental Spain, the work claims superior cluster detection accuracy relative to SaTScan and improved model fit and predictive performance (lower DIC, WAIC, and logarithmic scores) compared with both SaTScan-based and standard BYM2 models.

Significance. If the reported gains survive proper accounting for cluster-selection uncertainty and are shown to generalize beyond the fitting data, the approach could meaningfully advance spatial epidemiology by relaxing the smoothness assumptions of conventional BYM2 models while retaining their computational tractability. The combination of an efficient greedy scan with a Bayesian risk model addresses a practical gap in analyzing large spatial domains where standard scan statistics may miss flexible spatiotemporal patterns.

major comments (3)

[Abstract / Methods (cluster integration)] Abstract and Methods (two-step procedure): the detected clusters are treated as fixed, known discontinuities when inserted into the Bayesian spatiotemporal model, yet the manuscript provides no uncertainty quantification, Bayesian model averaging over candidate clusters, or joint likelihood that would propagate selection uncertainty from the greedy scan step. Because the same data are used for both cluster detection and subsequent risk estimation, any improvement in DIC, WAIC or log scores is vulnerable to post-selection overfitting; this directly undermines the central claim of superior performance.
[Results / Model comparison] Results (performance comparison): the abstract asserts better DIC, WAIC and logarithmic scores than SaTScan-based or standard BYM2 models, but supplies no information on validation design (held-out space-time units, cross-validation, or simulation benchmarks), sensitivity to modeling choices, or error propagation from the first-step clusters. Without these details the superiority claim cannot be verified and may reflect in-sample optimization rather than genuine improvement.
[Methods (cluster detection)] Methods (greedy scan-statistic): the premise that clusters identified by the greedy search accurately represent genuine spatiotemporal risk jumps rather than artifacts of the search procedure or noise is load-bearing for the entire pipeline, yet no simulation study or external validation against known discontinuities is described to support this assumption.

minor comments (2)

[Methods] The description of the greedy search within the scan window would benefit from explicit pseudocode or a numbered algorithmic outline to ensure reproducibility.
[Model specification] Notation for the spatiotemporal random effects and the precise functional form by which cluster indicators enter the linear predictor should be stated explicitly (e.g., as an additional fixed or random term) rather than left implicit.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We are grateful to the referee for providing a detailed and constructive report. The comments highlight important aspects regarding uncertainty quantification and validation that we will address in the revision. We respond to each major comment below.

read point-by-point responses

Referee: [Abstract / Methods (cluster integration)] Abstract and Methods (two-step procedure): the detected clusters are treated as fixed, known discontinuities when inserted into the Bayesian spatiotemporal model, yet the manuscript provides no uncertainty quantification, Bayesian model averaging over candidate clusters, or joint likelihood that would propagate selection uncertainty from the greedy scan step. Because the same data are used for both cluster detection and subsequent risk estimation, any improvement in DIC, WAIC or log scores is vulnerable to post-selection overfitting; this directly undermines the central claim of superior performance.

Authors: We agree that treating the clusters as fixed introduces a limitation by not accounting for the uncertainty in the cluster detection step. This two-step approach prioritizes computational efficiency for large spatial domains, as a fully joint model would be computationally prohibitive. In the revised manuscript, we will add a dedicated subsection in the Discussion to explicitly discuss the potential for post-selection bias and its implications for the reported performance metrics. We will also conduct additional sensitivity analyses by re-running the Bayesian model with perturbed cluster boundaries or alternative cluster sets to assess robustness. While we cannot implement a full joint likelihood or Bayesian averaging within the current framework without major redesign, these additions will provide a more balanced presentation of the results. We will adjust the claims in the abstract and conclusions to reflect that the improvements are conditional on the detected clusters. revision: partial
Referee: [Results / Model comparison] Results (performance comparison): the abstract asserts better DIC, WAIC and logarithmic scores than SaTScan-based or standard BYM2 models, but supplies no information on validation design (held-out space-time units, cross-validation, or simulation benchmarks), sensitivity to modeling choices, or error propagation from the first-step clusters. Without these details the superiority claim cannot be verified and may reflect in-sample optimization rather than genuine improvement.

