{"id":"85419d8a-bf5f-4583-9e3f-22352cb9509d","arxiv_id":"2506.06491","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new boxplot fence coefficient, derived by equating boxplot fences to Chauvenet's criterion thresholds via normal quantiles, yields sample-size-dependent outlier detection that is claimed to be more robust than either method alone.","lead":"This paper proposes a new outlier detection rule, the Chauvenet-type boxplot, which adjusts the boxplot's fence length according to sample size. The authors claim it keeps the simplicity of Tukey's boxplot while being more robust than Chauvenet's criterion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'exact outside rate' claim is not established: Eq. (9) is derived by replacing sample quartiles with normal population quantiles, and Theorem 1 only gives asymptotic threshold equivalence.","rationale":"The paper's central claim is that the Chauvenet-type boxplot with the fence coefficient in Eq. (9) provides an exact control of the outside rate per observation. The derivation of this coefficient is transparent and the proposed rule is simple, with an available R package and a clear real-data demonstration. However, the exactness claim rests on replacing sample quartiles by their normal population limits, and the only formal support offered is Theorem 1, which is asymptotic. Finite-sample randomness of Q1 and Q3 means the outside probability is not exactly the intended value, and the paper provides no error analysis. This is a load-bearing gap because the headline contribution is precisely the claimed exact rate control. The concern is not that the method is useless; it is that the strongest claim is overstated. A Monte Carlo check of the outside rate across sample sizes would settle whether the error is negligible in practice. The reader already flagged the same approximation as the weakest assumption, and the recommended conditional verdict remains appropriate.","tokens_in":20142,"tokens_out":5598,"duration_ms":62439,"concrete_test":"Run a Monte Carlo study with at least 10^4 replications for pure N(0,1) data at n = 10, 30, 72, 100, 500, and 5000. In each replication compute the fences in Eq. (8) with Eq. (9), record the number of sample observations outside, and also estimate the probability that an independent N(0,1) observation falls outside the fences. Compare the average false-positive count and the outside probability with the claimed values of 0.5 and 0.5/n. If the average count is not close to 0.5 and does not approach 0.5 as n grows, the word 'exact' should be replaced by an asymptotic or approximate statement; if it is close, the remaining issue is only the unquantified finite-sample bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 derives k_n^Chau = Phi^{-1}(1 - 0.25/n)/1.35 - 0.5 by setting Q1 + Q3 approximately equal to 2 times the sample mean and Q3 - Q1 approximately equal to 1.35 times the sample standard deviation, then equating the resulting boxplot fences to the Chauvenet interval. This is a population-quantile heuristic: in finite samples Q1 and Q3 are random, so the actual probability that a new observation falls outside the fences in Eq. (8) depends on the joint distribution of the sample quartiles and is not exactly 0.25/n per tail. The paper's Theorem 1 is asymptotic and establishes only that the two threshold values have ratio tending to 1, not that the per-observation outside rate is controlled at the claimed level. The same overstatement applies to the original Chauvenet interval in Eq. (7), because the sample mean and standard deviation are estimated from the same data, so even there the expected number of outside observations is not exactly 0.5 for every n. Thus the central claim of an 'exact control of the outside rate per observation' is unsupported as stated; what is actually supported is an asymptotic approximation whose finite-sample error is unquantified. The Section 4.1 simulation is a single-seed illustration and does not supply the replication needed to estimate that error.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sample-size-dependent boxplot fence coefficient, k_n^Chau = Phi^{-1}(1 - 0.25/n)/1.35 - 0.5, obtained by equating the boxplot fences to Chauvenet's criterion under normal-population approximations for the quartiles. It claims that this 'Chauvenet-type boxplot' provides an exact control of the outside rate per observation, is asymptotically equivalent to Chauvenet's criterion (Theorem 1), has a 25% breakdown point, and performs well in simulations and in a Hong Kong civil service pay-adjustment example. A non-normal extension using estimated quantile functions is also proposed, and an R package 'ChauBoxplot' is released on CRAN.","tokens_in":20566,"tokens_out":8301,"duration_ms":85357,"significance":"If rigorously supported, the proposal would fill a practical gap: it offers an explicit, simple, sample-size-adjusted, and quartile-based robust boxplot. The formula