{"id":"a3fedac6-e58a-4d51-86cd-e83b39ee6b7f","arxiv_id":"1908.09410","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Pairwise odds ratios, and their spatial surfaces, are proposed as interpretable summaries of species dependence in joint species distribution models, where latent correlations are argued to be uninformative about co-occurrence strength.","lead":"This paper argues that the correlations used in joint species distribution models tell ecologists little about how species actually co-occur, and proposes odds ratios, including odds-ratio maps across a landscape, as a clearer alternative. It demonstrates the idea on data from South Africa's Cape Floristic Region with 639 plant species at 662 sites.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Odds-ratio surfaces inherit the fitted spatial JSDM's low-rank/common-decay assumptions; without a sensitivity check, the empirical illustration is not secure.","rationale":"The reader's weakest assumption is that the empirical odds-ratio surfaces are meaningful only if the fitted spatial JSDM is an adequate generative model for the presence/absence data. I agree, and I sharpen the concern by locating it in the explicit common-decay/low-rank modeling choice in Section 5.1. The central mathematical contribution, the monotonicity theorem in the Appendix, is sound: Slepian's theorem gives monotonicity of the bivariate normal CDF in the correlation, and the corollary about signs follows. Therefore the main methodological point is not threatened by this concern. What is threatened is the empirical demonstration: the surfaces are presented as illustrating how dependence varies over the region, but under the fitted model the within-site latent correlation is constant, and the surface variation comes from covariate-driven means. If the common spatial decay assumption is wrong, the estimated factor structure can be biased, and the post hoc selection of five species with correlations in a wide range makes the illustration vulnerable to cherry-picking. The paper itself admits the data analysis is not the focus, but it still uses the CFR surfaces as evidence for the utility of the proposal. A sensitivity analysis with more flexible spatial ranges or a different number of factors would settle whether the surfaces are robust. Because this is an addressable empirical concern rather than a flaw in the central argument, the reader's CONDITIONAL verdict remains appropriate; no verdict change is needed.","tokens_in":14406,"tokens_out":14553,"duration_ms":166108,"concrete_test":"Refit the Cape Floristic Region data with the same dimension reduction but allow r distinct spatial decay parameters for the latent factor processes, and separately try r = 5 and r = 10 factors. Recompute the ten log-odds-ratio surfaces for the five displayed species. If the surfaces change qualitatively, e.g., sign flips or magnitude changes by more than a factor of 2 in data-rich regions, the common-decay/low-rank assumption is load-bearing. In addition, run a posterior predictive check of pairwise co-occurrence frequencies; if the fitted JSDM systematically under- or over-predicts observed co-occurrence, the surfaces are not externally supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Appendix theorem is correct and supports the methodological claim that latent correlations are poorly interpretable for dependence strength. The load-bearing weak point is the empirical demonstration. Section 5.1 assumes a dimension-reduced factor model with a common exponential decay parameter for all latent factor processes, and asserts that species-specific spatial ranges are 'expected to be negligible' without supporting evidence. Under this assumption, the within-site species covariance is constant across the region, so the reported log-odds-ratio surfaces vary only through covariate-driven means. If true species-specific spatial ranges differ, estimated factor loadings and implied pairwise correlations can be distorted. Moreover, the five displayed species are selected post hoc (Section 6), and no uncertainty is shown for the surfaces. Thus the dramatic features, including log odds ratios as low as -26.1, may reflect modeling artifacts rather than ecological dependence. This does not refute the theorem, but it makes the practical value of the proposed spatial odds-ratio surface conditional on an unverified adequacy assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that in joint species distribution models (JSDMs) for presence/absence data, the latent multivariate normal correlation matrix is not a reliable guide to pairwise species dependence. Instead, the authors advocate reporting the induced odds ratios