Authors: We appreciate this observation and will revise the Results section to provide a clear description of the validation strategy employed. Specifically, we will detail how the logarithmic scores were computed using a temporal hold-out validation, where data from the last time period were withheld for prediction. We will also include cross-validation results across spatial regions and sensitivity checks to key hyperparameters in the Bayesian model. Regarding error propagation, we will note in the text that the reported metrics are conditional on the first-step clusters and discuss this as a limitation. These clarifications will strengthen the verifiability of our performance claims. revision: yes
Referee: [Methods (cluster detection)] Methods (greedy scan-statistic): the premise that clusters identified by the greedy search accurately represent genuine spatiotemporal risk jumps rather than artifacts of the search procedure or noise is load-bearing for the entire pipeline, yet no simulation study or external validation against known discontinuities is described to support this assumption.

Authors: The primary validation for the cluster detection comes from the real-world application to Spanish cancer mortality data, where the detected clusters align with known epidemiological patterns and outperform SaTScan in identifying relevant areas. However, we recognize that a simulation study would provide stronger evidence for the method's ability to recover true discontinuities. In the revised manuscript, we will incorporate a simulation study section that generates synthetic data with known spatiotemporal clusters and evaluates the greedy scan-statistic's detection accuracy, false positive rates, and comparison to SaTScan. This will directly address the concern about artifacts versus genuine patterns. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external benchmarks

full rationale

The paper describes a two-step procedure consisting of a greedy scan-statistic cluster detector followed by insertion of the resulting clusters as fixed adjustments inside a Bayesian spatiotemporal model. All reported performance gains are quantified via direct comparisons to SaTScan (for cluster detection accuracy) and to both SaTScan-augmented and standard BYM2 models (via DIC, WAIC and logarithmic scores). No equation or modeling step equates the claimed improvement to a redefinition of the input clusters themselves, nor does any central result reduce to a self-citation chain or an ansatz smuggled from prior work by the same authors. The evaluation therefore rests on independent external references rather than on quantities that are definitionally identical to the fitted clusters.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit list of fitted hyperparameters, background assumptions, or new postulated entities; all such elements remain unknown.

pith-pipeline@v0.9.0 · 5704 in / 1286 out tokens · 39124 ms · 2026-05-18T14:25:04.025476+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/ArithmeticFromLogic.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We introduce GscanStat, an efficient scan-statistic-based clustering algorithm that employs a greedy search within the scan window... integrate these detected clusters into a Bayesian spatiotemporal model... Equation (5) with fixed effects β_j for each cluster C_j

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

[1]

Adin, A., Lee, D., Goicoa, T., and Ugarte, M. (2019). A two-stage approach to estimate spatial and spatio-temporal disease risks in the presence of local discontinuities and clusters. Statistical Methods in Medical Research , 28(9):2595--2613

work page 2019
[2]

Adin, A., Orozco-Acosta, E., and Ugarte, M. (2025). Scalable Bayesian Disease Mapping Models for High-Dimensional Data . R package version 0.5.7

work page 2025
[3]

Anderson, C., Lee, D., and Dean, N. (2014). Identifying clusters in Bayesian disease mapping . Biostatistics , (15):457--469

work page 2014
[4]

Besag, J., York, J., and Molli \' e , A. (1991). Bayesian image restoration, with two applications in spatial statistics . Annals of the Institute of Statistical Mathematics , 43(1):1--20

work page 1991
[5]

and Raftery, A

Gneiting, T. and Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association , 102(477):359--378

work page 2007
[6]

D., and Hodges, J

Goicoa, T., Adin, A., Ugarte, M. D., and Hodges, J. S. (2018). In spatio-temporal disease mapping models, identifiability constraints affect PQL and INLA results . Stochastic Environmental Research and Risk Assessment , 32(3):749--770

work page 2018
[7]

Knorr-Held, L. (2000). Bayesian modelling of inseparable space-time variation in disease risk. Statistics in Medicine , 19(17--18):2555--2567

work page 2000
[8]

Kulldorff, M. (1997). A spatial scan statistic. Communications in Statistics - Theory and Methods , 26(6):1481--1496

work page 1997
[9]

