{"id":"f83b5737-ca82-4cf0-b72a-b303e40a703b","arxiv_id":"2509.01871","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"The inferred epidemic transmission tree changes depending on whether sampling times, location, or genetic distance data are used, with genetic data implying fewer and location data implying more unsampled hosts.","lead":"A new Monte Carlo method combines sampling times, locations, and viral genome distances to infer who infected whom in an outbreak. Applied to early COVID-19 cases in Australia, it shows that the type of data used changes the inferred transmission tree, including the estimated number of infected people never tested.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6)'s genetic likelihood is not a normalized probability and Eq. (7)'s location factor is unspecified; the reported data-type dependence of NU may be an artifact of these pseudo-likelihood components rather than of the data.","rationale":"The reader's weakest assumption already identifies the pseudo-likelihood factors as the main risk, and I agree that this is the load-bearing point. The paper has real independent support: the code is released, and the MCMC sampler is checked against an exactly solvable synthetic case (Sec. IV.C). However, that check only verifies that the sampler converges to Eq. (8) as coded; it does not validate Eq. (8) as a correct posterior for real transmission data. The printed genetic likelihood in Eq. (6) is not a normalized probability mass function, and the location likelihood in Eq. (7) is left unspecified. These are not merely simplifying assumptions; they can change the relative weights of trees and thereby produce the observed NU shifts. The reported inconsistency between NU=17.55 and its own 80% CI [5,14], and the absence of an effective-sample-size report, further reduce confidence, but the pseudo-likelihood issue is primary. Because the problem is addressable by re-running with corrected likelihoods, the reader's CONDITIONAL verdict is appropriate. I recommend no change to that verdict. If the corrected rerun reproduced the same ordering, the claim would be substantially strengthened; if it did not, the central claim should be rejected.","tokens_in":13646,"tokens_out":8616,"duration_ms":102196,"concrete_test":"","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Sec. V.C and Conclusions) is that adding genetic/location data changes NU and other tree statistics. This comparison is meaningful only if the MCMC targets the posterior P(T|D) in Eq. (8). The genetic and location factors in that posterior are not valid normalized probabilities. Eq. (6) gives P_genetic(j'→i') = μΔt e^{-μΔt d_{i',j'}}. If d is an integer SNP count, this is not a pmf: summing over d ≥ 0 yields μΔt/(1-e^{-μΔt}) ≠ 1, and the Poisson factorial is absent; if d is treated as continuous, it is an exponential density, not a probability. The omitted normalization depends on Δt, hence on the tree, so the Metropolis-Hastings ratio in Eq. (9) does not sample from the stated posterior. Eq. (7) similarly introduces a constant A 'fixed by normalization' without specifying the normalization, and both factors use a heuristic 'closest non-downstream sampled host' (Secs. III.D, III.E) rather than the actual infector, so they are not derived from a generative model of sequence evolution or movement. Because the reported differences in NU (17.55 → 14.28, 25.68) are exactly the quantities affected by these factors, the direction and magnitude of the data-type effect could be an artifact of the pseudo-likelihood. The paper's own conclusions acknowledge the models are 'simplistic,' but the issue here is not simplicity: the posterior target is not properly normalized.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a generative model and MCMC sampler to infer transmission trees from sampling times, genetic distances, and locations, and applies it to NSW COVID-19 data. The central claim is that network properties such as the number of unsampled hosts depend sensitively on which data types are used.","tokens_in":14011,"tokens_out":3817,"duration_ms":35709,"significance":"If the central claim holds, the paper provides a practical tool for designing surveillance strategies. The authors provide code and data, and the MCMC convergence checks against analytically computed probabilities in a small synthetic setting are a strength. However, the validity of the posterior target is undermined by pseudo-likelihood factors that are not properly normalized, so the data-type dependence may be an artifact.","major_comments":[{"comment":"The genetic likelihood P_genetic(j'→i') = μΔt e^{-μΔt d} is not a valid probability mass function over SNP count d. Summing over d≥0 gives μΔt/(1-e^{-μΔt}) ≠ 1, and the factorial term of a Poisson is missing. If d is treated as continuous, it is an exponential density, not a probability. The normalization depends on Δt, hence on the tree, so the Metropolis-Hastings ratio in Eq. (9) does not sample from the stated posterior in Eq. (8). Since the comparison of NU across data types is the paper's central claim, this is a load-bearing