and joint occurrence probabilities, which have direct ecological interpretation. The central technical contribution is an appendix theorem, based on Slepian's theorem, showing that for fixed species means the odds ratio is non-decreasing in the latent correlation, equals 1 at zero correlation, and has the same sign as the correlation, yet its magnitude is strongly controlled by the mean occurrence probabilities. The paper proposes a spatial log-odds-ratio surface obtained from a spatial JSDM and illustrates it on Cape Floristic Region data for five selected species. It also discusses implications for species richness and sketches an extension to ordinal abundance data.","tokens_in":14548,"tokens_out":8578,"duration_ms":83384,"significance":"If the claims hold, the paper provides a useful corrective to the widespread practice of interpreting latent species correlations in JSDMs, and the proposed odds-ratio surface is a practical tool for ecologists. The appendix theorem is correct and cleanly supports the central interpretational message, and Figure 1 effectively demonstrates that latent correlation alone carries little information about dependence strength. The spatial odds-ratio surface is a constructive proposal that goes beyond merely criticizing the correlation view. However, the empirical demonstration is conditional on the adequacy of the fitted spatial JSDM, and one ancillary claim about maximum likelihood and expected richness is stated without proof. The paper is clearly written and the theorem is correctly derived.","major_comments":[{"comment":"The sentence \"omitting details, maximum likelihood estimation will produce expectations that agree\" is an unsupported assertion that is load-bearing for technical point (i) in the Introduction. Under the JSDM specification in Section 4, the marginal probability of presence is not generally equal to Phi(L^F); after integrating out the random effects it becomes E[Phi(L^F + L^R)], which depends on the random-effect variance. Unless the model is parameterized so that the marginal probabilities match those of the independent probit model, the expected richness under the joint and independent maximum likelihood fits need not agree. Please provide a proof or an explicit condition, qualify the claim, or remove it.","section":"Section 3"},{"comment":"The empirical odds-ratio surfaces are computed under a fitted model that assumes a common exponential spatial decay parameter for all latent factor processes. The paper asserts that the implications of this separable-model simplification \"are expected to be negligible\" (Section 5.1) but provides no supporting evidence or sensitivity analysis. Because the within-site species covariance is constant under this assumption, the reported log-odds-ratio surfaces in Figures 4 and 5 vary only through covariate-driven means; if true species-specific spatial ranges differ materially, the estimated factor loadings and implied pairwise correlations could be distorted. Please add a sensitivity check (e.g., fitting with r decay parameters, or comparing a subset of species under species-specific ranges) or explicitly temper the ecological conclusions drawn from the figures.","section":"Section 5.1 and Section 6"},{"comment":"The five species displayed in Figures 4 and 5 are selected post hoc to span a range of posterior correlations, and the surfaces are shown only as posterior means. Section 4.1 explains how to obtain posterior distributions for the odds ratios, yet no credible intervals or posterior probabilities of sign are reported. As a result, the qualitative claims in Section 6 (e.g., \"large positive log theta(s) is observed in the western part\") and the extreme value of -26.1 cannot be assessed for statistical precision or robustness to the species-selection rule. Please provide uncertainty summaries or at least explicitly note that these are illustrative posterior means without quantified uncertainty.","section":"Section 6"}],"minor_comments":[{"comment":"Typo: \"no direction connection\" should be \"no direct connection\".","section":"Section 4.1"},{"comment":"The in-text citation \"Wilkinson2018\" should be formatted as \"Wilkinson et al. (2018)\".","section":"Section 1"},{"comment":"In the allopatry paragraph, the sentence \"Switching the species, we have p11/p1. < p.1\" appears to repeat the previous inequality; it should presumably read \"p11/p.1 < p1.