Pith. sign in

REVIEW 2 major objections 4 minor 19 references

A Bayesian Joint Model for Multiple Point Processes with Application to Presence-Only Data

T0 review · 2 major / 4 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Bayesian network of Poisson processes with latent unobserved points yields exact joint inference on dependence, sampling bias, and missing occurrences for presence-only spatial data.

desk verdict Solid, usable extension of the authors’ exact presence-only Poisson work to a BN-structured multivariate setting; the tractable augmentation and η-mediated dependence are real, the fixed-DAG/distance kernel is the main modeling caveat, and it deserves referee time. read the letter →

arxiv 2607.02786 v1 pith:MUR555FL submitted 2026-07-02 stat.ME stat.AP

Laura C. M. Teixeira , Dani Gamerman , Vinicius Peripato , Carolina Levis This is my paper · ORCID
classification stat.MEstat.AP MSC 62M3062F1562H11
keywords multivariatepointprocessesBayesiannetworksPoissonprocesspresence-onlydatajointspeciesdistributionaugmentationAmazonforest
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper shows that multiple presence-only point patterns can be modeled jointly by placing them in a Bayesian network of inhomogeneous Poisson processes whose intensities depend on environmental covariates, observability covariates, and explicit summaries of parent processes. The usual intractable integrals disappear once each process is augmented with two latent processes—one for unobserved occurrences and one auxiliary—so that the superposed process is homogeneous and the likelihood becomes fully tractable. Blocked Gibbs sampling then recovers the dependence parameters, the environmental and bias coefficients, and the number and locations of the missing points. Simulations recover the true parameters even under heavy missingness, and an Amazon application finds species-specific associations with earthworks and Amazonian Dark Earths that match prior ecological claims while also mapping thousands of still-hidden archaeological sites.

What carries the argument

The data-augmented likelihood (Eqs. 3–8): each process Xi is thinned from a homogeneous Poisson process of intensity λ*i via intensity and detection functions qi and pi; the resulting superposed process removes all intractable integrals and supplies the latent points X′i that both complete the dependence graph and are themselves targets of inference.

What would settle it

Generate synthetic multi-type patterns from a non-DAG or non-distance interaction, fit the model under the paper’s assumed graph and d(·), and check whether the posterior for η still covers the true (possibly zero) dependence and whether the recovered latent maps match the known missing points.

Watch

Extended reading notes

Core claim

A Bayesian network of inhomogeneous Poisson processes, each augmented by latent unobserved and auxiliary processes, produces an exactly evaluable likelihood that supports joint posterior inference on dependence parameters, environmental and observability coefficients, and the full spatial distribution of unobserved occurrences for multiple presence-only patterns.

Load-bearing premise

Dependence is assumed to act only through a pre-specified directed acyclic graph and a hand-chosen distance summary of parent points; if the true interaction is non-directional, non-distance-based, or acts outside that summary, the estimated relationships mis-specify the real joint structure.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes a Bayesian joint model for multiple inhomogeneous Poisson point processes whose dependence is encoded by a directed acyclic graph (Bayesian network). Intensity functions include environmental covariates and a finite-dimensional summary d(s, Pa(Xi)) of parent processes, so that interaction strength is parameterized by coefficients η. For presence-only data the model introduces, for each type, a latent unobserved process X'i and an auxiliary process Ui; their superposition with the observed process is homogeneous with intensity λ*i, collapsing the intractable integral in the likelihood to |D|λ*i and yielding an exactly evaluable augmented likelihood (Eqs. 3–8). Blocked Gibbs MCMC (Algorithm 1) samples λ*, regression coefficients (via Pólya–Gamma or Albert–Chib augmentation), and the latent processes under a topological order. Simulations (two graph structures, four scenarios, 50 replicates) recover parameters with near-nominal 90% coverage and correctly cover η=0 when edges are absent. An Amazon application jointly models earthworks, Amazonian Dark Earths, and three tree species, producing estimates of η, observability bias, and maps/counts of unobserved occurrences that are consistent with prior ecological literature.

