{"id":"2336954c-0343-47b5-9c8f-34a04a5091bd","arxiv_id":"1908.06622","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A Bayesian mixture model jointly estimates time-varying means and spectra of multiple nonstationary time series, with covariate-dependent weights and missing-data handling.","lead":"AdaptSPEC-X is a Bayesian method that models many time series at once, estimating how their average level and cyclical patterns change over time and with covariates such as location. It handles missing data and can predict at unobserved covariate values, demonstrated on Australian rainfall and US measles records.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Gaussian missing-data imputation in Section 3.3 is the weakest link: the simulation uses Gaussian data, so it cannot validate missing-data handling for the zero-inflated and count-valued applications; a non-Gaussian missing-data simulation would settle whether this concern lands.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the missing-data imputation in Section 3.3 treats the Whittle likelihood as an exact multivariate normal distribution, and this assumption is untested for the non-Gaussian data in the applications. My reading of the paper confirms this. The simulation study is the only quantitative evidence for the missing-data machinery, and it uses Gaussian AR(2) data, so it cannot reveal bias caused by zero-inflation, nonnegativity, or heavy tails. The imputed values are then used to update the spectral parameters and the mixture allocations, so the assumption is structurally important, not a minor detail. The paper does cite Guinness (2019) and argues parsimony, but that is not a substitute for direct validation on non-Gaussian data with known missingness. Because the reader's conditional verdict already reflects this concern, I do not recommend changing the verdict; the appropriate next step is the concrete non-Gaussian missing-data simulation described above.","tokens_in":23378,"tokens_out":5997,"duration_ms":71714,"concrete_test":"Run the Section 5 simulation study with the same design but generate each series from a non-Gaussian process with known time-varying mean and spectrum, for example a zero-inflated version of (14) or a Gaussian AR(2) with lognormal or t innovations, while retaining 10% missingness. Compare AdaptSPEC-X estimates of mean and log spectrum to the truth, and also compare to a complete-case analysis or a fit using the exact Toeplitz likelihood. If median MSE or the posterior probabilities for the change-point events shift materially relative to the Gaussian case, the Gaussian imputation assumption is load-bearing; if performance is essentially unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AdaptSPEC-X 'handles missing values' rests on Section 3.3, where the Whittle likelihood (1) is treated as an exact multivariate normal density for the entire series and missing values are drawn from the conditional distribution in (11)-(13). This is a genuine model assumption, not a harmless asymptotic convenience: the Whittle likelihood is only an approximation for finite n, and the implied circulant covariance is not the exact covariance of the observed process. More importantly, in the two applications the data are far from Gaussian: monthly Australian rainfall is nonnegative and zero-inflated (9,933 zero months in the paper's own summary), and measles incidence rates are nonnegative and 16% zero. The imputation in (11) is unbounded and can impute negative rainfall or incidence values, and it cannot represent point masses at zero. Because Step 1 imputations feed directly into Step 2 (updating the spectra) and Step 3 (updating mixture allocations z), any systematic imputation bias propagates to the estimated time-varying spectra and to predictions at unobserved covariate values. The simulation study uses Gaussian AR(2) data with 10% missingness, so it does not exercise this assumption. Section 7 acknowledges Whittle small-sample inefficiency but does not acknowledge the Gaussian imputation assumption or test it on non-Gaussian data. The result is that the paper's second headline contribution is unverified for the data types in its own applications.