Evidence level: C
阈值定在分布第几分位,不是一个能自由拧的旋钮;它是被样本量锁死的秩序统计量。
一、这个问法预设了一个不存在的旋钮
上一轮把验证顺序倒过来:先看 ≥2 落在第几分位,再决定二分值不值得留。听起来合理,但它默认「距众数多远」可以自由选——想落 85 分位就落 85 分位。
分位位置不是自由参数。95 分位在 n=20 里就是第 19 个秩序统计量,其值由排序后倒数第二个观测单独决定;换个人、换个带位,这个观测换成另一个数,切点就整个搬家。横轴看似连续,实际只有几个格点,每格自带一份由 n 决定的抖动。
二、二分的账先记清
连续变量切成两类会丢信息、掉功效,这是有关系统研究的老结论:多变量回归中把连续预测因子分组被判为「坏主意」(PMID 16217841);整群随机试验里二分连续结局的功效损失可算得出(PMID 22733454);二分的统计性质与连续不可互换(PMID 18389492);连年龄二分后得到的比数比都有偏(PMID 17066378)。
替代路线也已存在:同一批原始研究用分布法可同时给出连续与二分两种估计,且推断一致(PMID 24920271)。方向提示很清楚——不是「二分怎么切更稳」,而是能不能不切。
三、切点偏倚的实测账:主因是样本量
真正量过这件事的一项模拟研究设敏感性、特异性真值都是 84%,用 Youden 指数数据驱动选切点(PMID 18258670)。n=40 时估成约 90%,n=200 时约 86%,偏倚与样本量成反比。
同一段结果里还有一句更该记住:样本量固定时,结果分布形态对该偏倚几乎没有影响。这句直接顶掉「按分布形态分别定线」这个前置任务——至少在这类偏倚上,形状不是自变量。研究还试了三种更稳的做法(假定分布、留一法、平滑 ROC),它们更不易偏,但假设不满足时性能会崩。对应的修补工具是:选完切点后的效应须用收缩加自助法才能恢复置信区间覆盖率(PMID 15160403)。
四、小样本里这两个百分位分辨不出来
还有更基础的问题:这个 n 下,70 与 85 分位能不能分开。
分位数与百分位的置信区间有专门算法(PMID 30591808);参考限本质是对极端百分位的估计,精度由样本量决定,且有现成的最少受试者数算法(PMID 25029084)。落到格点:n=20 时 85 分位只有 3 个观测在其上,95 分位只剩 1 个。两个切点值的区间几乎必然重叠——「放 85 还是放 95」在这个样本量下不是可执行的操作,是同一个模糊地带的两种叫法。
小样本参考区间的实践建议同向:把数值画成点图,用多种方法各算参考限,看是否互相靠得住(PMID 19473330);个体内变异大、强偏态、样本又小时,分布估计尤须谨慎(PMID 21731012)。
五、「双峰」这个前提自己先要过关
「距众数多远」隐含一个形态判断。双峰意味存在天然分界,二分才有道理,但形态判断很脆。
三种常用度量(双峰系数 BC、Hartigan dip、单成分对双成分的 AIC 差)被系统比对过(PMID 22806703):距离影响总强于比例,比例的影响在三种度量间差异极大;偏度会虚高双峰性检测,某些情形引发反常交互;AIC 差极度敏感,几乎把一切分布判成双峰。BC 与 dip 大体一致,但存在若干重要分歧。
意思是:一个偏态单峰分布很可能被读成双峰,于是你以为该用双峰线,其实数据从头到尾只有一个峰。把形态放在推理链最上游,等于用一个判别力有限的判断去支撑全部结论。
六、把横轴换成稀有侧计数 k
改造的入手点:百分位难用,因为它把距离和样本量搅在一起;把横轴换成阈值以上的样本个数 k,两者就分开了。
同一条 15% 经验线,n=100 是 15 人,n=20 只剩 3 人。前者切点稳、分半后每半仍有七八人;后者每半只剩一两人,Jaccard 的分子分母都由这几个观测决定。这解释了为什么同一条经验线在不同研究里表现迥异——它根本不是同一个门槛。
换成 k 后得到一个可检验推论:若稳定性主要由此决定,则四种分布形态的「k → Jaccard」曲线应当收敛。
七、反转与四条判据
反转是:不问「阈值该离众数多远」,改问「稀有侧至少留几个样本才够」。
四条判据:横轴换 k 后四形态最大差 <0.05,则分位与形态皆降为伪变量,规则改写成「稀有侧 k≥X」;换 k 后仍明显分形态(>0.05),则形状真起作用、须按形态分线,但形态判定须先做偏度校正、三种度量并列报;k 固定而 Jaccard 仍随 n 下降,则问题在个体内相关与分半结构,回到分半单位那条线;算出 70 与 85 分位切点的置信区间发现重叠,则该样本量下任何「前 15-20%」经验线都不可执行,只能报 k 与它的实际人数。
证据与局限
已验证: 数据驱动选最优切点致敏感性特异性高估,偏倚随样本量减小放大,n=40 时真值 84% 估成约 90%(PMID 18258670);样本量固定时形态对该偏倚影响很小(同前);选切点后置信区间覆盖不足,须收缩加自助法(PMID 15160403);二分损信息、掉功效、产生有偏比数比(PMID 16217841、PMID 22733454、PMID 18389492、PMID 17066378);偏度虚高双峰检测,三种度量结论不一致(PMID 22806703);极端百分位估计精度由样本量决定,有最少受试者数算法(PMID 25029084、PMID 30591808)。
