{"id":"22deabcb-edf5-4888-8b69-c9b739d7778c","arxiv_id":"1908.02806","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Bayesian Markov model with Pólya-Gamma sampling estimates behavior transition probabilities from accelerometer classifications and finds habitat, diurnal, and weather effects in migrating geese.","lead":"This paper builds a Bayesian Markov model to estimate how often individual geese switch among flying, feeding, stationary, and walking behaviors, using accelerometer data with habitat and weather covariates. It shows that habitat, time of day, and weather change these transition rates, and it accounts for classification uncertainty with multiple imputation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) misstates the multinomial likelihood; the Pólya-Gamma sampler may target a different posterior.","rationale":"The reader's weakest assumption (miscalibrated random forest probabilities) is real but secondary. The deeper problem is that the paper's likelihood derivation appears internally inconsistent. The authors define the transition probabilities via a multinomial logit (Eq. 2) but then in Eq. (5) replace the multinomial likelihood by a product of binary logistic terms. This replacement is exact only for J=2; for J=4 (the geese application) it introduces an extra factor that depends on all covariates, so the MCMC in S.3 samples from a different target posterior. If this is correct, the central claim that the model estimates transition probabilities and reveals significant covariate effects is unsupported, regardless of imputation distribution. The paper should either provide a correct derivation (perhaps following Polson et al. 2013's actual multinomial augmentation) or the simulation test above. I recommend REJECT unless the authors can demonstrate that the implemented sampler targets the multinomial posterior.","tokens_in":24583,"tokens_out":16760,"duration_ms":164880,"concrete_test":"Run a simulation: generate N=10 individuals, T=1000 transitions each, from a first-order Markov model with J=3 states and covariates (e.g., one binary habitat) using the softmax transition probabilities in Eq. (2) with known β. Then fit the model exactly as described in S.3 (same Gibbs updates, same offset C). Check whether the 95% posterior intervals for β_ij cover the true values across many replicate simulations. If they do not, the sampler is not targeting the multinomial posterior. As a secondary check, compare the posterior distribution of the transition probabilities from the PG sampler with that from a generic MCMC (e.g., Stan) using the exact multinomial likelihood on the same data; if the two posteriors differ systematically, Eq. (5) is misspecified.","verdict_should_be":"REJECT","load_bearing_attack":"Equation (5) claims the multinomial likelihood for the Markov transition model is proportional to a product of binary logistic functions with η_{nijt} = ψ_{nijt} - log(Σ_{k≠j} e^{ψ_{nikt}}) and indicators y_{nijt}. For J>2 this identity is false. For a single transition with y_{ij*}=1, the product equals logistic(η_{ij*}) × Π_{j≠j*}(1 - logistic(η_{ij})), which for J=3 is p_{j*} × Π_{j≠j*} (Σ_{k≠j} e^{ψ_k})/D, not p_{j*} = e^{ψ_{j*}}/D. The ratio depends on ψ, so no constant of proportionality exists. The Gibbs updates in S.3 are exactly those of independent binary logistic regressions (one per non-reference category), so the sampler appears to target the product-of-Bernoulli posterior rather than the multinomial logistic posterior defined by Eq. (2). Consequently, all reported coefficients, odds ratios, and significance statements may be based on a misspecified likelihood. This is more load-bearing than the imputation-calibration issue: even with perfectly calibrated classification probabilities, inference is invalid if the likelihood is wrong.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage Bayesian framework for inferring animal behavior transition probabilities from accelerometer (ACC) data. A random forest classifier assigns four behavior states (flying, feeding, stationary, walking) from ACC features, and multiple imputation (M = 200 imputed data sets) is used to propagate classification uncertainty. A first-order Markov model with multinomial logistic transition probabilities relates transitions to habitat, weather, and time-of-day covariates, with Pólya-Gamma latent variables used for Gibbs sampling of the coefficients. The method is applied to six greater white-fronted geese during March 2018, and the authors report significant habitat, diurnal, and weather effects on transition rates, with odds-ratio summaries and pairwise habitat comparisons.","tokens_in":24825,"tokens_out":16365,"duration_ms":188404,"significance":"If