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces AdaptSPEC-X, a Bayesian covariate-dependent infinite mixture model for a panel of nonstationary time series. Each mixture component is an AdaptSPEC model; the mixture weights are driven by time-independent covariates through a logistic stick-breaking process with a thin-plate Gaussian process prior on the log odds; and AdaptSPEC is extended with (i) an unknown time-varying segment mean and (ii) missing-data imputation based on treating the Whittle likelihood as a multivariate normal density (Section 3.3, Eqs. 11-13). Estimation is by MCMC combining data augmentation, reversible-jump moves, a label-swapping step, and a Riemann manifold Hamiltonian Monte Carlo update of the spline coefficients, for which the metric tensor is shown to be constant in the coefficients (Appendix A.1.2). The method is evaluated on a simulation with Gaussian AR(2) data from four known covariate regions, with 10% missingness, and applied to Australian monthly rainfall and US measles incidence. The simulation shows recovery of the true time-varying mean and log spectrum at most locations, and the applications produce the reported drought and vaccination signals.","tokens_in":23683,"tokens_out":7525,"duration_ms":77449,"significance":"If the claims hold, the paper makes a substantial methodological contribution: it packages the per-series AdaptSPEC model into a panel model with covariate-dependent weights, which is a natural and useful extension, and it demonstrates predictive inference at unobserved covariate values. The technical execution is careful: the truncated-normal conditional for the segment mean (Appendix A.1.1) and the constant metric tensor of the RMHMC step (Eq. A.4) are derived correctly, including the observation that only the zero-frequency periodogram depends on the mean. The paper also ships reproducible code and data on GitHub, and the Discussion honestly lists the main limitations (smooth stationary GP for the log odds, Whittle inefficiency for short segments, measurement error, discrete counts, time-varying covariates). The main weakness is that the headline claim of handling missing values is validated only for Gaussian data, while the applications involve nonnegative, zero-inflated series for which the Gaussian imputation assumption is most questionable.","major_comments":[{"comment":"The imputation step treats the Whittle likelihood as an exact multivariate normal density for the whole segment and draws missing values from the implied conditional Gaussian. This is a genuine modeling assumption, not an asymptotic convenience, for finite n. In the two applications the data are nonnegative and zero-inflated (9,933 zero months among the rainfall series and 30,439 zero weekly counts among the measles series), so the unbounded Gaussian imputation can produce negative rainfall or incidence values with positive probability and cannot represent point masses at zero. Because the simulation study of Section 5 uses Gaussian AR(2) data with only 10% missingness, it does not exercise the assumption for the data types used in the applications, and the claim that AdaptSPEC-X 'handles missing values' is therefore unverified for exactly the cases where it matters most. I would like to see either a simulation with non-Gaussian and/or zero-inflated data under artificially induced missingness, or a diagnostic in the applications (for example, the proportion of imputed draws that are negative or the distribution of imputed values near zero), together with a discussion of the consequences if that proportion is non-negligible. Section 7 acknowledges Whittle small-sample inefficiency but does not acknowledge this specific assumption or the validation gap.","section":"Section 3.3, Eqs. (11)-(13)"},{"comment":"The simulation study reports absolute recovery errors but contains no comparison with any existing method, even though the paper itself cites close competitors (Bruce et al. 2018, Krafty et al. 2017) and the natural baseline of fitting AdaptSPEC separately to each series. Without such a comparison it is difficult to judge what the covariate-dependent mixture and the LSBP weights add over simpler alternatives, and the notably larger MSE at D2/T2 (median 0.09 for the mean and 0.34 for the spectrum versus below 0.02 and 0.08 elsewhere) is the only quantitative evidence about where the modeling assumptions bite. Adding at least one baseline to the replicated simulation would make the empirical evaluation proportionate to the model's complexity and would substantially strengthen the paper's claims.","section":"Section 5, Figure 3"}],"minor_comments":[{"comment":"The text says '1N is an n×1 vectors of ones', but the vector should be N×1; this is a typographical error that should be corrected.","section":"Section 3.2, Eq. (9)"},{"comment":"The notation wh is used both for the log-odds function wh(·) and for the stacked vector (wh(u1),...,wh(uN))'; a bold or subscript convention would avoid ambiguity in Eqs. (9)-(10).","section":"Section 3.2"},{"comment":"The abstract states that software is available in the R package BayesSpec, while the footnotes say the latest CRAN version does not contain AdaptSPEC-X and the code is 'available from the authors'; since the Reproducibility section provides a GitHub link, the footnotes should direct readers there and clarify the CRAN status.","section":"Footnotes 1-2"},{"comment":"The authors note that the full-period spectra in Figure 12 look almost identical across states because of the wide power range; since Figure 14 is needed to see geographic heterogeneity, a per-panel