部分验证: 分布法可在同一批原始研究中同时给出连续与二分估计且一致(PMID 24920271),但属对已发表研究的二次分析,非前瞻验证;小样本参考区间须多方法并列(PMID 19473330)、个体内变异大与强偏态须谨慎(PMID 21731012),均为方法学建议而非本参数的实测。
待验证: 「阈值分位 → Jaccard 中位数」曲线从未在丹田五参数上画过,拐点在 80-85 分位只是待检验假设;换成计数 k 后四形态曲线是否收敛,是本篇核心可检验命题,尚无数据;70 与 85 分位切点置信区间是否实际重叠,也尚未算。
相关阅读
- Jaccard 掉到 0.5,是规则脆还是样本被砍了一半:分半稳定性不能做减法分解
- 阈值自己稳不稳:信噪比≥2 筛出的集合,换一半人就换一批
- 给参数发入场券的阈值本身在抖:信噪比筛选规则的可复现性
- 环境-生命频率接口
证据等级: C
参考文献: PMID 18258670, PMID 15160403, PMID 16217841, PMID 22733454, PMID 18389492, PMID 17066378, PMID 24920271, PMID 22806703, PMID 25029084, PMID 30591808, PMID 19473330, PMID 21731012
L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md节点表仅按易学/中医侧、物理侧、生理/体验侧三类域分派(原文「| 节点 | 各域的概念/量(五行、五脏、频率、颜色…) | 按域分派(见下表) |」),三类域均无「筛选阈值 / 切点 / 分位位置 / 分布形态判定 / 入选集合」这类测量学维度节点;边表止于 E-050(「六十甲子纳音 | 对应 | 60 个频率」,状态 open),无任何以阈值、切点或分布形态为源或目标的边。EDGE_TABLE_science-yixue.md亦仅有概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇 / 频率 407 篇),无 L1/L2/L3 层级与 active/open/conflict 状态判定。
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
How Far Should the Threshold Be from the Median for Binary Classification: What Really Determines Jaccard Similarity Is Not the Percentile but How Many Samples Remain on the Rar…
Evidence level: C
Setting a threshold at a specific percentile is not a freely adjustable knob; it is an order statistic locked by sample size.
I. This Question Presupposes a Nonexistent Knob
In the previous round, we reversed the verification sequence: first check where $\ge 2$ falls in the percentile distribution, then decide that a binary split is not worth retaining. This sounds reasonable, but it assumes that “how far from the mode” can be freely selected—choosing the 85th percentile if desired.
Percentile positions are not free parameters. The 95th percentile with $n=20$ is the 19th order statistic, a value determined solely by the second-to-last observation in the sorted sequence; changing an individual or the grouping shifts this observation, causing the cutoff point to relocate entirely. The horizontal axis appears continuous, but it actually consists of only a few grid points, each carrying its own jitter determined by $n$.
