{"id":"e00414d5-2f47-408b-a4c5-f5a9eaddd7fa","arxiv_id":"2509.10668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new framework approximates thinned count autoregressions with a normal-normal model and a latent Gaussian transform, enabling simple Bayesian reconstruction of under-reported epidemic curves.","lead":"This paper shows that ignoring under-reporting in disease count models distorts transmission estimates, then proposes a Gaussian approximation plus a latent integer transform to make epidemic curve reconstruction much simpler. If the approximation holds, epidemiologists can fit under-reported models with standard software instead of custom MCMC.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inverse-CDF reconstruction step is validated only by pointwise quantile agreement; the joint serial dependence of the reconstructed X_t is asserted but never checked.","rationale":"Good-faith reading: the paper is proposing a practical approximation and is honest about some limitations (small counts, MCMC difficulties). The Section 2 moment results are a useful extension, the simulation design is reasonable, and the real-data applications are plausible demonstrations rather than validation. My concern is not that the approximation is wrong in all regimes; for large counts a Gaussian CLT argument is plausible. The issue is that the specific claim 'Z*_{1:t} ≈ Z^X_{1:t}' is presented as the justification for the reconstruction transform, yet the evidence offered is limited to marginal/pointwise agreement. The joint distribution of a time series is not determined by its marginal quantiles; an autocorrelation diagnostic is the natural check. If that check fails, the method might still be useful as an approximate smoothing tool, but the paper's central claim of accurate epidemic-curve reconstruction from the normal-normal approximation would need to be qualified. I therefore keep the CONDITIONAL verdict; the authors can resolve the concern by adding the autocorrelation diagnostic, stating how negative latent values are handled, and releasing the promised code. The reader's weakest assumption already pointed to the serial-dependence question, so I agree.","tokens_in":18013,"tokens_out":8757,"duration_ms":102018,"concrete_test":"Re-run the Section 4 simulation for, e.g., the φ=0.8, π=0.4 scenario with T=50. For each posterior draw from the approximate model, form X_t = F^{-1}_{X_t}(Φ(Z*_t)) and compute the sample lag-1 autocorrelation of the resulting path; build the 90% posterior predictive interval for this autocorrelation. Do the same for the exact binomially thinned Poisson model. If the approximate interval covers the true simulated ACF(1) substantially less often than the exact interval does, the §3.3 assertion is unsupported and the reconstruction claim must be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central reconstruction claim rests on §3.3's assertion that Z*_{1:t} ≈ Z^X_{1:t}, the latent Gaussian series from the Jia et al. construction, so that X_t = F^{-1}_{X_t}(Φ(Z*_t)) yields a posterior close to the exact thinned Poisson autoregression. What Section 4 actually validates is pointwise agreement of posterior medians and 90% credible intervals for X_t; Table 1's 'perfect match rate' is a per-time-point criterion. Pointwise agreement does not imply joint agreement. In the approximating model, Z*_t are defined as independent standard normal innovations; their posterior can acquire dependence through the observed data, but the paper never compares the serial dependence of the reconstructed X_t with that of the true X_t. Since epidemic-curve reconstruction concerns a coherent trajectory, an undetected distortion of the lag structure would bias reconstructed epidemics even if marginal quantiles line up. The approximation is also admitted to degrade for mean counts below 5, but no regime characterization or safeguard is given. Relatedly, model (9) contains variance terms sqrt(π(1−π)Z_t) and sqrt(λ_t); as written it is undefined when Z_t or λ_t is negative, and the paper does not state the positivity constraint used in the fits.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies count-valued autoregressions with under-reporting. Section 2 derives moment-based results characterizing how ignoring binomial thinning biases estimates of the autoregressive parameter φ and the exogenous component ν, with two propositions on when ν is overestimated. Section 3 proposes an approximate Bayesian framework: replace a binomially thinned Poisson autoregression with a normal-normal model, standardize the latent process as Z*, and reconstruct integer counts via the latent Gaussian transformation X_t = F^{-1}_{X_t}(Φ(Z*_t)). The approximation is evaluated in a simulation study and applied to rotavirus data in Saarland and COVID-19 data in England. The paper argues that the approach retains the mechanistic appeal of thinned