in Eq. (9) is parameter-free and easy to implement, the accompanying R package is a concrete deliverable, and the paper is clearly written. The main theoretical contribution, however, is currently an asymptotic approximation, not the claimed exact finite-sample control; this distinction is load-bearing and needs to be corrected before the paper can be recommended for publication.","major_comments":[{"comment":"The central claim that the coefficient in Eq. (9) 'provides an exact control of the outside rate per observation' is not established. The derivation in Section 3.2 assumes Q1 + Q3 ≈ 2Xbar and Q3 − Q1 ≈ 1.35S, which are normal-population identities (or asymptotic approximations), not equalities that hold for a finite sample. With random Q1 and Q3, the probability that a normal observation falls outside the fences in Eq. (8) is a function of the joint distribution of these sample quartiles and is not exactly 0.25/n per tail. Theorem 1 establishes only that the thresholds are asymptotically equivalent in ratio; it does not imply exact finite-sample control. The same caveat applies to the original Chauvenet interval in Eq. (7), since Xbar and S are estimated from the same data. The abstract, Section 3.2, and Section 6 should replace 'exact control' with a qualified statement about an asymptotic approximation, or a genuine finite-sample result with quantified error should be provided.","section":"§3.2, Eq. (9); also abstract and §6"},{"comment":"The simulation evidence for normal data is a single run with set.seed(1863) and two fixed contaminating values (5 and 6). It contains no Monte Carlo replication, so the reported counts cannot substantiate the claims of 'extremely well regardless of the sample size' or of exact control of the outside rate. I recommend reporting averages and standard errors over many replications for each n of (i) the number of genuine observations flagged as outliers, (ii) the number of contaminating observations detected, and (iii) the empirical outside rate for uncontaminated normal samples, which would allow a direct check of the claimed 0.25/n per-tail rate.","section":"§4.1"},{"comment":"The non-normal extension is also supported only by a single generated data set of size 50,000, and the assertion that the method 'can perfectly do its job' is stronger than the evidence. The coefficients in Eq. (12) are derived by replacing sample quartiles with F̂^{-1}(0.25) and F̂^{-1}(0.75), so the finite-sample outside rate again depends on estimation error in the parameters and quartiles, which is not quantified. In addition, the t-distribution estimator ν̂ = 2S²/(S²−1) is undefined when S² < 1. These limitations should be acknowledged and the claims scaled back accordingly.","section":"§5.2, Eqs. (12)–(14)"}],"minor_comments":[{"comment":"Typos: 'standard derivation' should be 'standard deviation'; this error appears in the abstract and in the introduction.","section":"Abstract and §1"},{"comment":"The expression 'Φ−1(1 − α/2)1/n)' has a misplaced parenthesis; it should be 'Φ−1((1 − α/2)1/n)'.","section":"§2.3"},{"comment":"'ploted' should be 'plotted'; in §2.4, 'expect for' should be 'except for'; in §4.2, 'Fore more details' should be 'For more details'.","section":"§3.3"},{"comment":"The statement 'kChau_n = 3 when n = 217, 282' is ambiguous; if two sample sizes are intended, they should be written as 'n = 217 and n = 282'.","section":"§6"},{"comment":"The sentence that the fences 'remain unchanged' when fewer than 25% of observations are outliers is imprecise: quartile-based statistics have a positive breakdown point, but they can still move under contamination below that fraction. A precise definition of the finite-sample replacement breakdown point for the fences would be helpful.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The proposal is simple and potentially useful, but the 'exact control' language is a pervasive overstatement that affects the abstract, introduction, Section 3.2, and conclusion. The simulation section is illustrative rather than evidential. I would support publication after the claims are scaled back to asymptotic or approximate control and after a proper Monte Carlo evaluation is added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is a single formula: k_n = Phi^{-1}(1 - 0.25/n)/1.35 - 0.5, which turns Tukey's boxplot into a sample-size-aware rule that asymptotically matches Chauvenet's criterion. That is a neat, practical idea, and I don't know of another paper that has written down this exact coefficient. The literature review of sample-size-adjusted boxplots is genuinely useful, and the R package on CRAN is a real plus for practitioners. The toy example showing Chauvenet's criterion failing on contaminated data is well chosen, and the non-normal extension via a known parametric distribution is reasonable.\n\nThe soft spot is the central claim. The paper says the Chauvenet-type boxplot provides 'exact control of the outside rate per observation.' That is not what the derivation shows. The derivation in Section 3.2 replaces sample quartiles with normal population quantiles (Q1+Q3 approx 2*Xbar, Q3-Q1 approx 1.35*S), equates the resulting fences to the Chauvenet interval, and then solves for k_n. That is a population-quantile heuristic, not an exact finite-sample result. Theorem 1 only establishes asymptotic equivalence between the two threshold values, not exact control of the per-observation outside rate. The finite-sample error from using random quartiles is never quantified. The same criticism actually applies to the original Chauvenet interval, since Xbar and S are estimated from the data, so 'exact' is too strong there too.\n\nThe simulation is also thinner than it looks: single seeds, no replication, no standard errors. The real-data analysis is anecdotal, which is fine for illustration but not for the 'performs extremely well regardless of sample size' conclusion.\n\nNone of this is fatal. The paper is transparent about its construction, and the proposed rule is simple and likely to work reasonably well in practice. The fix is to tone down 'exact' to 'asymptotic' or 'approximate,' add a few lines explaining the finite-sample approximation, and run Monte Carlo with replications to show the actual outside rate is close to 0.25/n per tail. That is refereeable work.\n\nI would send this to review. It is a practical applied-statistics contribution with real code, and the overclaim is correctable without changing the method.","headline":"A genuinely simple and useful boxplot tweak, but the 'exact control' claim is not backed by the math as written; the paper deserves review with revisions, not desk rejection.","tokens_in":725,"tokens_out":824,"would_cite":false,"duration_ms":24395,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G30","62G35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a sample-size-dependent boxplot fence coefficient that flags half a genuine point per normal sample while staying resistant to 25% contamination.","keywords":["box-and-whisker plot","Chauvenet's criterion","Chauvenet-type boxplot","fence coefficient","outlier detection","sample size","robust statistics","interquartile range"],"falsifier":"Simulate standard normal samples of sizes $n=20,72,500,5000$ many times, and count how many genuine observations fall outside $Q_1-k_n IQR$ and $Q_3+k_n IQR$. The paper's exact-control claim predicts an average of 0.5 flagged observations per sample at every $n$; observed averages differing from 0.5 by more than simulation error, particularly at small $n$, would refute it.","tokens_in":19949,"feed_emoji":"📊","tokens_out":18896,"duration_ms":153376,"temperature":0.7,"pith_summary":"Tukey's boxplot places its whiskers at 1.5 interquartile ranges beyond the quartiles, regardless of sample size, so with large normal samples it flags many genuine points as outliers. Chauvenet's criterion adjusts for sample size but uses the sample mean and standard deviation, so a few large outliers can widen its interval and hide real anomalies. This paper fuses the two: keep the boxplot's quartile-based fences, but set the fence coefficient to $k_n=\\Phi^{-1}(1-0.25/n)/1.35-0.5$. The result, the Chauvenet-type boxplot, is claimed to flag on average half a genuine observation per normal sample of any size, and to leave its fences unchanged when fewer than 25% of the observations are contaminated. That would give practitioners a simple, sample-size-aware alternative to both classic tools.","feed_headline":"New boxplot rule flags half a genuine point per normal sample","feed_subtitle":"Quartile fences with a sample-size knob: half a false flag per sample, even with 25% contamination.","key_machinery":"The central object is the fence coefficient $k_n^{\\mathrm{Chau}}=\\Phi^{-1}(1-0.25/n)/1.35-0.5$, a sample-size-dependent multiplier for the interquartile range in the standard boxplot fences. It does the work by converting Chauvenet's boundary $c_n=\\Phi^{-1}(1-0.25/n)$ into an IQR multiplier through the normal-data relations $Q_3-Q_1\\approx 1.35 S$ and $(Q_1+Q_3)/2\\approx \\bar{X}$. A single coefficient then controls whisker length: below Tukey's 1.5 for $n<72$, equal at $n=72$, and growing with $n$ afterward, matching the per-observation false-outside rate that Chauvenet's criterion targets.","core_discovery":"For normal data, the paper defines new fences $LF_n=Q_1-k_n\\times IQR$ and $UF_n=Q_3+k_n\\times IQR$ with $k_n=\\Phi^{-1}(1-0.25/n)/1.35-0.5$, obtained by equating the boxplot fences to Chauvenet's thresholds $\\bar{X}\\pm c_n S$ under the normal identities $IQR\\approx 1.35 S$ and $(Q_1+Q_3)/2\\approx \\bar{X}$. It claims this coefficient gives an exact control of the outside rate per observation, that the resulting outlier region is asymptotically equivalent to Chauvenet's criterion, and that because the fences are built on quartiles rather than the sample mean and standard deviation, the rule has