Kulldorff, M., Athas, W., Feuer, E., Miller, B., and Key, C. (1998). Evaluating cluster alarms: A space-time scan statistic and brain cancer in los alamos. American Journal of Public Health , 88:1377--1380

work page 1998
[10]

Orozco-Acosta, E., Adin, A., and Ugarte, M. (2023). Big problems in spatio-temporal disease mapping: Methods and software. Computer Methods and Programs in Biomedicine , 231:107403

work page 2023
[11]

and Song, G

Quick, H. and Song, G. (2024). Reliable event rates for disease mapping. Journal of official statistics , 40(2):333--347

work page 2024
[12]

E., Etxeberria, J., and Ugarte, M

Retegui, G., Gelfand, A. E., Etxeberria, J., and Ugarte, M. D. (2025). On prior smoothing with discrete spatial data in the context of disease mapping. Statistical Methods in Medical Research , 0(0):09622802251362659

work page 2025
[13]

H., Simpson, D., and Rue, H

Riebler, A., S rbye, S. H., Simpson, D., and Rue, H. (2016). An intuitive B ayesian spatial model for disease mapping that accounts for scaling. Statistical Methods in Medical Research , 25(4):1145--1165

work page 2016
[14]

and Held, L

Rue, H. and Held, L. (2005). Gaussian Markov Random Fields: Theory and Applications . CRC Press, Boca Raton

work page 2005
[15]

Santafé, G., Adín, A., Lee, D., and Ugarte, M. (2021). Dealing with risk discontinuities to estimate cancer mortality risks when the number of small areas is large . Statistical Methods in Medical Research , 30(1):6--21

work page 2021
[16]

S rbye, S. H. and Rue, H. (2014). Scaling intrinsic gaussian markov random field priors in spatial modelling. Spatial Statistics , 8:39--51

work page 2014
[17]

J., Best, N

Spiegelhalter, D. J., Best, N. G., Carlin, B. P., and Van Der Linde, A. (2002). Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 64(4):583--639

work page 2002
[18]

Takahashi, K., Kulldorff, M., Tango, T., and Yih, K. (2008). A flexibly shaped space-time scan statistic for disease outbreak detection and monitoring. International Journal of Health Geographics , 17(7):583--639

work page 2008
[19]

and Takahashi, K

Tango, T. and Takahashi, K. (2002). A flexibly shaped spatial scan statistic for detecting clusters. International Journal of Health Geographics , 11(4):583--639

work page 2002
[20]

Watanabe, S. (2010). Asymptotic equivalence of B ayes cross validation and widely applicable information criterion in singular learning theory . Journal of Machine Learning Research , 11(Dec):3571--3594

work page 2010
[21]

Yin, X., Anderson, C., Lee, D., and Napier, G. (2025). Risk estimation and boundary detection in bayesian disease mapping. International Journal of Biostatistics . Published online May 22, 2025

work page 2025
[22]

Yin, X., Napier, G., Anderson, C., and Lee, D. (2022). Spatio-temporal disease risk estimation using clustering-based adjacency modelling. Statistical Methods in Medical Research , 31(6):1184--1203

work page 2022

[1] [1]

Adin, A., Lee, D., Goicoa, T., and Ugarte, M. (2019). A two-stage approach to estimate spatial and spatio-temporal disease risks in the presence of local discontinuities and clusters. Statistical Methods in Medical Research , 28(9):2595--2613

work page 2019

[2] [2]

Adin, A., Orozco-Acosta, E., and Ugarte, M. (2025). Scalable Bayesian Disease Mapping Models for High-Dimensional Data . R package version 0.5.7

work page 2025

[3] [3]

Anderson, C., Lee, D., and Dean, N. (2014). Identifying clusters in Bayesian disease mapping . Biostatistics , (15):457--469

work page 2014

[4] [4]

Besag, J., York, J., and Molli \' e , A. (1991). Bayesian image restoration, with two applications in spatial statistics . Annals of the Institute of Statistical Mathematics , 43(1):1--20

work page 1991

[5] [5]

and Raftery, A

Gneiting, T. and Raftery, A. E. (2007). Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association , 102(477):359--378

work page 2007

[6] [6]