\".","section":"Section 2"},{"comment":"The captions use \"parenthesis\" where \"parentheses\" is meant.","section":"Figures 4 and 5"},{"comment":"The display of the cumulative odds ratio definition is typeset ambiguously; the denominator should be a ratio of conditional probabilities for category k+1, not a product as it currently appears.","section":"Section 7"},{"comment":"The paper states that log odds ratios are on the base-10 scale in Section 2, but Figures 4 and 5 do not specify the base; please state which logarithm is used in the figures so that the reported values (e.g., -26.1) are interpretable.","section":"Section 2 and Figures 4/5"}],"recommendation":"major_revision","confidential_remarks":"The central theorem is correct and the paper makes a worthwhile methodological point. The main risks are the unsupported expected-richness claim in Section 3 and the unverified model-adequacy assumptions behind the empirical surfaces. Both are fixable with additional analysis or tempering, so the paper is suitable for major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core claim holds up. The paper argues that the latent correlations in JSDMs are a poor handle on pairwise species dependence, and that induced odds ratios (and joint occurrence probabilities) carry the interpretable information. That argument is correct as far as it goes. The appendix theorem, that the odds ratio is non-decreasing in the latent correlation with sign matching, is a clean application of Slepian's theorem, and the corollary about sign is immediate. Figure 1 makes the stronger point well: for a fixed positive rho you can get log odds ratios anywhere from near zero to very large, depending on the marginal means. That is the real message, and it is useful.\n\nWhat's new: applying the odds-ratio view to spatial JSDMs, producing odds-ratio surfaces over the region, and framing the interpretation problem in this way. The extension to ordinal abundance classes is a sensible pointer, though not worked out. The paper also clears up a muddle between functional and conditional specifications of probit JSDMs.\n\nSoft spots: the empirical section is the weakest part, but it is mainly illustrative. The model from Shirota et al. (2019) assumes a common spatial decay for all latent factor processes, and the paper hand-waves that species-specific ranges are 'expected to be negligible.' That deserves a sensitivity check or at least a caveat. The five species are chosen post hoc to show a range of correlations, and the surfaces are shown only as posterior means, with no uncertainty. The extreme log odds ratios like -26.1 are plausibly artifacts of the model's low-rank structure. Also, Section 3 asserts without proof that maximum likelihood produces equal expected richness under independent and joint models; that is a side comment, but if it stays it needs a reference or a derivation. No code or data are provided, which limits reproducibility, though the paper is explicitly methodological.\n\nNone of this undermines the central claim. If the goal is to change how ecologists report dependence from JSDMs, the paper makes a solid case. The limitations are addressable in revision.\n\nWho should read it: ecologists who fit JSDMs and wonder what to do with the latent correlation matrix; statisticians working on multivariate binary spatial models. I would take it to a reading group and would cite it as the place that makes the odds-ratio interpretation explicit.\n\nRecommendation: send to peer review. The core argument is correct and the methodological suggestion is practical. The empirical illustration needs to be either strengthened or explicitly framed as a demonstration, and the common-decay assumption should be sensitivity-checked.","headline":"The core argument about latent correlations being uninformative for pairwise species dependence is correct and useful; the empirical odds-ratio surfaces are the soft spot but are illustrative rather than load-bearing.","tokens_in":15129,"tokens_out":3241,"would_cite":true,"duration_ms":28122,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H20","62M30","62P12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that latent species correlations in joint species distribution models tell ecologists little about co-occurrence; the interpretable dependence lives in the induced odds ratios and their spatial surfaces.","keywords":["joint species distribution model","odds ratio","presence-absence data","latent variables","spatial dependence","Gaussian process","species co-occurrence","Cape Floristic Region"],"falsifier":"Numerical integration of the bivariate normal probabilities in (3) at fixed means and increasing $\\rho$ would immediately falsify the theorem if it were not monotone; for the applied claim, a test is to fit the spatial JSDM and compare model-based odds-ratio surfaces against smoothed empirical odds ratios from the same sites: systematic sign reversals would indict the probit-threshold