{"id":"60f1a620-3cfb-4e24-b2a6-a211a4983706","arxiv_id":"2506.08760","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For segmented regression models, the penalty per change-point in AIC is 2 when the regression function is continuous and 6 when it is discontinuous.","lead":"This paper derives a corrected Akaike information criterion for segmented (piecewise) regression, giving a penalty of 2 per change-point when the fitted curve is continuous and 6 per change-point when the curve jumps. It matters because segmented regression is widely used in epidemiology, and the choice of penalty changes which trends and turning points are reported.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Discontinuous-case AIC penalty 6 depends entirely on Assumption 2's shrinking-jump asymptotics; fixed-jump settings, including the paper's own simulations, have no proven universal penalty, so the headline claim is narrower than stated.","rationale":"The reader's weakest assumption identifies exactly the load-bearing point. The derivation of 6 is a valid result for the stated local asymptotic experiment, but the paper's claim that this is the AIC penalty for discontinuous segmented regression in general relies on Assumption 2 being the right asymptotic regime. The paper explicitly adopts and defends Assumption 2, so this is not an internal contradiction; it is a scope limitation. The fixed-jump simulations are the natural finite-sample check, but because they do not vary the penalty or use shrinking jumps, they do not test the Assumption 2 regime or the optimality of 6. I also agree with the reader that the continuous-case simulation parameters appear inconsistent with continuous truth (for example, Table 2 rows have nonzero jumps at tau*=0.5), which weakens the numerical evidence for the 2-penalty, but this does not invalidate the theoretical Theorem 3. The verdict should remain conditional pending a penalty-sensitivity analysis and corrected simulations.","tokens_in":28066,"tokens_out":25995,"duration_ms":312264,"concrete_test":"Using the Table 3 fixed-jump designs (n=100,200,300), simulate the model and compute the Monte Carlo mean KL divergence between the true distribution and the fitted AIC-selected model for discontinuous change-point penalties c=0,0.5,1,...,12, keeping the theta penalty at 2 and the continuous-change-point penalty at 2. If the c minimizing mean KL is not within roughly 1 of 6, then the fixed-jump setting does not support the Assumption 2-derived penalty 6.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5 derives E[climit]=3m+p(m+1) and hence the AIC penalty 6m+2p(m+1) only under Assumption 2, which requires true jumps to shrink as alpha_n^{-1/2} with alpha_n = o(n). This is a local alternative sequence: as n grows, the true model converges to continuity. For a fixed discontinuous model, the jump does not shrink; the change-point estimator is n-consistent, and the asymptotic bias in the KL risk depends on the jump size and error distribution, not on the universal Brownian constant 3 per change-point. The paper's Section 5 simulations use fixed jumps (e.g., Table 3), so they cannot confirm the 6 penalty in the regime where it was derived; they only show that 6 happens to be reasonable for the particular jump sizes tried. The abstract and Section 7 present 6 as the penalty for discontinuous segmented regression without this caveat, overstating the scope. The continuous-case Theorem 3 is not vulnerable in the same way, since its penalty 2p(m+1) is the regular-model dimension count.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives Akaike information criteria for segmented regression models in two settings: regression functions that are continuous at the change-points and those that are discontinuous. In the continuous case, the authors prove that the change-point estimator is asymptotically normal and that the AIC penalty for each change-point is 2, so the total penalty is 2p(m+1). In the discontinuous case, following the change-point asymptotics of Ninomiya (2015), they show under a local shrinking-jump assumption (Assumption 2) that the per-change-point penalty is 6, giving a total penalty of 6m+2p(m+1). The paper presents Monte Carlo experiments and a real-data analysis on COVID-19 infection counts, arguing that the proposed AIC yields lower Kullback-Leibler divergence than BIC and than a naive AIC, and that model selection can differ between continuous and discontinuous formulations.","tokens_in":28271,"tokens_out":10460,"duration_ms":118429,"significance":"If the theoretical results hold, the paper makes a substantive contribution by providing a KL-optimal model selection criterion for segmented regression, extending earlier work on change-point AIC to generalized linear and nonlinear piecewise-continuous models. The distinction between penalties of 2 and 6 for continuous versus discontinuous change-points is practically important for fields such as epidemiology, where both types of models are used. A strength of the paper is that the penalties are derived, not tuned: the proofs of the key bias evaluations (Theorems 3 and 5) are