II. Keep the Binary Ledger Clear First
Splitting a continuous variable into two categories loses information and reduces statistical power; this is an established conclusion in systems research: grouping continuous predictors in multivariate regression has been deemed a “bad idea” (PMID 16217841); power loss in binary continuous outcomes in cluster randomized trials can be calculated (PMID 22733454); the statistical properties of binary variables are not interchangeable with continuous ones (PMID 18389492); even odds ratios derived from binary age groups are biased (PMID 17066378).
Alternative routes already exist: the same original studies can use distribution methods to simultaneously provide both continuous and binary estimates with consistent inference (PMID 24920271). The directional hint is clear—it is not “how to cut more stably,” but rather whether one can avoid cutting altogether.
III. Empirical Accounting for Cutoff Bias: Main Driver is Sample Size
One simulation study that truly quantified this set true sensitivity and specificity at 84%, using data-driven selection via the Youden index (PMID 18258670). With $n=40$, the estimate became approximately 90%; with $n=200$, it was about 86%, indicating bias inversely proportional to sample size.
Also worth remembering from the same results: when sample size is fixed, the distribution shape has little effect on this bias. This directly negates the prerequisite task of “drawing lines based on distribution shape”—at least regarding this type of bias, shape is not the independent variable. Studies also tested three more robust approaches (assuming distribution, leave-one-out, and smoothed ROC); these are less prone to bias but performance collapses when assumptions are violated. The corresponding correction tool is that effect estimates after selecting a cutoff must use shrinkage with bootstrapping to restore confidence interval coverage (PMID 15160403).
IV. These Percentiles Are Indistinguishable in Small Samples
There is a more fundamental issue: can the 70th and 85th percentiles be distinguished at this $n$.
Confidence intervals for quantiles and percentiles have specific algorithms (PMID 30591808); reference limits are essentially estimates of extreme percentiles, their precision determined by sample size, and there are existing algorithms for minimum subject numbers (PMID 25029084). On the grid: with $n=20$, the 85th percentile rests on only 3 observations, while the 95th percentile rests on just 1. The intervals for these two cutoff values almost certainly overlap—”placing at 85 or 95″ under this sample size is not an executable operation but two names for the same fuzzy zone.
Practical suggestions for small-sample reference intervals align: plot values as dot plots, calculate reference limits using multiple methods to see if they hold up (PMID 19473330); when within-individual variability is large, distributions are strongly skewed, and sample sizes are small, distribution estimation requires extreme caution (PMID 21731012).
V. The “Bimodal” Premise Must First Pass Its Own Test
“How far from the mode” implies a morphological judgment. Bimodality suggests a natural boundary exists, making binary splitting logical, but morphological judgments are fragile.
Three common metrics (bimodal coefficient BC, Hartigan dip, and AIC difference for single vs. two components) have been systematically compared (PMID 22806703): distance affects overall strength more than proportion, while the effect of proportion varies drastically across metrics; skewness can artificially inflate bimodality detection, causing anomalous interactions in certain cases; AIC difference is extremely sensitive, almost classifying all distributions as bimodal. BC and dip are generally consistent but have several important divergences.
The implication is: a skewed unimodal distribution is likely to be read as bimodal, leading one to think a bimodal line should be used, when in fact the data has only one peak from start to finish. Placing morphology at the very top of the reasoning chain equals using a judgment with limited discriminative power to support all conclusions.
VI. Replace the Horizontal Axis with Rare-Side Count $k$
The point of intervention: percentiles are difficult to use because they conflate distance and sample size; replacing the horizontal axis with the count $k$ of samples above the threshold separates the two.
The same 15% empirical line: with $n=100$, it is 15 people; with $n=20$, only 3 remain. The former yields a stable cutoff with roughly seven or eight people in each half; the latter leaves only one or two per half, meaning Jaccard’s numerator and denominator are determined by these few observations. This explains why the same empirical line performs so differently across studies—it is not the same threshold.
Replacing with $k$ yields a testable inference: if stability is primarily determined by this, then the “$k \to$ Jaccard” curves for the four distribution shapes should converge.