Significance. If the technical claims hold, the paper supplies a coherent exact-inference framework for multi-type presence-only point patterns in which dependence is an explicit, estimable parameter rather than a residual correlation or shared random effect. The data-augmentation construction (superposition to homogeneous PP, thinning interpretation) is carefully derived and extends the univariate exact-Bayes approach of Moreira & Gamerman (2022) to a BN-structured multivariate setting; the same augmentation also delivers posterior inference on the number and spatial distribution of unobserved points, which is scientifically useful. Simulation design (known ground truth, coverage of null edges) and the fully observed special case (Section 3.4) strengthen the methodological contribution. The Amazon illustration shows that the machinery can be run at continental scale and yields interpretable η estimates, though scientific conclusions remain conditional on the chosen DAG and distance kernel.

major comments (2)
  1. Section 3.1 and Eqs. (9), (17)–(19): dependence enters solely through a pre-specified DAG and a hand-chosen finite-dimensional summary d(s, Pa(Xi)) (here truncated inverse min-distance with a 25 km cutoff and +0.1 offset). Simulations recover η only under the same generative d; the paper does not examine robustness to misspecified kernels (e.g., counts in a radius, non-truncated distance, or non-distance interactions). Because the central scientific claim is inference on the existence and magnitude of relations via η, a sensitivity analysis or explicit discussion of how alternative d choices affect posterior conclusions is needed before the Amazon η estimates can be treated as more than illustrative.
  2. Section 5.1, Table 1 and Scenarios 3–4: empirical coverage for the intercept β0 of the child process falls to 66% and 58% when the parent is heavily unobserved. The paper notes the drop but does not diagnose whether it reflects finite-sample bias, prior influence, or partial non-identifiability of the intensity level under preferential sampling. Clarifying this (e.g., by reporting bias/RMSE of posterior means or by a targeted simulation with varying prior strength) is load-bearing for the claim that the scheme recovers true parameters under realistic missingness.
minor comments (4)
  1. Section 3.3: the two-stage structure-learning proposal (start dense, prune by credibility intervals on η) is useful but informal; a short note on multiple-testing or on the risk of canceling effects when d aggregates several parents would help readers use it safely.
  2. Algorithm 1 and Appendix A: the binary success/failure coding for ζ and δ is clear, but the notation n_x̃i for the observability regression is easy to miss; a one-line definition in the main text would improve readability.
  3. Figure 4 and Table 2: posterior densities and CIs for η are informative; adding the corresponding posterior means of the total number of unobserved points (already given in text) to a small table would make the latent-process results easier to cite.
  4. Throughout: a few typographical inconsistencies (e.g., “Ignoring then can also limit” in the Introduction; occasional spacing in λ*i) should be cleaned in revision.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; only expected methodological self-citation of the univariate augmentation foundation, which does not force the multivariate BN claim or the η inference.

full rationale

The core derivation (BN factorization of the joint PP likelihood in Eq. 1, data-augmented intensities in Eqs. 3–5 whose superposition collapses the intractable integral to the closed form |D|∑λ*_i in Eq. 8, and the blocked Gibbs sampler of Algorithm 1) is self-contained and rests on standard Poisson thinning/superposition properties plus the DAG topological order. The dependence parameters η and the summary d(s,Pa(Xi)) are free model components estimated from data under stated priors; simulations recover known ground-truth values (including correct coverage of η=0 when edges are absent) and the Amazon application is presented as an illustration whose signs and magnitudes are compared to external literature rather than being forced by construction. The sole self-citation of Moreira & Gamerman (2022) (and Gonçalves & Gamerman 2018) is ordinary lineage for the univariate presence-only augmentation that is being extended; it is not load-bearing for the new joint BN structure, the multi-process latent inference, or any claimed prediction. No self-definitional loop, no fitted-input-called-prediction, no uniqueness theorem imported from the authors, and no renaming of a known empirical pattern appear in the derivation chain.

Assumptions & free parameters 6 free parameters · 6 assumptions · 3 invented entities

The central claim rests on standard Poisson/BN math plus domain modeling choices (DAG known or over-connected then pruned; intensity/observability logistic or probit links; distance-based interaction kernel; weak Gamma/Normal priors). Free parameters are the usual regression and intensity-bound coefficients plus a few hand-chosen interaction-kernel constants. Latent X′ and U processes are computational constructs with scientific interpretation for X′, not new physical entities.