autoregressions while substantially simplifying inference.","tokens_in":18320,"tokens_out":6599,"duration_ms":70764,"significance":"If the central approximation is valid, this is a practically useful contribution: it offers a software-friendly route to epidemic curve reconstruction in multivariate, covariate-adjusted settings where exact MCMC is infeasible. The Section 2 results on the direction of bias from ignoring under-reporting are also of independent interest and extend earlier moment-matching work. Strengths include the reproducible analysis scripts, the two real-data applications, and the explicit connection to the latent Gaussian count time series construction of Jia et al. (2023). However, the paper's central claim that the approximate posterior is 'virtually identical' to the exact posterior is currently supported mainly by pointwise comparisons, and several implementation details of the approximate model are underspecified.","major_comments":[{"comment":"Proposition 1 states a result for 'any consistent estimator' of ν in the misspecified model, but the proof only uses the moment equation \\hatν = (1-\\hatϕ)\\tilde μ. Under misspecification, different estimators (e.g., MLE vs. method of moments) converge to different pseudo-true values, and the Poisson AR MLE need not satisfy this moment equation asymptotically. The claim is therefore overbroad. Please restrict the proposition to estimators based on the moment equations, or prove that all consistent estimators share the same pseudo-true value. The same issue propagates to Proposition 2.","section":"Section 2 / Appendix A.1"},{"comment":"The normal-normal model uses variances sqrt(π(1−π)Z_t) and sqrt(λ_t), and model (12) uses a binomial probability of sum_{k=0}^{13} Z_{i,d-k}/pop_i. Since Z_t is Gaussian, it is negative with positive probability, making these expressions undefined or out of range. No positivity or truncation constraint is stated. This is not cosmetic: the approximate posterior is not well defined for negative Z. Please specify how the implementation handles negative Z (truncation, abs(), softplus, or similar) or reparameterize to guarantee positivity.","section":"Eq. (9) and Eqs. (11)-(12)"},{"comment":"The central claim Z*_{1:t} ≈ Z^X_{1:t} is validated only by pointwise agreement. Table 1's 'perfect match rate' and Figure 4 compare posterior medians and credible intervals time point by time point; they do not compare the joint serial dependence of the reconstructed X_t (e.g., ACF, partial ACF, or joint distributions of (X_t, X_{t-1})). Since epidemic curve reconstruction is about a coherent trajectory, a distortion of the lag structure would bias reconstructed epidemics even if marginal quantiles line up. Please add diagnostics for joint/autocorrelation structure.","section":"Section 3.3 and Section 4"},{"comment":"The reconstruction algorithm is not fully specified. The paper recommends computing X_t = F^{-1}_{X_t}(Φ(Z*_t)) with Z* the standardized innovations, but the rotavirus application writes X_t = F^{-1}_{X_t}(Φ(Z_t)). Also, F_{X_t} should be the conditional CDF of X_t given the relevant past and λ_t; this dependence is not stated. Please clarify the exact mapping used in the code and in each application.","section":"Section 3.3 and Section 5.1"}],"minor_comments":[{"comment":"The sentence 'maps accurately maps this continuous process back to the integers' contains a duplicated phrase.","section":"Abstract"},{"comment":"Typo: 'thehhhmodel' should be 'the hhh model'.","section":"Section 1"},{"comment":"The sentence about 'attenuate the correlation between X and π' should presumably read 'between Z and π'.","section":"Section 3.2"},{"comment":"Minor typos: 'from from' in the Figure 3 caption, and 'mispecified' in Section 5.1. Also, the text says simulations with Rhat>1.05 were removed, but Table 1 says Rhat<1.01 was required; please make the filtering criterion precise.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising and potentially publishable core, but the current version overclaims the fidelity of the approximation. The positivity issue in the defining equations and the lack of joint-distribution validation are fixable, and a revision addressing them would be worth reconsidering. I am not recommending rejection because the central idea is novel and the applications are relevant."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper deserves a serious referee, but not because the central claim is settled. The genuinely new pieces are the explicit limiting expressions in Section 2 for what happens when you fit a mis-specified Poisson autoregression to binomially thinned counts, and the normal-normal approximation plus inverse-CDF transform for reconstructing the integer-valued latent process. The former is a real extension of Bracher and Held (2021), giving clean intuition about when ignoring under-reporting inflates the exogenous component. The latter could lower the computational barrier for fitting thinned autoregressions in Stan, which is practically valuable.