{"id":"c4748ca2-3d1d-4a53-82de-e5084c0d8c3c","arxiv_id":"2607.05579","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Unnormalized power-prior MCMC samples of a0 are proportional to C(a0), so mixture (plus optional extrapolation) approximations enable normalized power priors in standard Bayesian software.","lead":"A sampling trick lets analysts run normalized power priors inside ordinary Bayesian software by estimating the hard normalizing constant from unnormalized MCMC draws. This could make historically informed Bayesian analyses more routine in clinical and observational work without custom marginal-likelihood code.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The load-bearing risk is that mixture-plus-log-linear extrapolation of C(a0) can still distort the NPP posterior for θ when UnPP mass is near zero and log C is strongly convex.","rationale":"The Reader correctly isolates the weakest practical assumption: that a mixture fitted only where UnPP places mass, patched by log-linear extrapolation to a Laplace estimate of log C(1), leaves the posterior of θ essentially undistorted when C decays rapidly. The identity itself is sound and the binomial closed-form check is clean; the paper is transparent about the pathology and supplies a usable SAS workflow. No stronger internal inconsistency appears. The residual concern is therefore exactly the one the Reader named—empirical accuracy of the hybrid approximation for θ under strong convexity/skew—and it already justifies CONDITIONAL rather than unconditional ACCEPT. The concrete test above would settle whether that residual error is negligible or material on the authors’ own hard example, without requiring new theory. Verdict remains CONDITIONAL; no upgrade or downgrade is warranted from this second pass.","tokens_in":12961,"tokens_out":762,"duration_ms":6566,"concrete_test":"On the same linear-regression simulation of Section 3.2, recompute the NPP posterior for (b0,b1,b2,s2) three ways: (i) true closed-form C(a0), (ii) mixture + linear extrapolation with a*_0 = amax0, and (iii) a dense-grid bridge-sampling dictionary for C(a0). Report max absolute and relative differences in posterior means, SDs, and 95% HPD endpoints for each coefficient; if any coefficient mean or HPD endpoint shifts by more than ~5–10% of its true-NPP width under (ii) relative to (i)/(iii), the “mild impact” claim fails for this regime.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the identity that UnPP marginals of a0 (uniform prior, powered historical likelihood as pseudo-likelihood) are proportional to C(a0), so a parametric mixture (plus anchors at 0 and Laplace log C(1) with linear log-scale extrapolation past a*_0) yields a plug-in usable inside generic MCMC. That identity is correct under the stated conditions (Section 2.2). The soft spot is the accuracy claim for scientific parameters θ when C(a0) decays extremely rapidly (Sections 3.2–3.4, Figures 2–6). Because log C is known to be convex (Carvalho & Ibrahim 2021, cited by the authors), a linear bridge from a*_0 (chosen as amax0 or 0.9 amax0) to log C(1) systematically mis-specifies the tail. The paper shows that a pure mixture underestimates the tail and inflates posterior mass on large a0 (Figure 4), while the piecewise fix improves visual overlap for θ (Figure 6) yet still leaves the a0 marginal visibly off. No quantitative bound is given on how large the residual error in log C may become before the posterior of θ is materially shifted, nor is the method compared against bridge/importance sampling on the same regression example. Thus the practical guarantee that “impact on model parameters is relatively little” remains an unquantified assumption precisely in the regime the authors flag as hardest.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper addresses the practical barrier to using the normalized power prior (NPP): the intractable normalizing function C(a0) = ∫ L(θ|D0)^a0 π0(θ) dθ. It exploits the identity that, under a uniform prior on a0 and treating the powered historical likelihood as a pseudo-likelihood, the marginal of a0 under the unnormalized power prior (UnPP) is proportional to C(a0). MCMC samples from UnPP are therefore used to fit a parametric mixture for C(a0) (or a weighted/transformed version), which is then plugged into a standard NPP analysis inside generic Bayesian software. For the common case of rapid decay of C(a0) near zero, the authors add endpoint anchors (log C(0)=0 and a Laplace estimate of log C(1)) and log-linear extrapolation past a breakpoint a*_0, with optional reparameterization and weighting. Closed-form binomial and conjugate linear-regression examples illustrate agreement when the mixture is adequate and partial recovery of θ posteriors after extrapolation when pure mixtures fail.","tokens_in":13347,"tokens_out":1154,"duration_ms":8789,"significance":"If the approximation is reliable enough for the scientific parameters θ, the contribution is practically important: NPP becomes usable inside PROC MCMC, Stan, JAGS, BUGS, or NIMBLE without external bridge/importance sampling or model-specific marginal-likelihood code. The identity itself is correctly stated and underutilized; the binomial closed-form