free parameters (6)
  • λ*_i intensity upper bounds
    Positive scale parameters for each process; Gamma priors with shape/rate 0.001; posterior means drive total expected (including unobserved) counts.
  • β_i environmental intensity coefficients
    Fitted Normal-prior regression coefficients linking covariates Z_i to q_i; central to habitat interpretation.
  • δ_i observability coefficients
    Fitted coefficients for sampling-bias surface p_i (roads, tree cover); required to separate bias from occurrence.
  • η_i parent-process interaction coefficients
    Load-bearing free parameters for edge strength; posterior intervals used both for scientific claims and graph pruning.
  • 25 km distance truncation and +0.1 offset in d_ij
    Hand-chosen kernel constants in the Amazon application (Eq. 19), motivated by Levis et al. but not estimated; change the definition of proximity that η multiplies.
  • Prior hyperparameters (N(0,10), Gamma(0.001,0.001))
    Weakly informative choices fixed across simulations and application; influence posterior scale especially for sparse processes.
assumptions (6)
  • domain assumption Each process is an inhomogeneous Poisson point process conditionally on parents and parameters.
    Stated throughout §2–3; excludes clustering beyond intensity covariates and parent summaries.
  • domain assumption Joint distribution factorizes according to a directed acyclic Bayesian network (Eq. 1).
    §2–3; inference is conditional on G; cycles forbidden; topological order required for Gibbs.
  • standard math Superposition of (Xi, X′i, Ui) is homogeneous PP with constant intensity λ*_i, enabling exact likelihood (Eq. 8).
    Standard thinning/superposition identity used as in Gonçalves–Gamerman / Moreira–Gamerman.
  • ad hoc to paper Parent influence enters only through a finite-dimensional summary d(s, Pa(Xi)) of parent locations (observed and latent).
    §3.1; min-distance or truncated inverse distance chosen by the authors; not derived from first principles.
  • domain assumption Observability is independent thinning with probability p_i(s) depending only on W_i covariates, not on other processes.
    Eqs. 3–5, 10; separates detection from biotic dependence.
  • ad hoc to paper Prior independence across processes for λ*, ζ=(β,η), and δ.
    §3.2; simplifies blocked Gibbs factorization.
invented entities (3)
  • Latent unobserved process X′_i
    purpose: Represent missing occurrences under preferential sampling; enter parent sets so dependence uses full processes; enable counts/maps of unobserved points.
    Standard data-augmentation construct with scientific interpretation; not a new physical object.
  • Auxiliary process U_i
    purpose: Complete the thinning construction so intensities sum to λ*_i and likelihood becomes tractable.
    Purely computational latent field without direct ecological meaning.
  • Deterministic union nodes E and A (earthworks/ADE full realizations)
    purpose: Let species intensities depend on complete archaeological patterns in the Amazon graph (Fig. 2).
    Graphical convenience nodes equal to X∪X′; not independent entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Bayesian Joint Model for Multiple Point Processes with Application to Presence-Only Data." pith.science (2026). https://pith.science/paper/MUR555FL

@misc{pith2026260702786,
  author       = {Pith},
  title        = {Pith review of: A Bayesian Joint Model for Multiple Point Processes with Application to Presence-Only Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MUR555FL}},
  note         = {Machine review of arXiv:2607.02786}
}
read the original abstract

Joint modeling of multiple point processes is relevant in applications where relationships among processes are of interest, such as in ecological and archaeological studies. Statistical inference becomes particularly challenging when multiple processes are analyzed jointly and the observed data correspond to presence-only patterns, which are subject to preferential sampling and partial observability. This paper proposes a Bayesian joint model for multiple point processes, with application to the presence-only setting. The dependence between processes is explicitly incorporated into the probabilistic specification of the model using Bayesian networks. Direct use of the likelihood leads to intractable likelihood functions. Latent data processes are then introduced so that the augmented likelihood function becomes tractable and can be exactly evaluated. This formulation also enables direct inference on the number and the spatial distribution of unobserved occurrences of any of the point patterns. Inference is carried out using Markov chain Monte Carlo with blocked Gibbs sampling. Simulation studies demonstrate that the proposed inferential scheme is able to recover the true model parameters. The proposed model is applied to real presence-only data of archaeological sites and tree species from Amazonia, as part of the study of the effect that pre-Columbian Indigenous presence might have on the occurrences of relevant tree species. The results are consistent with the findings reported in the literature. They also illustrate how the proposed model enables inference on the existence and on the magnitude of the relation between processes, in addition to their association with environmental covariates.