the methodology were correct, the paper would make a useful contribution to movement ecology: it explicitly models temporal dependence in behavior classifications, propagates classifier uncertainty via multiple imputation, and offers a computationally convenient Pólya-Gamma sampler for a nonhomogeneous Markov model. The full conditional distributions are provided, and the supplementary material includes sensitivity analyses for the number of imputations and posterior predictive checks. However, the central statistical claim is undermined by an incorrect likelihood representation in Eq. (5) and by full conditionals in the supplementary material that implement a different model. The ecological conclusions, while plausible and of interest, are currently not supported by the reported inference.","major_comments":[{"comment":"Equation (5) is not a valid algebraic representation of the multinomial logistic likelihood for J > 2. For a single transition with observed category j*, the right-hand product is logistic(η_{j*}) × Π_{j≠j*} (1 − logistic(η_j)). For J = 3 with reference category 3, this equals e^{ψ_{j*}}(e^{ψ_{j*}}+1)/D^2, where D = 1 + e^{ψ_1} + e^{ψ_2}; the extra factor (e^{ψ_{j*}}+1)/D depends on the parameters and on which category was observed, so the right side is not proportional to the multinomial probability e^{ψ_{j*}}/D. Consequently, the posterior sampled by the proposed algorithm is not the posterior of the multinomial logistic model in Eq. (2). Because every reported coefficient, odds ratio, and significance statement in Tables 1–2 and Figures 2–3 is produced by this sampler, the central inference of the paper is currently unsupported.","section":"Model Fitting, Eq. (5)"},{"comment":"The full conditionals in S.3 are exactly the updates for J−1 independent binary logistic regressions (one for each non-reference category), with response Y_ij and offset C_nijt. In a correct Pólya-Gamma sampler for the softmax likelihood, the update for β_ij must also account for the dependence of the other categories' offsets C_nik (k ≠ j) on β_ij; those terms are absent here. Thus the Markov chain targets the product-of-Bernoulli posterior rather than the multinomial logistic posterior. The reference to Theorem 1 of Polson et al. (2013) does not justify this multinomial construction; the authors need to use a valid multinomial Pólya-Gamma representation (for example, logistic-softmax or multivariate Pólya-Gamma) and rerun all analyses.","section":"Supplementary Material S.3"},{"comment":"Equation (6) assumes that the imputation distribution [S|A] is exactly the random forest classification probability, but no evidence is provided that these probabilities are calibrated, and the assumption that behavior labels are conditionally independent in time given ACC data is stated but not checked. If the random forest probabilities are miscalibrated or the classification errors are temporally correlated, the imputed behavior sequences are drawn from the wrong distribution and the bias propagates directly into all transition estimates. The sensitivity analysis in S.6 only varies the number of imputations, not the correctness of the imputation distribution, so it cannot address this concern. A calibration check on held-out ACC fixes, or an alternative imputation model with posterior predictive calibration, is needed.","section":"Multiple Imputation, Eq. (6)"},{"comment":"Habitat and weather covariates for each ACC fix are taken from the most recent GPS fix, with GPS at 30-minute intervals and ACC at 6-minute intervals (Dataset section). Last-observation-carried-forward can misassign habitat during flights or short stops, and no sensitivity analysis is reported for this choice. Since the paper's main ecological conclusions concern habitat-specific transition rates and weather effects, this measurement-error issue is load-bearing; at minimum, a comparison with interpolated locations or an analysis restricted to GPS-aligned fixes is required.","section":"Dataset"}],"minor_comments":[{"comment":"The notation in the definition of C_nijt is garbled: the summation should be over k ≠ j, not over 'k, j'. Please correct the typesetting and make the notation consistent.","section":"Equation (5) and S.3"},{"comment":"The text describing the corn coefficient as a 'mean effect' of 0.42 should say 'posterior mean of the log-odds coefficient' or 'log odds ratio', rather than 'mean effect', to avoid ambiguity.","section":"Results, Table 1"},{"comment":"The main text says the only transition without significant habitat differences was flight to stationary, while S.5 states that 'grazing to flight' was the only transition without significant pairwise differences. These statements should be reconciled, and the transit","section":"Supplementary S.5 vs Results"},{"comment":"The caption says 'the upper triangular values and lower triangular values sum to 1'; this should clarify that corresponding row/column pairs (not the entire triangles) sum to 1.