normalization or a shared scale restricted to the pre-vaccine period would make Figure 12 more informative.","section":"Section 6.2, Figures 12-14"},{"comment":"The four regions have very different membership counts (41, 8, 18, and 33 time series), so the larger MSE at D2/T2 conflates cluster size with covariate structure; a balanced design, or reporting results conditional on cluster size, would separate these two explanations for the degraded performance.","section":"Section 5, Process (14)"},{"comment":"The Discussion candidly lists Whittle inefficiency and measurement error as limitations but does not mention the Gaussian missing-data imputation assumption of Section 3.3; a sentence acknowledging this assumption and pointing to the suggested validation would help readers calibrate the method's scope.","section":"Section 7, Discussion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well within the journal's scope and the derivations are careful; the central risk is the unvalidated Gaussian imputation for zero-inflated application data, which I believe a focused simulation or diagnostic can resolve. The absence of any baseline comparison in the simulation is a second, addressable weakness. The paper's reproducible code and honest limitation section are strengths. If the authors supply the missing-data validation and at least one baseline comparison, I would expect the paper to be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a careful read if you work on nonstationary spectral methods. The real contribution is a single framework that handles panels, time-varying means and spectra, multiple covariates, and missing data, and it ships working code and a GitHub repo. The appendices are coherent: the truncated normal for segment means and the constant metric tensor for RMHMC are correct. The simulation recovers the known time-varying mean and log spectrum at most locations, and the discussion in Section 7 honestly lists several limitations.\n\nThe softest spot is the missing-data imputation in Section 3.3. The Whittle likelihood is used as an exact multivariate normal for the whole series, so missing values are drawn from a Gaussian whose covariance is circulant. That is reasonable for the Gaussian simulation with 10% missingness, which is exactly the case the paper demonstrates. But the applications are monthly rainfall, which is zero-inflated, and measles incidence rates, counts with 16% zeros. The imputation can impute negative values and cannot represent zeros. Those imputed values feed directly into the spectrum and allocation updates, so bias would propagate. Section 7 notes the Whittle approximation is inefficient for small samples, but it does not flag the Gaussian imputation assumption itself. That is a genuine gap, and it is fixable: a simulation with count-valued or zero-inflated data and known missingness would settle it.\n\nLess serious but worth noting: the simulation has no baseline comparison to Bruce et al. or Cadonna et al., and the applications report posterior probabilities without MCMC diagnostics or Monte Carlo error. The tuning choices (H, M, J, B, tmin) are reasonable but not tested for sensitivity. These are addressable.\n\nFor whom: applied researchers analyzing panels of nonstationary environmental or epidemiological series, and methodologists wanting a flexible Bayesian toolkit. The central machinery holds up; the missing-data question is the one I'd want resolved. I'd send it to peer review and ask the authors to add the non-Gaussian missing-data simulation.","headline":"A competent synthesis that extends AdaptSPEC to panels with covariates and missingness; the missing-data imputation is the load-bearing assumption that needs testing.","tokens_in":24265,"tokens_out":3353,"would_cite":true,"duration_ms":29349,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M15","62F15","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"AdaptSPEC-X jointly estimates time-varying means and spectra for a panel of nonstationary time series, handles missing values, and makes predictions at unobserved covariate values.","keywords":["nonstationary time series","spectral estimation","covariate-dependent mixture","logistic stick breaking process","missing data","smoothing splines","Bayesian nonparametrics","locally stationary processes"],"falsifier":"Simulate panels from a known non-Gaussian, zero-inflated process, such as monthly aggregated rainfall with a known time-varying spectrum and 10-30% missingness, then compare AdaptSPEC-X's imputed values and estimated log spectra against the truth; systematic bias in imputations or median log-spectrum MSE well above the Gaussian-case values reported in Section 5 would show that the multivariate normal missing-data assumption is the failing component.","tokens_in":23109,"feed_emoji":"📈","tokens_out":7315,"duration_ms":69058,"temperature":0.7,"pith_summary":"AdaptSPEC-X is