Figures

Figures reproduced from arXiv: 2607.02786 by the authors.

Figure 1
Figure 1. Bivariate example illustrating the augmented graphical structure. (a) A graph with a single directed edge [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Application augmented graph. The deterministic nodes [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Presence-only occurrences across the Amazonian region. Panels (A), (B), and (C) display the locations of [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (a) Graph representing the modeled dependency structure between archaeological sites (ADEs and earthworks) [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Estimated response curves relating the presence probability in a 1 km [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Posterior predictive probability of occurrence of the latent processes [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 16 canonical work pages

  1. [1]

    doi:10.1111/j.0006-341X.1999.01051.x

    ISSN 0006-341X. doi:10.1111/j.0006-341X.1999.01051.x. URL https://doi.org/10. 1111/j.0006-341X.1999.01051.x. Christophe Botella, Alexis Joly, Pierre Bonnet, François Munoz, and Pascal Monestiez. Jointly estimating spatial sampling effort and habitat suitability for multiple species from opportunistic presence-only data.Methods in Ecology and Evolution, 12...

    work page doi:10.1111/j.0006-341x.1999.01051.x 1999

  2. [2]

    doi:10.1214/24-BA1460. W. D. Kissling, Carsten F. Dormann, Jürgen Groeneveld, Thomas Hickler, Ingolf Kühn, Greg J. McInerny, José M. Montoya, et al. Towards novel approaches to modelling biotic interactions in multispecies assemblages at large spatial extents.Journal of Biogeography, 39(12):2163–2178,

    work page doi:10.1214/24-ba1460

  3. [3]

    Rachel R

    doi:https://doi.org/10.1111/j.1365-2699.2011.02663.x. Rachel R. Fern, Michael L. Morrison, Hsiao-Hsuan Wang, William E. Grant, and Tyler A. Campbell. In- corporating biotic relationships improves species distribution models: Modeling the temporal influence of competition in conspecific nesting birds.Ecological Modelling, 408:108743,

    work page doi:10.1111/j.1365-2699.2011.02663.x 2011

  4. [4]

    doi:https://doi.org/10.1016/j.ecolmodel.2019.108743

    ISSN 0304-3800. doi:https://doi.org/10.1016/j.ecolmodel.2019.108743. A. M. Latimer, S. Banerjee, H. Sang Jr, E. S. Mosher, and J. A. Silander Jr. Hierarchical models facilitate spatial analysis of large data sets: a case study on invasive plant species in the northeastern United States.Ecology Letters, 12(2): 144–154,

    work page doi:10.1016/j.ecolmodel.2019.108743 2019

  5. [5]

    URL https://onlinelibrary.wiley

    doi:https://doi.org/10.1111/j.1461-0248.2008.01270.x. URL https://onlinelibrary.wiley. com/doi/abs/10.1111/j.1461-0248.2008.01270.x. 19 A Bayesian Joint Model for Multiple Point ProcessesA PREPRINT Laura J. Pollock, Reid Tingley, William K. Morris, Nick Golding, Robert B. O’Hara, Kirsten M. Parris, Peter A. Vesk, et al. Understanding co-occurrence by mode...

    doi:10.1111/j.1461-0248.2008.01270.x 2008

  6. [6]

    URL https://onlinelibrary

    doi:https://doi.org/10.1111/jbi.12540. URL https://onlinelibrary. wiley.com/doi/abs/10.1111/jbi.12540. C. Levis, F. R. C. Costa, F. Bongers, M. Peña-Claros, C. R. Clement, A. B. Junqueira, E. G. Neves, et al. Persistent effects of pre-Columbian plant domestication on Amazonian forest composition.Science, 355(6328):925–931,

    work page doi:10.1111/jbi.12540

  7. [7]

    Crystal N

    URLhttps://www.science.org/doi/abs/10.1126/science.aal0157. Crystal N. H. McMichael, Mark B. Bush, Hans ter Steege, Dolores R. Piperno, William D. Gosling, Majoi N. Nasci- mento, Umberto Lombardo, et al. Centuries of compounding human influence on Amazonian forests.Proceedings of the National Academy of Sciences, 122(47):e2514040122,

    work page doi:10.1126/science.aal0157

  8. [8]