given in the text and appendix, and the asymptotic normality for the continuous case is extended beyond the earlier linear-setting results. However, the numerical validation has a serious flaw in the continuous-case simulation design, and the scope of the discontinuous-case penalty is narrower than the abstract suggests.","major_comments":[{"comment":"The true parameter vectors used for the 'continuous case' simulations do not satisfy the continuity condition at the change-points. For example, Table 2 lists τ* = 0.5 and θ* = (0, 0, 2.5, -1.25); at x = 0.5 the left-hand limit is 0 and the right-hand limit is 2.5 - 1.25·0.5 = 1.875, so the generated data are discontinuous. The same inconsistency appears in the continuous columns of Table 1 and throughout Table 4. Thus the numerical experiments labeled 'continuous' are actually fitting a discontinuous model and cannot validate the continuous-case AIC (9). The authors must correct the true parameter vectors (for instance by imposing θ1^{[k+1]} = θ1^{[k]} + (θ2^{[k]} - θ2^{[k+1]}) τ^{[k]*}) and rerun the continuous-case experiments.","section":"Section 5, Tables 1, 2, and 4"},{"comment":"The penalty of 6 per discontinuous change-point is derived under the local asymptotics of Assumption 2, where the true jump size shrinks as α_n^{-1/2} with α_n = o(n). For fixed jump magnitudes, the asymptotic bias of the change-point estimator is not a universal constant; it depends on the jump size and error distribution. The numerical experiments in Section 5 use fixed jumps, so they do not test the 6 penalty in the regime in which it was derived. The abstract and Section 7 present 6 as 'the' penalty for discontinuous segmented regression without this caveat. I recommend qualifying the claim explicitly and either adding simulations under shrinking-jump designs or providing a heuristic or additional argument for the fixed-jump setting.","section":"Section 4, Theorem 5 and Assumption 2; Section 7"}],"minor_comments":[{"comment":"The stated penalty for continuous-case AIC naive, 8m+4, appears inconsistent with the model's parameter dimension p=2: assigning a penalty of 6 to each of m change-points plus 2p(m+1) for the regression parameters gives 4(m+1)+6m = 10m+4. Please check the formula and the values used in the tables.","section":"Section 5, first paragraph"},{"comment":"The two rows for n=300 with θ* = (0,0,1.2,0) and θ* = (0,0,1.8,0) are identical to the corresponding n=100 rows, which is likely a copy-paste error; the correct n=300 results should be reported.","section":"Table 3"},{"comment":"The evaluation E[sup_s V(s) - Ṽ(argmax_s V(s))] = 3/2 is cited from Ninomiya (2015) but not reproduced; including this calculation in the appendix would make the paper more self-contained, particularly because the definition of the bias differs slightly from that reference.","section":"Section 4, proof of Theorem 5"},{"comment":"There are several typographical errors: 'AIC navie' in Section 6, 'where where' near equation (10), and inconsistent spelling of 'naive'.","section":"Throughout"},{"comment":"The statement that 'it is more natural that they should be continuous unless there is some reason' is an interpretive remark; consider clarifying that this is a subject-matter preference, not a statistical conclusion from the model comparison.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The continuous-case simulation design is fundamentally flawed: the reported true parameter vectors violate the continuity constraint, so the corresponding Monte Carlo results cannot support the paper's claims. This is a fixable but serious issue. The Assumption 2 scope issue also warrants attention. I recommend major revision and a careful rerun of all numerical experiments with corrected continuous data-generating processes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper's theoretical contribution is real: it proves asymptotic normality for continuous change-point estimators in a generalized GLM segmented regression, then derives the AIC bias correction, recovering a penalty of 2 for continuous change-points and 6 for discontinuous ones. The continuous-case penalty 2 was used intuitively before, and the 6 came from Ninomiya (2015), so the novelty is the formal derivation in a broader model and the clean separation between the two cases. The proof structure follows standard change-point asymptotics and, as far as I can tell, holds up.\n\nThe soft spots are mostly in the empirical section, but one is conceptual. The penalty 6 is derived under Assumption 2, where true jumps shrink with sample size. For fixed jumps, which is what practitioners and the paper's own simulations use, no universal penalty is proven. The abstract and Section 7 state the 6 penalty as 'the' penalty for discontinuous segmented regression without this caveat, overstating the scope. That is a real qualification, not a fatal flaw.