D., and Hodges, J

Goicoa, T., Adin, A., Ugarte, M. D., and Hodges, J. S. (2018). In spatio-temporal disease mapping models, identifiability constraints affect PQL and INLA results . Stochastic Environmental Research and Risk Assessment , 32(3):749--770

work page 2018

[7] [7]

Knorr-Held, L. (2000). Bayesian modelling of inseparable space-time variation in disease risk. Statistics in Medicine , 19(17--18):2555--2567

work page 2000

[8] [8]

Kulldorff, M. (1997). A spatial scan statistic. Communications in Statistics - Theory and Methods , 26(6):1481--1496

work page 1997

[9] [9]

Kulldorff, M., Athas, W., Feuer, E., Miller, B., and Key, C. (1998). Evaluating cluster alarms: A space-time scan statistic and brain cancer in los alamos. American Journal of Public Health , 88:1377--1380

work page 1998

[10] [10]

Orozco-Acosta, E., Adin, A., and Ugarte, M. (2023). Big problems in spatio-temporal disease mapping: Methods and software. Computer Methods and Programs in Biomedicine , 231:107403

work page 2023

[11] [11]

and Song, G

Quick, H. and Song, G. (2024). Reliable event rates for disease mapping. Journal of official statistics , 40(2):333--347

work page 2024

[12] [12]

E., Etxeberria, J., and Ugarte, M

Retegui, G., Gelfand, A. E., Etxeberria, J., and Ugarte, M. D. (2025). On prior smoothing with discrete spatial data in the context of disease mapping. Statistical Methods in Medical Research , 0(0):09622802251362659

work page 2025

[13] [13]

H., Simpson, D., and Rue, H

Riebler, A., S rbye, S. H., Simpson, D., and Rue, H. (2016). An intuitive B ayesian spatial model for disease mapping that accounts for scaling. Statistical Methods in Medical Research , 25(4):1145--1165

work page 2016

[14] [14]

and Held, L

Rue, H. and Held, L. (2005). Gaussian Markov Random Fields: Theory and Applications . CRC Press, Boca Raton

work page 2005

[15] [15]

Santafé, G., Adín, A., Lee, D., and Ugarte, M. (2021). Dealing with risk discontinuities to estimate cancer mortality risks when the number of small areas is large . Statistical Methods in Medical Research , 30(1):6--21

work page 2021

[16] [16]

S rbye, S. H. and Rue, H. (2014). Scaling intrinsic gaussian markov random field priors in spatial modelling. Spatial Statistics , 8:39--51

work page 2014

[17] [17]

J., Best, N

Spiegelhalter, D. J., Best, N. G., Carlin, B. P., and Van Der Linde, A. (2002). Bayesian measures of model complexity and fit. Journal of the Royal Statistical Society: Series B (Statistical Methodology) , 64(4):583--639

work page 2002

[18] [18]

Takahashi, K., Kulldorff, M., Tango, T., and Yih, K. (2008). A flexibly shaped space-time scan statistic for disease outbreak detection and monitoring. International Journal of Health Geographics , 17(7):583--639

work page 2008

[19] [19]

and Takahashi, K

Tango, T. and Takahashi, K. (2002). A flexibly shaped spatial scan statistic for detecting clusters. International Journal of Health Geographics , 11(4):583--639

work page 2002

[20] [20]

Watanabe, S. (2010). Asymptotic equivalence of B ayes cross validation and widely applicable information criterion in singular learning theory . Journal of Machine Learning Research , 11(Dec):3571--3594

work page 2010

[21] [21]

Yin, X., Anderson, C., Lee, D., and Napier, G. (2025). Risk estimation and boundary detection in bayesian disease mapping. International Journal of Biostatistics . Published online May 22, 2025

work page 2025

[22] [22]

Yin, X., Napier, G., Anderson, C., and Lee, D. (2022). Spatio-temporal disease risk estimation using clustering-based adjacency modelling. Statistical Methods in Medical Research , 31(6):1184--1203

work page 2022