a Bayesian method for analyzing a panel of possibly nonstationary time series as a single object. The paper claims that modeling each series with a covariate-dependent infinite mixture of AdaptSPEC components, piecewise stationary processes whose segment log spectra are smoothing splines, allows time-varying means and spectra to be estimated jointly across series, with missing values handled inside the model and predictive inference available at unobserved covariate values. These features matter because single-series spectral estimates cannot be improved by finer sampling once the record is fixed, whereas a panel can borrow strength across similar series. In a simulation with four regimes and 10% missingness, the method recovers the true mean and log spectrum, with median mean-squared error below 0.02 for the mean and below 0.08 for the spectrum at almost all observed and unobserved locations. Applications to Australian rainfall and US measles incidence reproduce established drought and post-vaccine signals and add a new one: estimated rainfall variability has declined since 1950.","feed_headline":"AdaptSPEC-X estimates changing spectra across many series at once","feed_subtitle":"Handles missing values and predicts the spectrum at unobserved sites; validated on rainfall and measles panels.","key_machinery":"The central object is the AdaptSPEC-X model: a covariate-dependent infinite mixture (Equation 6) whose components are AdaptSPEC models, piecewise stationary processes in which the unknown number of segments and their boundaries are sampled by reversible-jump MCMC and each segment's log spectrum is a smoothing spline. Covariate dependence enters through logistic stick-breaking weights (Equation 8), whose log odds are modeled by a thin-plate Gaussian process prior on the covariates (Equation 9). Missing values are handled by treating the Whittle likelihood as an exact multivariate normal with a symmetric circulant precision matrix built from the estimated spectrum, so the conditional distribution of missing observations is multivariate normal (Equations 11-13). The MCMC scheme combines data augmentation for missing values, Polya-Gamma sampling for the stick-breaking coefficients, Riemann manifold Hamiltonian Monte Carlo for the spline coefficients, and a label-swapping move to improve mixing.","core_discovery":"The central claim is that a panel of nonstationary time series can be modeled jointly by an infinite mixture of AdaptSPEC components, with mixture weights driven by time-independent covariates through a logistic stick-breaking process. Each component segments a series into an unknown number of locally stationary pieces, estimates a time-varying mean and a smoothing-spline log spectrum per piece, and imputes missing values from the multivariate normal conditional implied by the Whittle likelihood. The paper argues that this structure simultaneously solves four problems that had been treated separately: multiple series, nonstationarity in both mean and spectrum, multiple covariates, and missing data. In the simulation study, the model recovers the true time-varying mean and log spectrum at observed and unobserved covariate values; in the applications, it detects the World War II and Millennium droughts and the post-1963 measles decline, and it estimates that rainfall variability across Australia has declined since 1950.","pith_inferences":["The missing-data assumption is the part most likely to limit the method: for zero-inflated rainfall-like series, a transformation or copula layer over the Gaussian imputation is a natural extension that could be tested by masking known values and comparing imputations.","The thin-plate GP prior on the log odds is smooth in covariate space, so regimes that change abruptly across space will be smoothed over; a piecewise or nonstationary GP would sharpen those boundaries.","Extending the LSBP to time-varying covariates, such as climate indices or policy changes, would let the mixture weights react to external drivers; the paper lists this as future work, but the model machinery suggests the extension is direct.","A head-to-head comparison with single-series AdaptSPEC on the same simulated panels would isolate how much accuracy is gained from borrowing strength across series, which the paper motivates but does not separately quantify."],"forward_implications":["Joint modeling across a panel improves local spectral estimation where the asymptotics of single-series locally stationary processes cannot be improved by further observation.","Panels with substantial missingness, 10% in simulation and 26% in the measles application, can be analyzed by nonparametric spectral methods without discarding or ad-hoc gap-filling the missing periods.","Spectral quantities can be predicted at unobserved covariate values, so features like the seasonal rainfall cycle can be interpolated to locations with no gauge