\n\nThe numerical section has three fixable problems. The continuous-case parameter vectors in Tables 2 and 4 are inconsistent: for example, with τ=0.5 and θ=(0,0,2.5,-1.25), the left and right segment functions do not meet, so the data generating process is discontinuous despite being labeled continuous. That undercuts those simulation comparisons. Table 3 duplicates the n=100 rows under n=300. There are no Monte Carlo standard errors, and no code or data are provided. None of this refutes the theory, but it means the empirical claims cannot be trusted as-is.\n\nThe paper is for statisticians working on segmented regression and model selection; epidemiologists using joinpoint regression may also benefit, though they would need a cleaned version. It deserves a serious referee: the theoretical result is new and likely correct, and the limitations are repairable. I would not reject it; I would ask for revision that fixes the simulations, adds error bars, and restates the scope of the discontinuous-case penalty.","headline":"The paper's theoretical AIC derivation for continuous and discontinuous change-points is real, but the simulations are sloppy and the scope of the 6 penalty is overstated beyond the shrinking-jump regime.","tokens_in":28813,"tokens_out":2280,"would_cite":false,"duration_ms":26101,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J02","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"For segmented regression, the AIC penalty is 2 per continuous change-point and 6 per discontinuous one, derived from the original Kullback-Leibler definition of AIC.","keywords":["AIC","segmented regression","change-point analysis","joinpoint regression","model selection","piecewise regression","Brownian motion","asymptotic theory"],"falsifier":"Simulate discontinuous segmented regression with fixed, non-shrinking jump sizes over a grid of sample sizes and estimate the exact bias $E[c_{\\mathrm{limit}}]$ by Monte Carlo: under the paper's Assumption 2 the bias should approach $3m+p(m+1)$, so any systematic departure that grows with jump magnitude would falsify the use of penalty 6 outside the local-asymptotics regime.","tokens_in":1966,"feed_emoji":"📈","tokens_out":2907,"duration_ms":128780,"temperature":0.7,"pith_summary":"This paper derives an Akaike information criterion for segmented regression models from the original definition of AIC: the criterion estimates the Kullback-Leibler divergence between the true distribution and the fitted model by adding an asymptotic bias correction to the maximized log-likelihood. The central result is that the bias equals $p(m+1)$ when the regression function is continuous at the change-points and $3m+p(m+1)$ when it is discontinuous, giving penalties $2p(m+1)$ and $6m+2p(m+1)$. In concrete terms, a continuous change-point costs the same as one ordinary parameter, 2, while a discontinuous change-point costs 6. This matters because only BIC-type criteria with model-selection consistency have been available, and those impose the same penalty on both kinds of change-points; the authors show that their AIC tends to produce smaller divergence than BIC in simulations, and that on UK COVID-19 new-infection data the continuous-versus-discontinuous choice changes the estimated trend.","feed_headline":"Segmented regression AIC: 2 per smooth break, 6 per jump","feed_subtitle":"Derived from the Kullback-Leibler definition of AIC, the new penalty separates continuous from discontinuous breakpoints on trend data.","key_machinery":"The machinery is the asymptotic-bias identity $c_{\\mathrm{limit}} = \\mathrm{dlim}_{n\\to\\infty}\\{\\sum_i g(\\hat\\xi;x_i,y_i)-\\sum_i g(\\hat\\xi;\\tilde x_i,\\tilde y_i)\\}$, the difference between the maximized log-likelihood and the same quantity at an independent copy, whose expectation supplies the AIC penalty. In the continuous case the work is done by the asymptotic normality of $\\hat\\theta$ and $\\hat\\tau$, which makes each change-point estimator a linear function of the score. In the discontinuous case the work is done by the two-sided random walk $Q^{(k)}(\\tau^{(k)})$ with negative drift; under Assumption 2 this converges to a two-sided Brownian motion $\\sigma W_s-\\sigma^2|s|/2$, a Gaussian process with independent increments and variance $|s|$ on both sides of zero. The expected value of $\\sup_s V(s)-\\tilde V(\\arg\\sup_s V(s))$ is 3 for each change-point, and the factor 2 in AIC doubles it to the penalty 6.","core_discovery":"Under Assumption 1, for a generalized linear segmented model, the maximum likelihood estimators are consistent; when the regression function is continuous at the breakpoints, $\\sqrt{n}(\\hat\\theta-\\theta^*)$ and $\\sqrt{n}(\\hat\\tau-\\tau^*)$ are asymptotically normal, so the change-point parameters behave like regular parameters and the asymptotic bias of the maximum log-likelihood is $p(m+1)$. When the regression function is discontinuous, the change-point estimators instead converge through a two-sided random walk, and under Assumption 2—true jump sizes of order $\\alpha_n^{-1/2}$ with $\\alpha_n=o(n)$—the bias becomes $3m+p(m+1)$. The proposed criteria are therefore $-2\\log L(\\hat\\xi)+2p(m+1)$ for continuous segmented regression and $-2\\log L(\\hat\\xi)+6m+2p(m+1)$ for discontinuous segmented regression; for a mixture of $m_1$ continuous and $m_2$ discontinuous change-points the penalty