\n\nNow the soft spots. Proposition 1 says “any consistent estimator” but the proof only covers moment-based estimators. Under misspecification, the MLE converges to the pseudo-true value, which need not equal the moment-based limit unless the score equations coincide with the Yule–Walker type equations. That is a fixable overclaim, but it should be corrected or proved.\n\nMore serious is the load-bearing claim in Section 3.3 that Z*_t ≈ Z^X_t, so the inverse-CDF transform yields a posterior close to that of the exact thinned Poisson autoregression. What Section 4 actually checks is pointwise agreement of posterior medians and 90% credible intervals; the “perfect match rate” is a per-time-point criterion. The paper never compares the serial dependence of the reconstructed X_t with that of the true X_t. For an epidemic curve, the joint trajectory matters; an undetected distortion of the lag structure would bias reconstructions even when marginal quantiles line up. The authors admit the approximation degrades for small counts, but they don’t characterize the regime or provide a safeguard.\n\nAlso, model (9) writes variance terms sqrt(π(1−π)Z_t) and sqrt(λ_t); the model is undefined when Z_t or λ_t is negative. No positivity constraint is stated in the paper. That is a small but embarrassing gap. And despite the text saying reproducible analysis scripts are included in the supplement, no code or data appear to be shipped with the preprint. That should be fixed before publication.\n\nThe real-data applications are illustrative rather than validating; they show the method runs and gives plausible curves. The discussion of identifiability is honest and useful.\n\nBottom line: worth refereeing. The Section 2 results are solid apart from the overclaim; the approximation idea is promising, but the authors need to demonstrate joint fidelity of the reconstructed process, not just marginal coverage. I’d engage with it and would cite the Section 2 results. Bring it to a reading group if you have a spare hour.","headline":"Useful new results on the pitfalls of ignoring under-reporting in Poisson autoregressions, but the key approximation claim rests on an unvalidated dependence assumption and a few loose ends.","tokens_in":18769,"tokens_out":3273,"would_cite":true,"duration_ms":35062,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that ignoring under-reporting in Poisson autoregressions biases reproduction numbers and offers a normal-normal approximation to recover true counts.","keywords":["under-reporting","epidemic curve reconstruction","Poisson autoregression","binomial thinning","latent Gaussian transformation","normal-normal approximation","state-space models","Bayesian inference"],"falsifier":"Run a long simulation (T ≥ 200) from a binomially thinned Poisson autoregression with small mean counts (e.g., about 3–5) and strong autocorrelation (ϕ ≈ 0.8). If the approximate-model credible intervals for the true epidemic curve have coverage far below nominal, or the lag-1 autocorrelation of Z*_t deviates materially from that of the true latent process, the central claim fails.","tokens_in":1529,"feed_emoji":"🦠","tokens_out":4973,"duration_ms":106335,"temperature":0.7,"pith_summary":"The paper establishes that fitting Poisson autoregressions to under-reported case counts systematically biases the reproduction number downward and often biases the exogenous component upward, and quantifies exactly when each bias occurs. To correct this, it introduces a Bayesian framework that replaces the binomially thinned Poisson autoregression with a normal-normal approximation of matching conditional moments, then maps the continuous latent states back to integers via the inverse-CDF transform X_t = F^{-1}(Φ(Z*_t)). The authors argue that this yields nearly the same posterior distribution for parameters and epidemic curve reconstructions as the exact thinned model, while being fast enough for standard gradient-based Bayesian software. If correct, this gives practitioners a general, software-friendly way to estimate hidden incidence while preserving the mechanistic interpretation of thinned autoregressions.","feed_headline":"Gaussian fit recovers true epidemic curves from under-reported counts","feed_subtitle":"The paper quantifies how ignoring under-reporting skews reproduction numbers and offers a faster way to estimate hidden incidence.","key_machinery":"The central mechanism is the latent Gaussian transformation X_t = F^{-1}_{X_t}(Φ(Z*_t)), applied to standardized latent variables Z*_t from a normal-normal approximation to a thinned count autoregression. The approximation replaces the binomial thinning Y_t | X_t ~ Bin(X_t, π) and the Poisson autoregression X_t | X_{t-1} ~ Pois(λ_t) with normal distributions that match