model, not the odds-ratio summary.","tokens_in":14144,"feed_emoji":"🌿","tokens_out":8820,"duration_ms":81236,"temperature":0.7,"pith_summary":"This paper tries to reset what ecologists should look at when fitting a joint species distribution model (JSDM). It argues that the latent correlation matrix, the usual summary of species dependence in these models, carries almost no usable information about whether two species actually tend to occur together at a site. The interpretable quantities are the odds ratio and the joint occurrence and joint absence probabilities that the fitted model induces for each pair of species at each site. For spatially explicit JSDMs, the paper converts those pair-specific summaries into surfaces over the study region, showing how dependence changes from place to place. A supporting theorem makes the point precise: for fixed marginal probabilities, the induced odds ratio is a monotone function of the latent correlation, but its magnitude is controlled by the species mean probabilities, so the correlation alone cannot indicate dependence strength.","feed_headline":"Latent correlations hide co-occurrence; odds ratios tell the story","feed_subtitle":"Dependence lives in a pair's odds ratio, and that ratio shifts across the landscape.","key_machinery":"The load-bearing object is the odds ratio for the $2 \\times 2$ table of joint presence/absence, $\\theta(s) = p_{00}(s)p_{11}(s)/(p_{10}(s)p_{01}(s))$ for a pair of species at location $s$, computed from probabilities induced by the latent model rather than modeled directly. In the spatial JSDM, presence is obtained by clipping a latent Gaussian process, $Y_j(s) = 1(Z_j(s) > 0)$, with $Z_j(s) = B x(s) + \\Lambda W(s) + \\epsilon(s)$, where a low-dimensional factor $W(s)$ follows independent Gaussian processes with a common exponential covariance; this is the dimension-reduced spatial specification that makes an $S$-by-$S$ covariance tractable and produces smooth probability surfaces. The appendix's monotonicity theorem, built on a classical bivariate-normal probability result, is what licenses the claim that latent correlation and odds ratio cannot have contradictory signs while also showing why the correlation magnitude is uninformative.","core_discovery":"The paper's central claim is that dependence in a JSDM should be reported and interpreted through the $2 \\times 2$ table of joint presence and absence for each species pair, not through the latent multivariate-normal correlation. For a pair of species at a site, let $\\theta = p_{11}p_{00}/(p_{10}p_{01})$ be the odds ratio induced by thresholding the latent bivariate normal. The paper proves that with the latent means fixed, $\\theta$ is non-decreasing in the latent correlation $\\rho$, equals $1$ at $\\rho = 0$, and the sign of $\\log \\theta$ matches the sign of $\\rho$. Yet the magnitude of $\\theta$ can be arbitrarily large or small for a fixed $\\rho$ as the species means vary, and the joint occurrence probability can be large even with negative correlation. Consequently, the author argues, the $S \\times S$ latent correlation matrix is a device for generating dependence but not an interpretable statement about it. The paper demonstrates the alternative by fitting a spatial JSDM to 639 tree species at 662 Cape Floristic Region sites and presenting posterior mean surfaces of $\\log \\theta$ for selected pairs, where positive-correlation pairs show regions of positive log odds ratios and negative-correlation pairs show regions of negative log odds ratios.","pith_inferences":["Editorial inference: the monotonicity theorem transfers to any multivariate probit model, so disciplines that summarize binary-item dependence with tetrachoric correlations could adopt the same odds-ratio reporting.","Editorial inference: a direct diagnostic for practitioners is to plot posterior odds-ratio surfaces against latent correlations; pairs with similar correlations but different prevalences should show visibly different dependence if the paper's caution is right.","Editorial inference: a testable extension is to standardize the odds-ratio surface by each pair's marginal prevalence, separating rarity-driven absence from genuine avoidance."],"forward_implications":["Ecologists can replace the uninterpretable latent correlation matrix with maps of pairwise odds ratios, so that positive values mean co-occurrence or co-absence is encouraged and negative values mean it is discouraged.","Because the odds ratio depends on the site-level mean probabilities, a single estimated correlation does not describe a species pair; the appropriate summary is a surface or a set of site-specific values.","Spatial JSDM fitting