is $2p(m+1)+6m_2$.","pith_inferences":["If the same Brownian-argmax mechanism carries other non-regular parameters selected by optimizing over a nuisance domain—threshold indicators, knot locations, regime switches—those parameters should also carry penalties larger than 2 when the target is predictive divergence.","The paper's own Table 1 shows the discontinuous bias approximation degrades as jump sizes grow, so a testable extension is a jump-size-dependent penalty that reduces to 6 under local asymptotics but grows for fixed, large jumps.","A practical implication the authors do not spell out: trend-surveillance software should report both the continuous and discontinuous AIC values before declaring a structural break, because the 6-versus-2 difference changes what counts as evidence.","The same derivation is likely to extend to partial-likelihood and survival settings, but that extension is not established by this paper."],"forward_implications":["In continuous joinpoint regression, the standard AIC penalty formula $2p(m+1)$ is valid, so treating change-point locations as ordinary parameters is justified when the goal is minimizing predictive divergence.","In discontinuous segmented regression, applying the ordinary penalty 2 to each change-point will overfit; the correct per-jump penalty is 6, so naive AIC is systematically too liberal.","Using the extended penalty $2p(m+1)+6m_2$, model selection can simultaneously determine the number of change-points and whether each one is continuous or discontinuous.","Because BIC-type criteria target consistency rather than divergence, AIC and BIC will sometimes select different models; the paper's COVID-19 analysis shows such disagreement on real data.","The gap between the continuous penalty 2 and discontinuous penalty 6 means that deciding whether a structural break is a jump or a smooth slope change is itself a model-selection problem with a well-defined AIC answer."],"supporting_citations":[{"why":"Derives AIC for the basic change-point model and computes the Brownian-motion expectation that yields the penalty 6, the method Section 4 adapts.","marker":"Ninomiya (2015)"},{"why":"Proves asymptotic normality for change-point estimators in segmented multiple regression, which Section 3 extends to generalized nonlinear models to get penalty 2.","marker":"Kim and Kim (2008)"},{"why":"Proposes a BIC-type criterion for segmented multivariate regression and establishes model-selection consistency, the baseline the paper contrasts with divergence minimization.","marker":"Liu et al. (1997)"},{"why":"Derives a modified BIC for change-point models and motivates different penalties for continuous versus discontinuous breaks, the intuitive starting point the paper makes rigorous.","marker":"Zhang and Siegmund (2007)"},{"why":"Provides the local-asymptotics and discontinuous change-point theory, including the shrinking-jump assumption used for the penalty 6.","marker":"Csörgő and Horváth (1997)"},{"why":"Supplies the COVID-19 piecewise-linear quantile trend model and the real dataset reanalyzed in Section 6 to compare selection criteria.","marker":"Jiang et al. (2022)"},{"why":"Introduces the joinpoint regression package and model that define the continuous case the paper targets.","marker":"Kim et al. (2000)"}],"fun_headline_variants":["AIC for segmented regression: 2 per smooth break, 6 per jump","Segmented regression AIC: continuous breaks 2, jumps 6","New AIC for segmented regression: penalty 2 or 6 per break","AIC for segmented regression: smooth breaks 2, jumps 6","Segmented regression model selection: AIC with jump penalty"],"cache_read_input_tokens":30976,"weakest_assumption_plain":"The discontinuous-case penalty of 6 rests on the assumption that the true jump sizes shrink with the sample size; if real-world jumps stay fixed in size, the same AIC correction may not minimize Kullback-Leibler divergence.","fun_headline_variants_meta":{"raw":{"variants":["AIC for segmented regression: 2 per smooth break, 6 per jump","Segmented regression AIC: continuous breaks 2, jumps 6","New AIC for segmented regression: penalty 2 or 6 per break","AIC for segmented regression: smooth breaks 2, jumps 6","Segmented regression model selection: AIC with jump penalty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000695,"raw_usage":{"total_tokens":3203,"prompt_tokens":1062,"completion_tokens":2141,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":678,"completion_tokens_details":{"reasoning_tokens":2045}},"tokens_in":678,"tokens_out":2141,"duration_ms":16031,"temperature":1.0,"reasoning_tokens":2045,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:05:37.429926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate discontinuous segmented regression with fixed, non-shrinking jump sizes over a grid of sample sizes and estimate the exact bias $E[c_{\\mathrm{limit}}]$ by Monte Carlo: under the paper's Assumption 2 the bias should approach $3m+p(m+1)$, so any systematic departure that grows with jump magnitude would falsify the use of penalty 6 outside the local-asymptotics regime.","supporting_citations":[],"review_version":1}