check and the honest documentation of UnPP skewness and tail failure modes are strengths. The appendix SAS code further lowers the barrier to adoption. The work therefore has clear applied value for dynamic historical borrowing, provided the residual approximation error on θ is shown to be controlled in the hard regime the authors themselves flag.","major_comments":[{"comment":"Sections 3.2–3.4 and Figures 3–6: the central practical claim is that mixture-plus-log-linear extrapolation yields NPP posteriors for θ that are not materially distorted even when C(a0) decays extremely rapidly. Because log C is known to be convex (Carvalho & Ibrahim 2021, cited), a linear bridge from a*_0 (amax0 or 0.9 amax0) to a Laplace log C(1) systematically misspecifies the tail. Figure 6 shows improved visual overlap for regression coefficients relative to the pure-mixture failure in Figure 4, yet the a0 marginal remains visibly off and no quantitative bound (e.g., total-variation or KL distance on the θ marginal, or coverage of credible intervals) is reported. Without such a metric, or a head-to-head comparison against bridge/importance sampling on the same regression example, the statement that “impact on the model parameters is relatively little” remains an unquantified assumpt","section":null},{"comment":"Section 3 and Discussion: empirical support is limited to two conjugate models (binomial with closed-form C(a0) and linear regression with closed-form C(a0)). The method is advertised for generic software and non-conjugate settings, yet no non-conjugate or higher-dimensional example is shown, nor is sensitivity to mixture component count, breakpoint choice, or Laplace accuracy for log C(1) systematically examined. At least one non-conjugate illustration (or a clear statement of the scope limitation) is needed to underwrite the claim of routine applicability.","section":null}],"minor_comments":[{"comment":"Throughout (e.g., Abstract, §2.2): “psuedo” should be “pseudo”; several other minor typos (“parametrically”, spacing around a0) should be cleaned.","section":null},{"comment":"Figure captions (Figs. 1–6) are informative but axis labels and legend fonts are small; ensure they remain legible in print.","section":null},{"comment":"§3.5–3.6 introduce transformed-scale sampling and weighting/tempering as remedies, yet provide no numerical illustration; either add a brief example or move the material to a short appendix so the main narrative stays focused on the mixture-plus-extrapolation pipeline that is actually demonstrated.","section":null},{"comment":"Appendix A: the SAS code is valuable; a short note on how the same plug-in would be written in Stan or JAGS would broaden accessibility without lengthening the main text.","section":null}],"recommendation":"major_revision","confidential_remarks":"The identity and software-portability angle are genuine and useful; the paper is closer to a solid methods note than to a reject. The main risk is over-claiming accuracy for θ under extreme skew without quantitative error metrics. If the authors add a non-conjugate example and a simple error table (or comparison to bridge sampling) on the regression case, the contribution would fit a computational-statistics or biostatistics methods journal well. Scope is appropriate for stat.ME; no novelty or citation concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is a workflow, not new theory: under a uniform prior on a0, UnPP draws on historical data alone are proportional to C(a0), so you fit a mixture (optionally with log C(0)=0 and a Laplace log C(1) anchor plus log-linear tail past a*_0) and drop log ˜C into any generic package. That is exactly the barrier that has kept NPP out of PROC MCMC / Stan / JAGS for most users.\n\nWhat they do well is keep the claim honest. The identity in §2.2 is correctly stated. The binomial closed form (Figure 1) shows the pipeline recovers both C and the joint posterior. The regression case is not papered over: pure mixture tails fail when UnPP mass collapses near zero (Figures 2–4), and the piecewise fix improves the θ marginals even while the a0 marginal stays visibly off (Figures 5–6). Appendix SAS is complete enough to reproduce. Citations to Ibrahim–Chen, Duan, Neuenschwander, and Carvalho–Ibrahim are the right ones; no invented entities.\n\nSoft spots are real but proportionate. Free parameters (component count, a*_0 = amax or 0.9 amax, optional weights/tempering) are acknowledged rather than hidden. Because log C is convex, the linear bridge is only a local fix; residual error in the tail is not bounded, and there is no head-to-head against bridge/importance sampling on the same regression example. Only two models are shown. Still, the authors never claim a universal guarantee—only that impact on scientific parameters is “relatively little” in the cases they tried—and the figures support that limited claim.