VII. Inversion and Four Criteria
The inversion is: do not ask “how far from the mode should the threshold be,” but rather “how many samples must remain on the rare side to suffice.”
Four criteria: if the maximum difference among the four shapes after switching to $k$ is $<0.05$, then percentiles and morphology are reduced to pseudo-variables, and the rule is rewritten as “rare-side $k \ge X$”; if they still clearly differ by shape ($>0.05$) after switching to $k$, then shape truly plays a role and lines must be drawn by shape, but morphological judgment must first undergo skewness correction and report all three metrics concurrently; if $k$ is fixed but Jaccard still declines with $n$, then the issue lies in within-individual correlation and splitting structure, returning to the split-unit line; if calculating confidence intervals for 70th and 85th percentile cutoffs reveals overlap, then any “top 15-20%” empirical line is unexecutable at this sample size, and only report $k$ and its actual count.
Evidence and Limitations
Verified: Data-driven selection of optimal cutoffs leads to overestimation of sensitivity and specificity, with bias amplified as sample size decreases ($n=40$ estimates true value of 84% as ~90%) (PMID 18258670); shape has little effect on this bias when sample size is fixed (same source); confidence interval coverage is insufficient after selecting a cutoff, requiring shrinkage with bootstrapping (PMID 15160403); binary splitting loses information, reduces power, and produces biased odds ratios (PMID 16217841, PMID 22733454, PMID 18389492, PMID 17066378); skewness artificially inflates bimodality detection with inconsistent conclusions across three metrics (PMID 22806703); precision of extreme percentile estimates is determined by sample size with existing algorithms for minimum subject numbers (PMID 25029084, PMID 30591808).
Partially Verified: Distribution methods can provide both continuous and binary estimates consistently from the same original studies (PMID 24920271), but this is secondary analysis of published studies, not prospective validation; small-sample reference intervals require multiple concurrent methods (PMID 19473330), and large within-individual variability with strong skewness require caution (PMID 21731012), both being methodological suggestions rather than empirical tests of this parameter.
To Be Verified: The “threshold percentile $\to$ Jaccard median” curve has never been plotted on the Dan Tian five parameters; the inflection point at the 80-85th percentile is merely a hypothesis awaiting testing; whether the four shape curves converge after switching to count $k$ is the core testable proposition of this paper, for which no data exists yet; whether confidence intervals for 70th and 85th percentile cutoffs actually overlap has also not been calculated.
Related Reading
- Jaccard drops to 0.5, is the rule fragile or was the sample halved: split-half stability cannot be decomposed by subtraction
- Is the threshold itself stable: signal-to-noise ratio $\ge 2$ filters a set, changing half the people changes the batch
- The threshold itself shakes to grant entry tickets to parameters: reproducibility of signal-to-noise ratio filtering rules
- Environment-Life Frequency Interface
Evidence Level: C
References: PMID 18258670, PMID 15160403, PMID 16217841, PMID 22733454, PMID 18389492, PMID 17066378, PMID 24920271, PMID 22806703, PMID 25029084, PMID 30591808, PMID 19473330, PMID 21731012
L1 Network Status: This topic has not yet entered the L1 corresponding network—theCORRESPONDENCE_NETWORK.mdnode table assigns nodes only by Yijing/TCM side, Physics side, and Physiological/Experiential side categories (original text “| Node | Concepts/Quantities in each domain (Five Phases, Five Zangs, Frequency, Color…) | Assigned by domain (see table below) |”), with no measurement dimensions nodes such as “selection threshold / cutoff / percentile position / distribution morphology judgment / selected set” in any of the three domains; the edge table stops at E-050 (“Six Heavenly Stems and Earthly Branchs Na Yin | Corresponds | 60 frequencies”, status open), with no edges originating or terminating at thresholds, cutoffs, or distribution shapes.EDGE_TABLE_science-yixue.mdonly contains concept counts (Dan Tian 51 articles / Qigong 48 articles / Breathing 189 articles / Frequency 407 articles), lacking L1/L2/L3 hierarchy and active/open/conflict status determination.
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply