{"id":"f976d460-fb41-49e7-bf3e-6ea4ba846fae","arxiv_id":"1908.07456","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The Cox partial likelihood and Breslow estimators have bounded moments of any order on high-probability events, under exponential moment conditions on the covariates.","lead":"This paper proves new moment bounds for the two standard estimators in Cox survival regression: the maximum partial likelihood estimator and the Breslow baseline hazard estimator. The bounds fill a gap in the literature that matters for analyzing shape constrained baseline hazard estimators and confidence bands.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's proof leaves an unbounded random-β* factor in the Cauchy-Schwarz step; under (A1)-(A4) it need not have finite mean.","rationale":"I read the paper as aiming to prove uniform moment bounds for the MPLE and Breslow estimator under (A1)-(A4). Theorems 1 and 2 state the bounds on events with probability tending to one, and Lemma 3 gives an unconditional empirical-process bound. The reader's weakest_assumption focuses on the strong exponential-moment condition (A3), which is a limitation but not a proof flaw. My stress-test identified a more targeted internal gap in the proof of Theorem 2, specifically the Cauchy-Schwarz step bounding term (13). The random β* appears in the exponential and is not controlled by (A4) uniformly over the sample space. The constructed example shows that under the paper's own assumptions, the second factor can have infinite expectation, so the proof as written fails to establish (13). This is more load-bearing than the entropy-rate error in Lemma 3, which is easily repairable because the correct entropy of monotone functions still yields a finite entropy integral. The central results may still be true, and the proof gap can plausibly be fixed by adding an event that keeps β* in the (A4) neighborhood and by using uniform-in-β moment bounds. Therefore the verdict should remain CONDITIONAL, matching the reader's overall assessment, but my identified concern differs from the reader's weakest assumption. The paper deserves credit for a useful contribution and a mostly rigorous derivation, but the random parameter in the Breslow proof needs explicit handling.","tokens_in":11498,"tokens_out":23650,"duration_ms":231059,"concrete_test":"Analyze the scalar model β0=1, Z=−Y, Y with density ∝ y^{-3} on [1,∞), and censoring chosen so (A1)-(A4) hold (e.g., administrative censoring at τG=1 with P(T=1)>0). Compute the Cauchy-Schwarz second factor E[(1/n)Σ Y_i^2 e^{-β*Y_i}]^p]. Show that for any p≥1, E[Y^{2p}e^{p a Y}]=∞ for any a>0, and that P(β*<0)>0 for finite n because β̂ has positive density on (−∞,0). Hence the unconditional expectation is infinite. Alternatively, run Monte Carlo for moderate n and check that the empirical average of (1/n)Σ Y_i^2 e^{-β*Y_i} is dominated by rare β*<0 samples and does not stabilize; then include the event {β*≥β0/2} in An and verify the repaired bound becomes finite.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 2, the bound for term (13) applies Cauchy-Schwarz to obtain a second factor E[(1/n)Σ Z_i^2 e^{β*'Z_i}]^p], where β* is a random point on the segment from β0 to β̂_n. The paper claims this factor is bounded by the proof of Theorem 1 using (A4). However, (A4) only controls expectations of Z^{2q}e^{qβ'Z} for |β−β0|≤ε, not for an arbitrary random β*. A concrete model satisfying (A1)-(A4) is β0=1, Z=−Y with Y∈[1,∞) having density ∝ y^{-3}. Then E[Z^{2q}e^{qβZ}]=E[Y^{2q}e^{-qβY}]<∞ for all β>0, so the assumptions hold, but E[Y^2e^{aY}]=∞ for any a>0. Since β̂ has a continuous distribution, P(β*<0)>0, and on that event e^{β*Z}=e^{-β*Y} with positive exponent, making the second factor have infinite mean. Thus the stated proof does not establish a finite bound for (13). The theorem might be repairable by intersecting An with an event that keeps β* in the (A4) neighborhood, but this is absent. This is the most load-bearing gap because it directly affects the Breslow estimator moment bound, while the entropy issue in Lemma 3 is a repairable quantitative error.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves uniform L^p moment bounds for the maximum partial likelihood estimator β̂_n and the Breslow estimator Λ_n in the Cox proportional hazards model, on events with probability tending to one, under assumptions (A1)–(A4), with (A3) required for the Breslow result. It also states Lemma 3, a uniform moment bound for the rescaled empirical process n^{1/2} sup_t |Φ_n(t;β0)−Φ(t;β0)| under the exponential moment condition (A3). The motivation is to provide tools for L^p global error analysis of shape-restricted estimators of the baseline hazard, where the usual CLTs for β̂_n and Λ_n are not enough. Theorem 1 controls the regression coefficient; Theorem 2 controls the Breslow estimator; Lemma 3 is auxiliary but of independent interest.","tokens_in":11775,"tokens_out":23185,"duration_ms":211975,"significance":"If the moment bounds are correct, they close a real gap in the Cox model literature: standard asymptotic results give weak convergence but not uniform boundedness of moments of arbitrary order, which is needed for L^p errors of shape-restricted baseline hazard estimators. The paper's high-probability-event formulation is honest and useful, and the proofs combine martingale calculations with empirical-process inequalities in a mostly careful way. The results are plausible and the statements are clean. However, two technical problems in the proofs—the random-β* factor in Theorem 2 and the covering-number claim in Lemma 3—must be repaired before the results can be considered established.","major_comments":[{"comment":"The control of term (13) is not established. After the Cauchy-Schwarz step, the proof requires a uniform bound on E[|n^{-1} Σ Z_i^2 e^{(β*)'Z_i}|^p], where β* is the random point from the mean value theorem. The text says this factor is bounded by the proof of Theorem 1 using assumption (A4), but (A4) only controls sup_{|β−β0|≤ε} E[Z_k^{2q} e^{qβ'Z}] for fixed β in a fixed neighborhood; it does not control an empirical average with a data-dependent β* that can leave that neighborhood on a set of positive probability for finite n. For example, β0=1 and Z=−Y with Y having density proportional to y^{-3} on [1,∞) satisfy (A1)–(A4), but E[Y^2 e^{aY}] = ∞ for every a>0, so the cited argument cannot yield a finite bound. The proof needs either an additional event forcing |β*−β0|≤ε (with the complement handled separately) or a different bound exploiting the event A2_n. As written, the proof of Theorem 2 is incomplete.","section":"§3, proof of Theorem 2, term (13)"},{"comment":"The covering-number bound N(ε||F||_{L2(Q)}, F, L2(Q)) ≲ 1/ε is not justified and is in general false for the class of monotone functions. Theorem 2.7.5 in van der Vaart and Wellner gives an upper bound of the form log N_{[]}(ε, F, L2(Q)) ≲ 1/ε, which does not imply N ≲ 1/ε; the covering number can grow exponentially in 1/ε. The proof of Lemma 3 (and the analogous argument for term (11)) should use the logarithmic bound instead. The entropy integral J(1,F) is still finite under this corrected bound, so the conclusions of Lemma 3 and of term (11) remain plausible, but the displayed inequality as written is wrong.","section":"§3, proof of Lemma 3 and term (11)"}],"minor_comments":[{"comment":"The phrase 'of the the Breslow estimator' contains a duplicated article; it should read 'of the Breslow estimator'.","section":"Abstract"},{"comment":"The definitions of D1_n(t;β) and D2_n(t;β) use e^{β'_0 Z_i}; the derivative with respect to β should contain e^{β' Z_i}.","section":"§3, proof of Theorem 1, equations (5)–(6)"},{"comment":"In the Cauchy-Schwarz display, the first factor contains e^{β'_0 Z_i} while the second contains e^{(β*)' Z_i}; both should contain e^{(β*)' Z_i} to match the preceding bound.","section":"§3, proof of Theorem 2, term (13)"},{"comment":"The sentence 'From Theorem 2.7.5 and in van der Vaart and Wellner (1996)' has an extra 'and in', and several integrals are written over R×Rp rather than R×R^d.","section":"§3, proof of Lemma 3"},{"comment":"The convergence sup_t |Φ_n(t;β*)−Φ(t;β0)| → 0 is asserted by analogy with a fixed-β lemma, but β* is random; a uniform-in-β argument over a neighborhood of β0 should be spelled out.","section":"§3, proof of Theorem 2, definition of A2_n"},{"comment":"Assumption (A3) requires finite exponential moments of every order and is substantially stronger than the moment conditions in Andersen and Gill (1982); a remark on how restrictive this is in applications would be helpful.","section":"Assumptions"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the central results are likely correct. The main problem is the random-β* factor in the proof of Theorem 2; this is a genuine proof gap rather than a cosmetic issue, so I recommend major revision rather than rejection. The authors should also correct the covering-number statement in Lemma 3."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper gives the first uniform moment bounds of all orders for the Cox maximum partial likelihood estimator and the Breslow estimator, on high-probability events. That fills a genuine gap: people working on shape-restricted baseline hazard estimation need these bounds to control Lp errors, and the existing literature mostly stops at convergence in distribution. The results are likely correct, and the proofs use standard empirical process and martingale machinery.\n\nWhat's good: Theorem 1's proof is careful. The martingale representation of the score, combined with Titu's lemma and assumption (A4), gives a clean bound on n^{-1/2}S'(β0). Lemma 3, the unconditional moment bound for the rescaled empirical process sup of Φ_n, is a useful standalone tool with an elementary proof. The motivation is honest, and the self-citations are technical or motivational, not circular.\n\nSoft spots, in proportion. First, the covering number claim in Lemma 3 is quantitatively wrong. The class of monotone functions on R to [0,1] does not have polynomial covering number N ≲ 1/ε; the correct bound is exponential, N ≲ exp(K/ε). The entropy integral still converges because sqrt(log N) ~ ε^{-1/2}, so the lemma's conclusion survives. The citation to Theorem 2.7.5 in van der Vaart and Wellner is misread. Minor and repairable.\n\nSecond, and more important, the proof of Theorem 2 has a gap in the term involving the random β*. After Cauchy-Schwarz, they need a uniform bound on E[(n^{-1}Σ Z_i^2 e^{β*'Z_i})^p], where β* lies on the random segment from β0 to β̂_n. The paper says this is bounded by the proof of Theorem 1 using (A4). But (A4) controls expectations for β in a fixed neighbourhood of β0, not for an arbitrary random β*. The event A_n they construct does not force β* into that neighbourhood. With a heavy-tailed covariate where β0'Z is negative, (A4) can hold while E[e^{β*'Z}] is infinite whenever β* strays below zero. The fix is simple: add an event {|β̂_n − β0| ≤ ε} (or equivalently |β* − β0| ≤ ε) to A_n; consistency gives probability tending to one, and then (A4) applies. As written, the proof is incomplete.\n\nThird, minor: the abstract says 'uniformly bounded moments' without mentioning the events. The theorems are truncated moment bounds on high-probability events. The paper is transparent about this in the main text, but the abstract oversells.\n\nBottom line: this deserves a serious referee. The theorems are credible and needed, Lemma 3 is useful, and the flaws are repairable without changing the substance. I'd send it to review with a request to fix Theorem 2's proof and correct the entropy claim in Lemma 3.","headline":"First uniform moment bounds for Cox estimators fill a real gap; Theorem 2's proof has a fixable gap on a random β* term.","tokens_in":12291,"tokens_out":5103,"would_cite":true,"duration_ms":48220,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62N01","62G20","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the proportional hazards regression model, the rescaled maximum partial likelihood and Breslow estimators have uniformly bounded moments of every order.","keywords":["Cox regression model","maximum partial likelihood estimator","Breslow estimator","uniformly bounded moments","counting process martingale","empirical process","baseline hazard","survival analysis"],"falsifier":"Set up the two-sample balanced Cox model with bounded covariates, unit exponential baseline, and independent censoring chosen so that $P(T = \\tau_G) > 0$, then compute analytically or by high-precision simulation the quantities $\\sup_n E[n^{p/2}|\\hat\\beta_n - \\beta_0|^p]$ for $p=2$ and $p=4$; if any of these is infinite or grows without bound in $n$, Theorem 1 is false.","tokens_in":11278,"feed_emoji":"📊","tokens_out":9498,"duration_ms":91127,"temperature":0.7,"pith_summary":"This paper fills a technical gap in the asymptotic theory of the proportional hazards regression model. It proves that the maximum partial likelihood estimator of the regression coefficients and the Breslow estimator of the cumulative baseline hazard, after centering and multiplying by the square root of the sample size, have $p$-th moments bounded uniformly in $n$, for every $p \\geq 1$. The bounds hold on events of probability tending to one and require, in addition to the usual Cox-model conditions, that the covariates have finite exponential moments in the direction of the true regression parameter. Uniform moment bounds of this kind are what one needs to turn pointwise or sup-norm convergence into convergence of $L_p$ errors, for example when studying shape-restricted estimators of the baseline hazard and constructing confidence bands.","feed_headline":"Cox model estimators have bounded moments of every order","feed_subtitle":"A missing moment bound for partial-likelihood and Breslow estimators underpins global error analysis.","key_machinery":"The argument turns on three objects. The score of the partial likelihood is written as a sum of stochastic integrals of predictable processes with respect to counting-process martingales, so its moments are controlled through its predictable variation process. The empirical process term appearing in the Breslow estimator is treated as a supremum over a class of functions $f_t(u,z) = 1_{u \\geq t} e^{\\beta_0' z}$, whose bracketing numbers grow like $1/\\varepsilon$ and whose envelope is $F(u,z) = e^{\\beta_0' z}$; exponential moment assumptions bound that envelope in $L_{2 \\vee p}(P)$. Finally, the Taylor expansion of the score around $\\beta_0$ is controlled on an event $E_n$ on which the normalized matrix of second derivatives is uniformly close to a nonsingular matrix, so that the inverse stays bounded and the estimation error is dominated by the normalized score.","core_discovery":"On the paper's own terms, the discovery is that the classical asymptotic normality results for the two standard estimators of the Cox model can be upgraded to fully quantitative moment bounds. Theorem 1 shows that under assumptions (A1), (A2), and (A4), with a continuous baseline hazard, there is a sequence of events $E_n$ with $P(E_n) \\to 1$ and a constant $K$ such that $\\limsup_n E[1_{E_n} n^{p/2} |\\hat\\beta_n - \\beta_0|^p] \\leq K$ for every $p \\geq 1$. Theorem 2 shows the analogous uniform bound for $\\sup_{t \\in [0,\\tau_G)} |\\Lambda_n(t) - \\Lambda_0(t)|$ under (A1)-(A4). Lemma 3 supplies an unconditional bound for the empirical process $n^{1/2} \\sup_t |\\Phi_n(t; \\beta_0) - \\Phi(t; \\beta_0)|$ under the exponential-moment assumption (A3). The event restriction is a device to isolate the random fluctuation of the matrix of second derivatives; because the events have probability tending to one, the bounds are sufficient for later convergence-in-distribution arguments.","pith_inferences":["A natural testable extension is to ask whether the exponential moment conditions (A3) and (A4) can be weakened to finite moments of sufficiently high order; the empirical-process peeling argument suggests that some higher-moment analogue should hold, but the present proof genuinely needs the exponential envelope.","The high-probability event restriction could likely be removed if one had direct moment control on the inverse of the observed information matrix; the paper's event device is one way to carry the argument, not an intrinsic obstruction.","The same martingale-plus-empirical-process template should transfer to other partial-likelihood settings, such as stratified or time-dependent covariate versions of the proportional hazards model, where the same missing moment bounds would be needed for global errors."],"forward_implications":["The normalized coefficient error $n^{1/2}(\\hat\\beta_n - \\beta_0)$ is uniformly integrable: its $p$-th moments are bounded for every $p$, not just for $p=1$ or $p=2$.","The sup-norm error of the Breslow estimator over $[0,\\tau_G)$ has bounded moments, so $L_p$ convergence of cumulative-hazard estimation follows at the parametric $\\sqrt{n}$ rate.","These bounds are exactly the ingredient needed to pass from sup-norm or pointwise asymptotics to global $L_p$ errors in shape-restricted baseline hazard estimation, such as isotonic-type estimators.","Because the bounds hold for every $p$ simultaneously, they also cover higher-moment terms in expansions, enabling uniform confidence bands via moment-based arguments.","Lemma 3 stands alone: the rescaled empirical process for the at-risk function $\\Phi_n$ has bounded moments without any event restriction, under only the exponential moment condition."],"supporting_citations":[{"why":"Supplies the large-sample theory that the maximum partial likelihood estimator solves the score equation and is asymptotically normal.","marker":"Tsiatis (1981)"},{"why":"Provides the counting-process martingale framework and the convergence of the normalized matrix of second derivatives to the nonsingular matrix used in the Taylor expansion.","marker":"Andersen and Gill (1982)"},{"why":"Defines the proportional hazards model and the partial likelihood for the regression coefficients.","marker":"Cox (1972)"},{"why":"Gives the martingale representation of the score function and the form of its predictable variation process used to bound its moments.","marker":"Kalbfleisch and Prentice (2002)"},{"why":"Supplies the maximal inequality and entropy bounds that turn the bracketing numbers of the monotone indicator class into a bounded empirical-process constant.","marker":"van der Vaart and Wellner (1996)"},{"why":"Provides the integral representation of the Breslow estimator and the sup-norm rate (4) used in the proof of Theorem 2.","marker":"Lopuhaa and Nane (2013a)"},{"why":"Used for the uniform convergence of the derivative processes $D^1_n$ and $D^2_n$ that appear in the score and information matrix.","marker":"Lopuhaa and Nane (2013b)"},{"why":"The motivating application: global $L_p$ errors of shape-restricted baseline hazard estimators require exactly these moment bounds.","marker":"Durot and Musta (2019)"}],"fun_headline_variants":["Cox estimator moments: bounded at every order","Moment bounds for Cox partial likelihood and Breslow estimators","Cox model: uniform moment bounds for standard estimators","Quantitative moment bounds for Cox regression estimators","All-order moment bounds for Cox model estimators"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing extra condition is that the covariate vector has finite exponential moments in the direction of the true regression parameter, $E[e^{q \\beta_0' Z}] < \\infty$ for every $q \\geq 1$, because without it the envelope $e^{\\beta_0' z}$ of the empirical process class is not controlled enough for the moment bounds; the positivity assumption $P(T = \\tau_G) > 0$ is also indispensable so that $\\Phi$ does not vanish at the end of the observation interval.","fun_headline_variants_meta":{"raw":{"variants":["Cox estimator moments: bounded at every order","Moment bounds for Cox partial likelihood and Breslow estimators","Cox model: uniform moment bounds for standard estimators","Quantitative moment bounds for Cox regression estimators","All-order moment bounds for Cox model estimators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1432,"prompt_tokens":855,"completion_tokens":577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":471,"tokens_out":577,"duration_ms":5399,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:19:52.277057+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set up the two-sample balanced Cox model with bounded covariates, unit exponential baseline, and independent censoring chosen so that $P(T = \\tau_G) > 0$, then compute analytically or by high-precision simulation the quantities $\\sup_n E[n^{p/2}|\\hat\\beta_n - \\beta_0|^p]$ for $p=2$ and $p=4$; if any of these is infinite or grows without bound in $n$, Theorem 1 is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the large-sample theory that the maximum partial likelihood estimator solves the score equation and is asymptotically normal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the counting-process martingale framework and the convergence of the normalized matrix of second derivatives to the nonsingular matrix used in the Taylor expansion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the proportional hazards model and the partial likelihood for the regression coefficients."},{"cited_title":"On the $L_p$-error of the Grenander-type estimator in the Cox model","cited_arxiv_id":"1907.06933","evidence_quote":"The motivating application: global $L_p$ errors of shape-restricted baseline hazard estimators require exactly these moment bounds."}],"review_version":1}