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (5) error is severe and invalidates the current numerical results. I recommend that the editor require a corrected multinomial Pólya-Gamma sampler and a complete reanalysis before any further consideration; if the authors cannot supply a valid implementation, the paper should be rejected. The current version's estimates should not be cited as evidence for the ecological conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the two things you need to know: the stress-test note is right, and it is more important than the imputation-calibration worry. Equation (5) claims a multinomial logistic likelihood is proportional to a product of binary logistic terms with η_{nijt} = ψ_{nijt} - log(Σ_{k≠j} e^{ψ_{nikt}}). For J > 2 that is false. The extra product over j≠j* contributes a factor Σ_{k≠j} e^{ψ_k}/D that depends on the same linear predictors, so there is no constant of proportionality. The Gibbs updates in S.3 are exactly the independent-binary-logistic updates, so the sampler is targeting the product-of-Bernoulli posterior, not the transition model defined by Eq. (2). Every odds ratio and significance statement is then about a different model. This is load-bearing: even with perfect classification probabilities, the inference is not for the stated multinomial Markov model.\n\nWhat the paper does well: the pipeline of supervised ACC classification + multiple imputation + a covariate-driven transition model is a sensible and clearly motivated idea for accelerometer data. The sensitivity analysis across 200, 100, and most-likely imputations is honest and useful. The writing is readable, and the ecological findings (feeding in crops, stationary on water at night) are consistent with prior goose biology.\n\nOther soft spots are secondary. Treating random forest probabilities as the true [S|A] is an assumption that deserves sensitivity testing, but the likelihood error dwarfs it. Carrying weather/habitat from the most recent GPS fix introduces timing error at the scale of tens of minutes, which is probably minor for these covariates.\n\nRecommendation: I would not cite this version, and the paper should not be accepted as is. But it deserves a serious referee rather than a desk reject: the flaw is subtle, specific, and fixable. A correct multinomial Pólya-Gamma sampler exists in the literature, or the authors could honestly reframe the model as independent binary regressions with an offset and adjust the interpretation. The application niche is real.","headline":"The ecological application is thoughtful, but the multinomial Pólya-Gamma sampler in Eq. (5) is invalid for four states, so the reported odds ratios likely do not correspond to the stated transition model.","tokens_in":25353,"tokens_out":4168,"would_cite":false,"duration_ms":47443,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62M05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Bayesian Markov model with Pólya-Gamma sampling estimates how habitat, time of day, and weather drive behavior transitions in accelerometer-classified geese.","keywords":["animal behavior","auxiliary variables","hierarchical models","multinomial logistic","multiple imputation","Pólya-Gamma","Markov model","accelerometer classification"],"falsifier":"Take the video-verified training behaviors, fit the same Markov model once to the true labels and once to sequences imputed from the random forest for the same ACC fixes, and compare posterior transition coefficients; if the imputation-based estimates shift systematically away from the truth-labeled estimates, the assumed $[S|A]$ is miscalibrated and the reported habitat, weather, and diurnal effects are not trustworthy. A simpler version: recalibrate the random forest probabilities on a validation set and check whether any odds-ratio posterior moves by more than its credible interval.","tokens_in":24415,"feed_emoji":"🦢","tokens_out":7547,"duration_ms":67755,"temperature":0.7,"pith_summary":"This paper proposes a two-stage Bayesian framework that turns accelerometer-based behavior classifications into estimates of how often individual animals switch among flying, feeding, stationary, and walking, and how those switch rates depend on habitat, time of day, and weather. The authors apply the model to six greater white-fronted geese during