conditional means and variances, yielding a continuous latent process Z_t that is easy to sample with gradient-based Bayesian software. The inverse-CDF transform then maps each posterior draw of Z*_t back to integer counts, and the paper asserts that the joint distribution of these transformed integers is close to the posterio","core_discovery":"The paper's central claim is that the posterior distribution of the parameters and latent true counts from a binomially thinned Poisson autoregression can be accurately approximated by fitting a normal-normal approximation—replacing each Poisson with a normal of matching conditional mean and variance, and each binomial thinning with a normal of matching conditional mean and variance—and then transforming each posterior draw of the standardized latent variable Z*_t via X_t = F^{-1}_{X_t}(Φ(Z*_t)). This transformation is key: it converts the continuous approximate latent process into integer counts with the correct marginal distributions while preserving the serial dependence structure, so the","pith_inferences":["If the normal-normal approximation with the latent Gaussian transform is as accurate as claimed for larger counts, it could serve as a fast surrogate for inference in other latent count models, such as ecology n-mixture or capture-recapture models, where discrete latent states create computational bottlenecks.","The theoretical result that under-reporting biases ϕ downward implies that apparent declines in reproduction numbers over time may be confounded with improvements in reporting coverage; a testable implication is that regions with better reporting should show higher estimated ϕ from naive fits.","The paper validates the approximation only for moderately large counts and a single series length; a formal error bound on the Gaussian approximation to the autocorrelation of Z*_t would strengthen the generalizability of the claim.","The authors' framework naturally supports time-varying reporting probabilities driven by covariates, as they demonstrate for Covid-19; validating reconstructions against independent serosurveys, as they do with REACT data, could become a standard model-checking step."],"forward_implications":["Epidemic curve reconstructions can be computed with standard gradient-based Bayesian software instead of custom particle filters or MCMC samplers over discrete latent states, enabling faster and more scalable analyses.","Published reproduction numbers from Poisson autoregressions fit to under-reported counts are likely to be systematically underestimated unless under-reporting is explicitly modeled.","The framework extends to multivariate spatiotemporal settings with covariates and additional data sources, as demonstrated by the England conurbation analysis that integrated random PCR testing data.","The method is designed for series with reasonably large counts; the paper notes the approximation degrades when the mean count falls below about 5.","Moment-based estimators for the thinned model are too variable for practical use, supporting the need for full posterior inference."],"fun_headline_variants":["Gaussian approximation recovers true incidence from under-counts","New framework for under-reported epidemic curves","Fast statistical method for under-reported incidence","Normal-normal approximation for thinned epidemic counts"],"cache_read_input_tokens":20096,"weakest_assumption_plain":"The method's central premise is that the standardized latent variable from the normal approximation carries essentially the same serial dependence as the latent Gaussian that would generate the true integer counts, an assumption the paper checks only in simulations with moderately large counts (means 6.7–40, series length 50) and explicitly concedes breaks down for very small counts.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian approximation recovers true incidence from under-counts","New framework for under-reported epidemic curves","Fast statistical method for under-reported incidence","Normal-normal approximation for thinned epidemic counts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00115,"raw_usage":{"total_tokens":4581,"prompt_tokens":696,"completion_tokens":3885,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":3829}},"tokens_in":440,"tokens_out":3885,"duration_ms":34651,"temperature":1.0,"reasoning_tokens":3829,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T17:42:02.403925+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a long simulation (T ≥ 200) from a binomially thinned Poisson autoregression with small mean counts (e.g., about 3–5) and strong autocorrelation (ϕ ≈ 0.8). If the approximate-model credible intervals for the true epidemic curve have coverage far below nominal, or the lag-1 autocorrelation of Z*_t deviates materially from that of the true latent process, the central claim fails.","supporting_citations":[],"review_version":1}