a 25% breakdown point. The paper also extends the construction to non-normal data by replacing normal quantiles with quantiles from an estimated parametric distribution, giving separate lower and upper fence coefficients for skewed or heavy-tailed data.","pith_inferences":["Beyond the paper: the same substitution of quartile-based location and scale for the mean and standard deviation could produce robust, sample-size-aware versions of any constant sigma-clipping rule, not only Chauvenet's criterion.","Beyond the paper: the claimed exact outside-rate control is a population-quantile design target; finite samples will show random fluctuation around 0.5 flagged genuine observations because the quartiles are random, so exactness is best read as the rule's intent rather than a per-sample guarantee.","Beyond the paper: the non-normal extension assumes the parametric family is known; feeding a nonparametric quantile estimate into the same fence formulas would remove that assumption and can be tested on the same simulation designs."],"forward_implications":["For normal data, the expected number of genuine observations flagged is about 0.5 regardless of sample size, whereas Tukey's fixed 1.5 fences flag roughly 0.7 percent of a large normal sample.","Because the fences use quartiles rather than the mean and standard deviation, the rule can detect outliers that classical Chauvenet misses when contamination inflates the standard deviation.","The new fence coefficient equals Tukey's 1.5 when the sample size is 72; below that the whisker is shorter, increasing detection power for small samples, and above that it lengthens to control false flags.","For non-normal data, replacing normal quantiles with quantiles from an estimated distribution gives separate lower and upper fence coefficients, which reduces false positives for skewed and heavy-tailed data.","The rule offers a single, sample-size-aware boxplot that can serve in place of both Tukey's boxplot and Chauvenet's criterion."],"supporting_citations":[{"why":"Defines the boxplot, the quartile-based fences, and the midhinge estimate that the new rule reuses.","marker":"Tukey (1977)"},{"why":"Establishes the criterion and the normal quantile threshold at probability 1 - 0.25/n that the new fence coefficient matches.","marker":"Chauvenet (1863)"},{"why":"Supplies the normal-data approximation that the interquartile range equals about 1.35 times the standard deviation, used to convert boxplot fences into mean-standard-deviation units.","marker":"Higgins et al. (2019)"},{"why":"Introduces the some-outside-rate control and outlier labeling rules that motivate sample-size-dependent fence coefficients.","marker":"Hoaglin et al. (1986)"},{"why":"Provides the exact-some-outside-rate boxplot and numerical fence coefficients that the new rule is compared against.","marker":"Sim et al. (2005)"},{"why":"Gives the asymptotic-fence boxplot and its smoothed sample-size-dependent coefficient, another comparison method.","marker":"Banerjee and Iglewicz (2007)"},{"why":"Supplies the definition of breakdown point used to state the new rule's 25 percent resistance.","marker":"Huber (2004)"},{"why":"Provides the adjusted boxplot for skewed data used as a benchmark in the non-normal simulations.","marker":"Hubert and Vandervieren (2008)"}],"fun_headline_variants":["Boxplot and Chauvenet merge into sample-size-aware fences","Exact outside-rate control in a quartile-based boxplot","Robust boxplot fences with a sample-size knob","Chauvenet-type boxplot: robust and sample-size adjusted","New fences hold outside rate constant per observation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation treats the population normal identities $IQR=1.35\\sigma$ and $(Q_1+Q_3)/2=\\mu$ as if they hold for sample quartiles, so if sample quartiles deviate from those relations, the per-observation outside rate is only approximately controlled.","fun_headline_variants_meta":{"raw":{"variants":["Boxplot and Chauvenet merge into sample-size-aware fences","Exact outside-rate control in a quartile-based boxplot","Robust boxplot fences with a sample-size knob","Chauvenet-type boxplot: robust and sample-size adjusted","New fences hold outside rate constant per observation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1705,"prompt_tokens":1029,"completion_tokens":676,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":645,"tokens_out":676,"duration_ms":7150,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:54:05.737712+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate standard normal samples of sizes $n=20,72,500,5000$ many times, and count how many genuine observations fall outside $Q_1-k_n IQR$ and $Q_3+k_n IQR$. The paper's exact-control claim predicts an average of 0.5 flagged observations per sample at every $n$; observed averages differing from 0.5 by more than simulation error, particularly at small $n$, would refute it.","supporting_citations":[],"review_version":1}