spring migration and report significant differences in transition rates among habitat types, between day and night, and under different weather conditions. If the approach works as claimed, it gives ecologists a way to study behavioral decisions at the native sampling scale of the data while accounting for classification uncertainty, which activity-budget summaries and resource selection models do not do.","feed_headline":"Markov model links goose behavior switches to habitat and weather","feed_subtitle":"Pólya-Gamma sampling and multiple imputation turn accelerometer labels into uncertainty-aware transition probabilities.","key_machinery":"The central object is a time-varying first-order Markov transition matrix $P_{nt}$ over four behaviors, with each row modeled by a multinomial logistic link, $\\psi_{nijt}=\\log(p_{nijt}/p_{niJt})=x_{nt}'\\beta_{ij}$, where walking is the reference category. Two augmentations carry the argument: Pólya-Gamma latent variables, which turn the multinomial logistic likelihood into conditionally conjugate Gaussian updates for the coefficients $\\beta_{ij}$, and multiple imputation, which draws $M=200$ behavior sequences from the random forest's classification probabilities $[S|A]$ and averages the Markov-model posterior over them. The machinery yields automatic Gibbs sampling without tuning and odds-ratio summaries for every covariate effect on each transition.","core_discovery":"The central claim is that behavior transition probabilities can be estimated directly from classified accelerometer data with a first-order Markov model whose transition probabilities follow a multinomial logistic link, with Pólya-Gamma latent variables making posterior sampling automatic and with multiple imputation carrying the random forest classifier's uncertainty into the posterior. For six greater white-fronted geese in March 2018, the model finds significant habitat effects on nearly every transition probability, a strong diurnal cycle (increased stationary behavior at night, reduced flight and feeding), and weather effects concentrated on flight and feeding transitions, including opposing effects of daily minimum and maximum temperature on the probability of continuing to fly. The authors present the framework as a unifying way to use both acceleration and GPS data for behavioral inference.","pith_inferences":["A natural extension the authors only gesture at is sharing coefficients across transitions for covariates like time of day, which appeared to act similarly across all from-states; a behavior-specific formulation would reduce parameter count and sharpen inference.","Habitat and weather values are assigned to each ACC fix from the most recent GPS location, so a 30-minute GPS gap is mapped onto a 6-minute behavior scale; interpolating locations to ACC times would test whether the reported habitat effects are robust to this timing assumption.","The sensitivity comparison suggests the direction of effects survives even using a single most-likely classification, implying the substantive conclusions may not hinge on the imputation count; a formal diagnostic for the number of imputations would strengthen the method.","The paper notes the same augmentation can enter hidden Markov models; that route would let movement ecologists replace tuned HMM samplers with Pólya-Gamma updates when behavioral states are not directly observed."],"forward_implications":["Habitat effects on behavior can be estimated without defining an availability distribution, so habitats used at similar rates can still be distinguished by the behaviors performed in them.","Classification uncertainty is propagated: compared with using only the most likely behavior label, multiple imputation shrinks coefficient estimates toward zero and keeps credible intervals at least as wide.","The Pólya-Gamma scheme removes the need to tune Metropolis-Hastings proposals for multinomial logistic transition models, making similar Bayesian behavior models easier to fit for other species.","Transition-specific coefficients can expose asymmetric responses, such as weather changing the probability of remaining in flight but not the probability of leaving stationary behavior.","Fitted transition matrices support posterior predictive simulation of behavior sequences, allowing comparisons of time allocation across habitats or weather scenarios."],"supporting_citations":[{"why":"Provides the Pólya-Gamma data augmentation that makes Gibbs sampling of the logistic regression coefficients automatic.","marker":"Polson et al. (2013)"},{"why":"Supplies the multiple-imputation MCMC algorithm the paper follows to marginalize over classifier uncertainty.","marker":"Scharf