issue.","section":"Eq. (6), Sec. III.D"},{"comment":"The location factor P_location = A if same location, and A(1-e^{-Δt/τ}) otherwise. The constant A is said to be 'fixed by normalization' but no normalization is specified. The factor does not sum to a constant over the binary outcome space; its normalization would depend on Δt and on the tree structure. Consequently, the posterior P(T|D) is not a well-defined probability distribution, and relative tree probabilities are not meaningful.","section":"Eq. (7), Sec. III.E"},{"comment":"The genetic and location factors are computed with respect to the closest sampled host that is not downstream of i, rather than the actual infector. These are heuristic pairwise-distance pseudo-likelihoods, not derived from a generative model of sequence evolution or host movement. The MCMC validation in Sec. IV.C computes 'theoretical' probabilities by integrating Eq. (8) numerically, but since Eq. (8) contains the same unnormalized pseudo-likelihoods, the validation only confirms that the sampler targets the stated (incorrect) density, not that the density is a correct posterior. The central claim about sensitivity to data type therefore rests on an unvalidated target distribution.","section":"Secs. III.D, III.E, Eq. (8)"},{"comment":"The paper reports 'NU = 17.55 – with 80% confidence interval CI 80% = [5, 14]' when only sampling time is used. A point estimate cannot lie outside its own 80% CI. This indicates a computational or reporting error in the estimation of NU or the CI. This is a concrete internal inconsistency that must be corrected.","section":"Conclusions, last paragraph"}],"minor_comments":[{"comment":"The symbol τ is reused: in Eq. (10) it denotes a minimum time between sampling and infection times, while in Eq. (7) it is the location-model time scale. Use different symbols to avoid confusion.","section":"Eq. (10)"},{"comment":"The text states NS = 49 cases, but the caption of Fig. 5 says 'NS = 50 sampled hosts'. Please reconcile.","section":"Sec. V.A / Fig. 5 caption"},{"comment":"The piecewise definition of the add/remove proposal ratio is difficult to parse due to missing braces in the displayed formula. Please clarify the notation and the factors involving (ki choose ku).","section":"Eq. (12)"},{"comment":"The definition of Δt_mutation and the host h ('the first host with no genetic information infected by j′ that is a predecessor of i′') is ambiguous. A clearer formal definition is needed.","section":"Sec. III.D, Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important question, but the pseudo-likelihood issue and the CI inconsistency are substantial. The revision should replace the genetic factor with a normalized Poisson pmf and explicitly define the location likelihood, or justify the pseudo-likelihood as an approximate posterior in a principled way (e.g., using composite likelihood theory). The CI error must be fixed. If these are addressed, the paper could be a valuable contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know this one: it proposes a tractable MCMC sampler for transmission trees using sampling times, locations, and pairwise SNP distances, and applies it to 49 NSW COVID-19 cases. The new bit is the systematic comparison of data types on tree topology (unsampled hosts, subtree count, degree distribution, Wiener index) in one framework, plus a new location factor. The code and data are out. That is real.\n\nThe good: the sampler is tested on a small exactly solvable case where they integrate Eq. (8) numerically and show the chain hits the right ratios. The NSW application, while small, is a genuine early-wave dataset with high coverage. The qualitative finding—that inferred network statistics depend on which metadata you feed in—is plausible and worth knowing.\n\nThe soft spots are not minor. Eq. (6) as printed is not a probability mass function over SNP distance d: the Poisson factorial is missing and the normalization depends on Δt, hence on the tree, so the Metropolis-Hastings ratio in Eq. (9) does not target the stated posterior. Eq. (7) has an 'A fixed by normalization' with no specification, and both genetic and location factors use the closest non-downstream sampled host rather than the actual infector. That makes them heuristic pseudo-likelihoods. Because the reported differences in NU (17.55 → 14.28, 25.68) are exactly what these factors shift, the direction and magnitude of the data-type effect could be an artifact. On top of that, the text reports NU = 17.55 with 80% CI [5,14]—a point estimate outside its own interval—and never gives the effective sample size for the real-data runs.\n\nThese are fixable: normalize the genetic factor (add the factorial and sum over d), specify Eq. (7)'s normalization, and recheck the CI. But as it stands, the central comparison between data types is not on solid ground.