    Vinicius Peripato, Carolina Levis, Guido A Moreira, Dani Gamerman, Hans Ter Steege, Nigel CA Pitman, Jonas G De Souza, et al

    doi:10.1073/pnas.2514040122. Vinicius Peripato, Carolina Levis, Guido A Moreira, Dani Gamerman, Hans Ter Steege, Nigel CA Pitman, Jonas G De Souza, et al. More than 10,000 pre-Columbian earthworks are still hidden throughout Amazonia.Science, 382 (6666):103–109,

    work page doi:10.1073/pnas.2514040122

Show all 19 references
  1. [9]

    doi:10.1080/24694452.2020.1843996. C. H. McMichael, M. W. Palace, M. B. Bush, B. Braswell, S. Hagen, E. G. Neves, M. R. Silman, et al. Predicting pre-Columbian anthropogenic soils in Amazonia.Proceedings of the Royal Society B: Biological Sciences, 281: 20132475,

    2020 doi

  2. [10]

    URLhttps://doi.org/10.1098/rspb.2013.2475

    doi:10.1098/rspb.2013.2475. URLhttps://doi.org/10.1098/rspb.2013.2475. R. Hills. Handroanthus serratifolius. The IUCN Red List of Threatened Species,

    2013 doi

  3. [11]

    e.T61985509A145677076

    URL https://dx.doi.org/ 10.2305/IUCN.UK.2021-3.RLTS.T61985509A145677076.en. e.T61985509A145677076. Americas Regional Workshop. Bertholletia excelsa. The IUCN Red List of Threatened Species,

    2021 doi

  4. [12]

    e.T32986A9741363

    URL https://dx.doi.org/10.2305/IUCN.UK.1998.RLTS.T32986A9741363.en. e.T32986A9741363. Robert S. Walker, John R. Ferguson, Ana Olmeda, Marcus J. Hamilton, Johan Elghammer, and Briggs Buchanan. Predicting the geographic distribution of ancient Amazonian archaeological sites with...

    1998 doi

  5. [13]

    Robert S Walker

    doi:10.7717/peerj.15137. Robert S Walker. Robertswalker/ancient_amazonia_archaeology: v1.0.0. Zenodo,

    doi

  6. [14]

    speciesLink Network

    Checklist dataset accessed via GBIF.org on 14 Dec 2025 forBertholletia excelsa,Dipteryx odorata, andHandroanthus serratifolius. speciesLink Network. speciesLink network,

    2025

  7. [15]

    Daiana C

    Accessed on 14 Dec 2025 viaspecieslink.net/search. Daiana C. M. Tourne, Maria V . R. Ballester, Patrick M. A. James, Lucieta G. Martorano, Marcelino Carneiro Guedes, and Evert Thomas. Strategies to optimize modeling habitat suitability of Bertholletia excelsa in the Pan-Amazon...

    2025

  8. [16]

    URL https://onlinelibrary.wiley.com/doi/abs/10.1002/ece3.5726

    doi:https://doi.org/10.1002/ece3.5726. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/ece3.5726. 20 A Bayesian Joint Model for Multiple Point ProcessesA PREPRINT Luciana Cristina Vitorino, Matheus S. Lima-Ribeiro, Levi Carina Terribile, and Rosane G. Collevatti. Demograph...

    doi

  9. [17]

    URL https://doi.org/10

    doi:10.1186/s12862-016-0779-9. URL https://doi.org/10. 1186/s12862-016-0779-9. Catarina S. Carvalho, Raquel Moura Machado, Maristerra R. Lemes, and Domingos Cardoso. Ecological niche modeling predicts alarming impacts of global climate change on economically important Neotropi...

    doi

  10. [18]

    URL https://onlinelibrary

    doi:https://doi.org/10.1002/ece3.72105. URL https://onlinelibrary. wiley.com/doi/abs/10.1002/ece3.72105. Jonas Gregorio de Souza, Denise Pahl Schaan, Mark Robinson, Antonia Damasceno Barbosa, Luiz E. O. C. Aragão, Ben Hur Marimon Jr., Beatriz Schwantes Marimon, et al. Pre-Colu...

    doi

  11. [19]

    Helaine Silverman and William H

    doi:10.1038/s41467-018-03510-7. Helaine Silverman and William H. Isbell, editors.Handbook of South American Archaeology. Springer, New York, NY ,

    doi

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.