smooths and interpolates these odds-ratio surfaces, so dependence at unsurveyed locations can be predicted.","For ordinal abundance classifications, local, global, and cumulative odds ratios extend the same interpretation to abundance dependence; the paper notes that non-spatial JSDMs with counts could be handled this way.","Stacked species distribution models, which assume independence, cannot represent sympatry or allopatry and will misstate the uncertainty in species richness even when expected richness agrees."],"supporting_citations":[{"why":"Supplies the monotonicity of bivariate normal rectangle probabilities in the correlation, which underpins the theorem that the odds ratio is non-decreasing in the latent correlation.","marker":"Slepian (1962)"},{"why":"Cited alongside Slepian for the multivariate normal probability-integral results used in the appendix.","marker":"Gupta (1963)"},{"why":"Supplies the spatial JSDM, the dimension-reduced Dirichlet-process factor specification, and the fitted Cape Floristic Region model from which the odds-ratio surfaces are drawn.","marker":"Shirota et al. (2019)"},{"why":"Provides the dimension-reduced factor approximation to the species covariance matrix that the spatial model builds on.","marker":"Taylor-Rodríguez et al. (2017)"},{"why":"Establishes the JSDM framework and the observation that latent correlations are only a device for creating dependence, the position the paper develops.","marker":"Clark et al. (2017)"},{"why":"Provides the latent-variable JSDM formulation with fixed and random effects whose pairwise correlations are under scrutiny.","marker":"Ovaskainen et al. (2016)"},{"why":"Supplies the odds-ratio machinery, including local, global, and cumulative odds ratios for the ordinal abundance extension.","marker":"Agresti (2012)"}],"fun_headline_variants":["Odds ratios reveal what latent correlations obscure","Species dependence lives in odds ratios, not correlations","Why latent correlations mislead JSDM ecologists","Dependence maps: odds ratios over correlations","See species co-occurrence through odds-ratio surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fitted probit-threshold latent-Gaussian model with a common spatial-decay parameter is a faithful generative description of the presence/absence data, so the odds-ratio surfaces it produces describe real ecological dependence rather than artifacts of the model.","fun_headline_variants_meta":{"raw":{"variants":["Odds ratios reveal what latent correlations obscure","Species dependence lives in odds ratios, not correlations","Why latent correlations mislead JSDM ecologists","Dependence maps: odds ratios over correlations","See species co-occurrence through odds-ratio surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1428,"prompt_tokens":1035,"completion_tokens":393,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":322}},"tokens_in":651,"tokens_out":393,"duration_ms":3983,"temperature":1.0,"reasoning_tokens":322,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:13:17.559907+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerical integration of the bivariate normal probabilities in (3) at fixed means and increasing $\\rho$ would immediately falsify the theorem if it were not monotone; for the applied claim, a test is to fit the spatial JSDM and compare model-based odds-ratio surfaces against smoothed empirical odds ratios from the same sites: systematic sign reversals would indict the probit-threshold model, not the odds-ratio summary.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the monotonicity of bivariate normal rectangle probabilities in the correlation, which underpins the theorem that the odds ratio is non-decreasing in the latent correlation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited alongside Slepian for the multivariate normal probability-integral results used in the appendix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the spatial JSDM, the dimension-reduced Dirichlet-process factor specification, and the fitted Cape Floristic Region model from which the odds-ratio surfaces are drawn."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the JSDM framework and the observation that latent correlations are only a device for creating dependence, the position the paper develops."},{"cited_title":"Abrego, P","cited_arxiv_id":null,"evidence_quote":"Provides the latent-variable JSDM formulation with fixed and random effects whose pairwise correlations are under scrutiny."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the odds-ratio machinery, including local, global, and cumulative odds ratios for the ordinal abundance extension."}],"review_version":1}