\n\nWho is this for? People working in phylodynamic inference or surveillance planning will want to read it, and the code is useful. It deserves a serious referee—the framework is sensible and the question matters—but my own verdict would be major revision, not acceptance, until the likelihood is properly normalized and the stats are fixed.","headline":"A useful but flawed comparison of data types for transmission-tree inference; the central claim is plausible but rests on an unnormalized pseudo-likelihood that needs fixing.","tokens_in":14517,"tokens_out":1639,"would_cite":false,"duration_ms":15469,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["92D30","62F15","60J22"],"pacs":[],"model":"deepseek-v4-flash","headline":"Inferred outbreak trees shift with the type of surveillance data available, not just with the outbreak itself.","keywords":["Transmission tree inference","Epidemic networks","Markov Chain Monte Carlo","Genomic epidemiology","Location metadata","Unsampled hosts","COVID-19 surveillance","Bayesian posterior sampling"],"falsifier":"Simulate an outbreak with a known transmission tree, known mutation process, and known movement process; run the same MCMC with each data-type combination; check whether genetic data still lowers and location data still raises the estimated N_U when the likelihood is computed against the true infector rather than the closest non-downstream sampled host. A more direct variant: on the NSW data, recompute the genetic factor using distances to arbitrary sampled hosts instead of the closest non-downstream host; if N_U moves substantially, the heuristic—not the data—drives the reported effect.","tokens_in":13504,"feed_emoji":"🧬","tokens_out":5627,"duration_ms":59224,"temperature":0.7,"pith_summary":"The paper sets out to show that the transmission trees inferred from an outbreak are not determined by the outbreak alone; they depend on which types of data are collected. The authors build a posterior distribution over transmission trees by multiplying five generative components—infection timing, sampling, offspring counts, genetic distance, and location proximity—and sample from it with a tailored Metropolis-Hastings Monte Carlo method. Applied to a Delta-wave COVID-19 dataset from New South Wales, the inferred number of unsampled hosts falls from 17.55 when only sampling times are used to 14.28 when genetic distances are added, and rises to 25.68 when location data are added; both metadata also narrow the inferred degree distribution while leaving the Wiener index unchanged. The significance of the claim is practical: judgments about hidden transmission and outbreak containment depend on which data a surveillance system happened to collect.","feed_headline":"Data choice shifts inferred COVID-19 transmission trees","feed_subtitle":"Genomic distance lowers estimated hidden cases; location data raises them — what you collect shapes the answer.","key_machinery":"The load-bearing object is the posterior P(T|D) in Eq. (8): a product over hosts of five conditional models—sampling (gamma-distributed detection delay), infection (gamma-distributed generation interval), offspring (negative binomial), genetic (Poisson in pairwise SNP distance against the closest non-downstream sampled host), and location (same-location advantage decaying with time)—times a constant prior, normalized by the evidence. The MCMC sampler moves through tree space with three proposals: time-shift of infection times, rewiring between chain and offspring topologies, and addition or removal of an unsampled host. The proposal ratios in Eqs. (10)-(12) keep the chain ergodic and reversi","core_discovery":"The central claim is that network properties important for outbreak management depend sensitively on the type of data used in the inference. Concretely, on 49 sampled COVID-19 cases from the early Delta wave in NSW, the estimated number of unsampled hosts N_U moves from a posterior-mean 17.55 with sampling dates alone to 14.28 when pairwise genetic distances are included and to 25.68 when location information is included, with 80% credible intervals [9,20] and [20,32] respectively; simultaneous inclusion of both metadata types narrows the inferred degree distribution but leaves the Wiener index essentially unchanged. The authors support this by validating an MCMC sampler on small synthetic p","pith_inferences":["The sign of the N_U shift (genetic lowers, location raises) may be specific to this dataset's cluster structure; the generalizable lesson is the sensitivity itself, which deserves replication on other outbreaks before guiding policy.","If the heuristic genetic and location likelihoods are valid, the approach becomes a surveillance-design calculator: one could choose which data types to fund by minimizing posterior variance of management-relevant statistics rather than maximizing point estimates.","A direct test would compare inferred parent-child links against known close-contact or household links in a contact-traced dataset, asking whether genetic or location metadata recovers true infectors more often.","The Wiener-index invariance suggests average pair distances might be credibly compared across studies that used different data, while N_U and degree statistics should not be."],"forward_implications":["Estimates of hidden, unsampled infections are conditional on data type; estimates of N_U should be reported together with the metadata that