\n\nThis is for applied Bayesian biostatisticians who already use power priors and want NPP without external marginal-likelihood machinery. It is not foundational, but it is a concrete, software-portable step that removes a real friction. I would send it to referees; the identity and the reproducible pipeline deserve a careful look, and the tail caveats are already on the page. Worth citing if you implement NPP yourself; bring to reading group only if the group is methods-for-trials.","headline":"Portable UnPP-to-mixture pipeline that lets ordinary MCMC software run NPP; identity is sound, binomial check clean, regression tails still the practical soft spot.","tokens_in":13921,"tokens_out":539,"would_cite":true,"duration_ms":5180,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62-08"],"pacs":[],"model":"grok-4.5","headline":"MCMC samples from the unnormalized power prior already contain the normalizing constant of the normalized power prior, so NPP can be fit inside ordinary Bayesian software.","keywords":["normalized power prior","unnormalized power prior","historical data borrowing","MCMC approximation","normalizing constant","Bayesian software","mixture density"],"falsifier":"In a linear-regression setting where the true closed-form C(a0) is known, replace the mixture-plus-extrapolation with pure mixture tails and check whether the resulting NPP posterior for the regression coefficients systematically over-borrows relative to the exact NPP posterior.","tokens_in":13836,"feed_emoji":"⚖️","tokens_out":675,"duration_ms":5545,"temperature":0.7,"pith_summary":"Normalized power priors let analysts borrow historical data while letting the data decide how much to borrow, but they require a hard-to-compute normalizing function C(a0). The paper shows that, under a uniform prior on the power weight a0, the marginal distribution of a0 obtained from an ordinary unnormalized power prior run on the historical data alone is already proportional to that C(a0). Consequently, one can draw MCMC samples from the easy unnormalized model, fit a simple mixture (or mixture-plus-log-linear tail) to those samples, and insert the resulting plug-in function into any standard Bayesian package. The method needs no external marginal-likelihood routines and works in PROC MCMC, Stan, JAGS, BUGS or NIMBLE. When C(a0) decays very sharply the authors add two analytic anchors—C(0)=1 and a Laplace estimate of C(1)—so that the approximation remains usable for the scientific parameters of interest.","feed_headline":"Unnormalized MCMC already holds the missing NPP constant","feed_subtitle":"Fit a mixture to those samples and run normalized power priors inside ordinary Bayesian packages","key_machinery":"The functional identity ˜π(a0|D0) ∝ C(a0) under a uniform prior on a0. It converts the intractable normalizing function into the marginal density of an ordinary MCMC run, so that a mixture (or piecewise mixture-plus-extrapolation) approximation of that density becomes a usable log C(a0) term inside standard Bayesian code.","core_discovery":"Under a uniform prior on a0, the marginal distribution of a0 induced by the unnormalized power prior (when the powered historical likelihood is treated as a pseudo-likelihood) is functionally proportional to the normalizing constant C(a0) of the normalized power prior. MCMC samples of a0 from that unnormalized model therefore supply a direct Monte-Carlo estimate of C(a0), which can be turned into a parametric plug-in and used inside any generic Bayesian software to fit the fully normalized power prior.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Unnormalized MCMC samples of a0 approximate NPP constant C(a0)","Marginal a0 from unnormalized power prior is proportional to C(a0)","Generic MCMC on unnormalized prior yields plug-in for normalized PP","Estimate C(a0) from unnormalized a0 draws for use in any Bayesian software","Ordinary unnormalized power prior MCMC enables full NPP fitting"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That a mixture fitted only where the unnormalized MCMC places mass, together with a straight-line log-scale extrapolation to a Laplace estimate of C(1), is accurate enough that the final posterior for the scientific parameters is not materially distorted when C(a0) drops off extremely fast.","fun_headline_variants_meta":{"raw":{"variants":["Unnormalized MCMC samples of a0 approximate NPP constant C(a0)","Marginal a0 from unnormalized power prior is proportional to C(a0)","Generic MCMC on unnormalized prior yields plug-in for normalized PP","Estimate C(a0) from unnormalized a0 draws for use in any Bayesian software","Ordinary unnormalized power prior MCMC enables full NPP fitting"]},"model":"grok-4.5","effort":"low","cost_usd":0.006018,"raw_usage":{"total_tokens":1537,"prompt_tokens":753,"num_sources_used":0,"completion_tokens":101,"cost_in_usd_ticks":60180000,"prompt_tokens_details":{"text_tokens":753,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":683,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":753,"tokens_out":101,"duration_ms":5565,"temperature":1.0,"reasoning_tokens":683,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T05:31:00.390624+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"In a linear-regression setting where the true closed-form C(a0) is known, replace the mixture-plus-extrapolation with pure mixture tails and check whether the resulting NPP posterior for the regression coefficients systematically over-borrows relative to the exact NPP posterior.","supporting_citations":[],"review_version":1}