et al. (2017)"},{"why":"Defines the 52 acceleration features and supervised classification workflow used to build the random forest behavior classifier.","marker":"Resheff et al. (2014)"},{"why":"Shows how to express multinomial logistic likelihoods as products of binary logistic functions, the transformation the Pólya-Gamma sampler exploits.","marker":"Holmes and Held (2006)"},{"why":"Demonstrates Pólya-Gamma sampling for transition probabilities in a Markov model, the template the paper adapts to animal behavior.","marker":"Holsclaw et al. (2017)"},{"why":"Provides the multiple-imputation framework for incorporating auxiliary data uncertainty into Markov models of animal movement.","marker":"McClintock (2017)"},{"why":"Establishes that Markov models improve time-allocation estimates relative to traditional activity budgets, motivating the transition-based approach.","marker":"Rugg and Buech (1990)"},{"why":"Supplies the ecological background that geese use different habitats for different behaviors, which the habitat comparisons are designed to test.","marker":"Krapu et al. (1995)"}],"fun_headline_variants":["Goose behavior switches driven by habitat, time, and weather","Bayesian model links goose transitions to habitat and weather","Pólya-Gamma sampling sharpens goose behavior transition estimates","Accelerometer data show habitat and weather shift goose behavior","Markov model quantifies goose behavior changes with environment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the random forest's predicted class probabilities give the true conditional distribution of behavior given the acceleration data, because behavior sequences imputed from those probabilities are then treated as the data for the Markov model; any miscalibration in the classifier flows directly into every transition estimate.","fun_headline_variants_meta":{"raw":{"variants":["Goose behavior switches driven by habitat, time, and weather","Bayesian model links goose transitions to habitat and weather","Pólya-Gamma sampling sharpens goose behavior transition estimates","Accelerometer data show habitat and weather shift goose behavior","Markov model quantifies goose behavior changes with environment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000705,"raw_usage":{"total_tokens":3156,"prompt_tokens":897,"completion_tokens":2259,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":2178}},"tokens_in":513,"tokens_out":2259,"duration_ms":16959,"temperature":1.0,"reasoning_tokens":2178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:33:25.786835+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the video-verified training behaviors, fit the same Markov model once to the true labels and once to sequences imputed from the random forest for the same ACC fixes, and compare posterior transition coefficients; if the imputation-based estimates shift systematically away from the truth-labeled estimates, the assumed $[S|A]$ is miscalibrated and the reported habitat, weather, and diurnal effects are not trustworthy. A simpler version: recalibrate the random forest probabilities on a validation set and check whether any odds-ratio posterior moves by more than its credible interval.","supporting_citations":[{"cited_title":"G., Scott, J","cited_arxiv_id":null,"evidence_quote":"Provides the Pólya-Gamma data augmentation that makes Gibbs sampling of the logistic regression coefficients automatic."},{"cited_title":"B., and Johnson, D","cited_arxiv_id":null,"evidence_quote":"Supplies the multiple-imputation MCMC algorithm the paper follows to marginalize over classifier uncertainty."},{"cited_title":"S., Rotics, S., Harel, R., Spiegel, O., and Nathan, R","cited_arxiv_id":null,"evidence_quote":"Defines the 52 acceleration features and supervised classification workflow used to build the random forest behavior classifier."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows how to express multinomial logistic likelihoods as products of binary logistic functions, the transformation the Pólya-Gamma sampler exploits."},{"cited_title":"M., Robertson, A","cited_arxiv_id":null,"evidence_quote":"Demonstrates Pólya-Gamma sampling for transition probabilities in a Markov model, the template the paper adapts to animal behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that Markov models improve time-allocation estimates relative to traditional activity budgets, motivating the transition-based approach."},{"cited_title":"L., Reinecke, K","cited_arxiv_id":null,"evidence_quote":"Supplies the ecological background that geese use different habitats for different behaviors, which the habitat comparisons are designed to test."}],"review_version":1}