entered the inference.","Adding genomic or location metadata narrows the distribution of plausible trees, so the same dataset yields less spread in inferred degree statistics.","The Wiener index appears insensitive to metadata, making it a candidate tree summary that may transfer across data regimes.","Because only pairwise distances are needed, the method can simulate hypothetical data-availability scenarios before sequencing or location data exist.","The framework is portable to other directly transmitted viruses whose mutation and serial intervals are on comparable timescales."],"supporting_citations":[{"why":"Supplies the base generative model for infection, sampling, and offspring components that this paper extends with genetic and location factors.","marker":"[28]"},{"why":"Establishes the combination of genetic and epidemiological data for transmission-tree inference, the problem this paper reframes around data-type dependence.","marker":"[14]"},{"why":"Provides Bayesian inference of transmission trees from whole-genome sequence data, informing the design of the simplified genetic model.","marker":"[31]"},{"why":"Supplies the NSW COVID-19 dataset of genomes, collection dates, and pairwise location distances used in the application.","marker":"[27]"},{"why":"Provides the COVID-19 epidemiological parameter values used to fix model parameters.","marker":"[32]"},{"why":"Releases the code and data that make the reported application and MCMC results reproducible.","marker":"[36]"}],"fun_headline_variants":["Data type shifts inferred unsampled host counts in epidemics","Location vs genetics: different data, different hidden case estimates","Epidemic network inference hangs on data choice","What you measure changes the outbreak tree you infer"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The genetic and location likelihoods are computed against the closest sampled host that is not downstream of the node rather than against the true infector, and the product of those pairwise pseudo-likelihoods is assumed to behave like a real generative model of sequence evolution and movement; if that heuristic fails, the reported data-type dependence is an artifact of the pseudo-likelihood.","fun_headline_variants_meta":{"raw":{"variants":["Data type shifts inferred unsampled host counts in epidemics","Location vs genetics: different data, different hidden case estimates","Epidemic network inference hangs on data choice","What you measure changes the outbreak tree you infer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000544,"raw_usage":{"total_tokens":2411,"prompt_tokens":684,"completion_tokens":1727,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":428,"completion_tokens_details":{"reasoning_tokens":1664}},"tokens_in":428,"tokens_out":1727,"duration_ms":18129,"temperature":1.0,"reasoning_tokens":1664,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:07:14.125815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate an outbreak with a known transmission tree, known mutation process, and known movement process; run the same MCMC with each data-type combination; check whether genetic data still lowers and location data still raises the estimated N_U when the likelihood is computed against the true infector rather than the closest non-downstream sampled host. A more direct variant: on the NSW data, recompute the genetic factor using distances to arbitrary sampled hosts instead of the closest non-downstream host; if N_U moves substantially, the heuristic—not the data—drives the reported effect.","supporting_citations":[{"cited_title":"Genomic infectious disease epidemiology in partially sampled and ongoing outbreaks","cited_arxiv_id":null,"evidence_quote":"Supplies the base generative model for infection, sampling, and offspring components that this paper extends with genetic and location factors."},{"cited_title":"Unravelling transmission trees of infectious diseases by combining genetic and epidemiological data","cited_arxiv_id":null,"evidence_quote":"Establishes the combination of genetic and epidemiological data for transmission-tree inference, the problem this paper reframes around data-type dependence."},{"cited_title":"Bayesian inference of infectious disease transmission from whole-genome sequence data","cited_arxiv_id":null,"evidence_quote":"Provides Bayesian inference of transmission trees from whole-genome sequence data, informing the design of the simplified genetic model."},{"cited_title":"Guiding the design of SARS-CoV-2 genomic surveillance by estimating the resolution of outbreak detection","cited_arxiv_id":null,"evidence_quote":"Supplies the NSW COVID-19 dataset of genomes, collection dates, and pairwise location distances used in the application."},{"cited_title":"Sars-cov-2 (Covid-19) by the numbers","cited_arxiv_id":null,"evidence_quote":"Provides the COVID-19 epidemiological parameter values used to fix model parameters."},{"cited_title":"Data and Python code used in this paper","cited_arxiv_id":null,"evidence_quote":"Releases the code and data that make the reported application and MCMC results reproducible."}],"review_version":1}