record.","Time-varying means are estimated within the model, so series need not be pre-centered, and low-frequency information is not removed by a preliminary detrending step.","The method scales to large panels, about 192,000 observations in the measles application, through a combination of basis truncation and efficient sampling."],"supporting_citations":[{"why":"Supplies the AdaptSPEC component model, adaptive segmentation of a nonstationary series with smoothing-spline log spectra, which each mixture component extends.","marker":"Rosen et al. (2012)"},{"why":"Provides the logistic stick-breaking process used to make mixture weights depend on covariates.","marker":"Rigon and Durante (2020)"},{"why":"Gives the Whittle likelihood that serves both as the approximate Gaussian likelihood for each segment and as the basis for the multivariate normal missing-data imputation.","marker":"Whittle (1957)"},{"why":"Precedent and computational machinery for spectral estimation with missing data via periodic embeddings, cited to justify the multivariate normal assumption for missing values.","marker":"Guinness (2019)"},{"why":"Introduces the Polya-Gamma latent variable expansion used to sample the logistic stick-breaking coefficients.","marker":"Polson et al. (2013)"},{"why":"Provides the reversible-jump MCMC framework that lets the number of segments in each AdaptSPEC component change dimension.","marker":"Green (1995)"},{"why":"Supplies the Riemann manifold Hamiltonian Monte Carlo step that accelerates sampling of the smoothing spline coefficients.","marker":"Girolami and Calderhead (2011)"}],"fun_headline_variants":["Covariate-dependent spectral modeling for multiple nonstationary series","AdaptSPEC-X: joint spectral analysis with missing data and covariates","Model changing spectra across series using covariate-driven mixtures","Bayesian spectral modeling for nonstationary panels with missing values"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's load-bearing premise is that the Whittle likelihood can be treated as an exact multivariate normal distribution for the entire series, so that missing values are drawn from a Gaussian whose covariance is built from the estimated spectrum; if that approximation is poor, for example with strongly non-Gaussian or zero-inflated rainfall, or with short segments where the Whittle error is large, the imputed values and every downstream estimate inherit the bias.","fun_headline_variants_meta":{"raw":{"variants":["Covariate-dependent spectral modeling for multiple nonstationary series","AdaptSPEC-X: joint spectral analysis with missing data and covariates","Model changing spectra across series using covariate-driven mixtures","Bayesian spectral modeling for nonstationary panels with missing values"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000579,"raw_usage":{"total_tokens":2725,"prompt_tokens":937,"completion_tokens":1788,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1719}},"tokens_in":553,"tokens_out":1788,"duration_ms":14673,"temperature":1.0,"reasoning_tokens":1719,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:38:56.259224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate panels from a known non-Gaussian, zero-inflated process, such as monthly aggregated rainfall with a known time-varying spectrum and 10-30% missingness, then compare AdaptSPEC-X's imputed values and estimated log spectra against the truth; systematic bias in imputations or median log-spectrum MSE well above the Gaussian-case values reported in Section 5 would show that the multivariate normal missing-data assumption is the failing component.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the AdaptSPEC component model, adaptive segmentation of a nonstationary series with smoothing-spline log spectra, which each mixture component extends."},{"cited_title":"and Durante, D","cited_arxiv_id":null,"evidence_quote":"Provides the logistic stick-breaking process used to make mixture weights depend on covariates."},{"cited_title":"(2019), Spectral density estimation for random fields via periodic embeddings, Biometrika, 106, 267--286","cited_arxiv_id":null,"evidence_quote":"Precedent and computational machinery for spectral estimation with missing data via periodic embeddings, cited to justify the multivariate normal assumption for missing values."},{"cited_title":"G., Scott, J","cited_arxiv_id":null,"evidence_quote":"Introduces the Polya-Gamma latent variable expansion used to sample the logistic stick-breaking coefficients."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the reversible-jump MCMC framework that lets the number of segments in each AdaptSPEC component change dimension."},{"cited_title":"and Calderhead, B","cited_arxiv_id":null,"evidence_quote":"Supplies the Riemann manifold Hamiltonian Monte Carlo step that accelerates sampling of the smoothing spline coefficients."}],"review_version":1}