三档判读里根本没有「取最大」这一步:那个七千的样本量门槛不是被非光滑顶上去的,是被 1/(1−p_e) 顶上去的

Written by

in

Language

Evidence level: C

前一篇把「TVD 是三条边际差取最大」当成精度损失的可疑来源,设计了一套平滑替代去压样本量门槛。数值做下来,前提不成立:三档时「取最大」在数学上就等于「绝对值求和的一半」,两者是同一个数,非光滑这一层根本不存在。换成任何平滑范数都压不动门槛,因为瓶颈不在估计量形式,而在那个把误差放大五到二十倍的除法。

正文

先回到问题现场。前一篇(#1395)给了一组让人不安的数字:判读一致性的天花板 κmax 只由两个量决定,总变差距离 TVD 与期望一致率 p_e,闭式是 κmax = (1 − TVD − p_e) / (1 − p_e)。它同时算出 TVD 的估计量在大样本下会系统性高估,而这个高估进入 κmax 时会被 1/(1 − p_e) 放大——p_e = 0.90 时十倍,0.95 时二十倍。结论是:要把 κ̂max 的 95% 区间压到半个判据带宽(0.05)以内,m̄ = 0.90 的研究需要七千例,而一致性研究的样本量中位数只有 71 例(PMID 36123642)。

于是很自然的一个追问:TVD 这个估计量是不是选错了?它的定义是「三条边际差取最大」,取最大是非光滑操作,在两条差相等处有尖点。文献对这类泛函有一致结论——非光滑泛函的插值估计量在光滑区外系统性偏、达不到极小极大率,需要多项式近似补救。既然如此,换成 Σ |Δp|^q 这类平滑族,正偏倚应当减小,所需样本量随之下降。推理链条完整,文献支撑现成——但前提一击即破。

一、三档时,「取最大」和「绝对值求和」是同一个数

三档判读里,p 与 q 都只有三个分量,两边各自求和等于 1,因此三条差之和恒为零——正的部分总量等于负的部分总量。总变差距离是正部分总和,也就是绝对值总和的一半。而在这个约束下,绝对值最大的那条差必然恰好等于正部分总和:它已经吃掉全部可正的质量,剩下两条只能分摊负的一半。

这不是近似。五十万对随机三档边际实测,违反次数为零,最大偏差 3.3×10⁻¹⁶,纯浮点误差。所以「非光滑造成的精度损失」在三档场合没有对应的数学对象:两种写法是同一个随机变量。

这个恒等式只在三档成立。四档以上确实分家——只要存在两条正向差和两条负向差,最大值就小于绝对值总和的一半。但一致性研究的主流设计就是二档三档,分家的恰恰不是我们讨论的场合。

二、平滑替代不是「减小偏倚」,是「换掉被估的量」

同一个真 TVD = 0.10 的数据,按三种次档构型造出来,乘性偏倚完全由构型决定:q = 1.5 时 1.43 倍与 1.59 倍;q = 2 时 1.22 倍与 1.41 倍;q = 3 时 1.08 倍与 1.26 倍。也就是说平滑范数估出来的量依赖「次档的差怎么分布」——而这件事在真实数据里正是未知的。拿它当 TVD 替代,等于把定义明确的量换成定义含混的量,再接受 8% 到 60% 的乘性误差。

蒙特卡洛同向。m̄ = 0.75、真 TVD = 0.06、n = 960 时:原始取最大偏倚 0.0000、标准差 0.0186;L2 型(除 √2 校准)偏倚 −0.0070、标准差 0.0160;softmax 平滑偏倚 +0.0087、标准差 0.0157。平滑确有 12%–16% 的方差收益,代价是凭空造出一个系统偏倚,而原始估计量在这个样本量下偏倚已降到零。用不随 n 缩小的常数偏倚,去换会随 n 缩小的标准差,在「区间压到 0.05」这个目标下是亏的。

三、真正的瓶颈是那个除法

m̄ = 0.75、TVD = 0.10 时,原始估计量的偏倚随 n 单调消退:

n 偏倚 标准差 κ̂max 半宽
60 +0.0129 0.0591 0.679
240 +0.0007 0.0351 0.400
960 −0.0001 0.0182 0.206
3840 0.0000 0.0091 0.104
7680 0.0000 0.0064 0.074

偏倚在 n = 240 已可忽略,而 κ̂max 半宽到 7680 还有 0.074,仍高于 0.05。偏倚不是瓶颈,方差才是;方差压不下去,因为它来自「两百例里逐档计数」的信息量上限。 半宽与 √n 成反比也验证了:960 加到 3840,比值 2.014(0.75)、1.811(0.88)、1.588(0.95)。越往高一致区走越偏离 2,因为 p_e 接近 1 时放大因子自己也在抖,这一项在方差里份额上升。

把 n × 标准差² 算出来:m̄ = 0.60 约 0.42,0.75 约 0.42,0.90 约 0.25,0.95 约 0.18。它不是常数,随 m̄ 下降——高一致区差值小,方差主要由噪声底决定。门槛的合式因此是 n ≈ C × (39.2/(1 − p_e))²,C 是随 m̄ 变化的方差系数。代入实算:0.75 需 2846 例、0.90 需 5263 例、0.95 需 7519 例。前一篇报的 3840 / 7680 / 15360 偏高但量级一致(它用 bootstrap、本篇用正态近似乘放大因子),两套口径都指向同一个结论:七千例是真的。

四、一个被证伪的「更聪明的估计」

既然三个差有确定结构关系,理论上可用反方差加权合成最优估计。标准差确实压到原始量的 0.31–0.47 倍,但偏倚同时炸了:−0.07 到 −0.17,且在所有 n 上都不消退。原因是结构假设只在特定构型成立;一旦真实构型是「主档差为零、分歧全在两个次档之间」,反方差加权会忠实估出零,而真 TVD 是 0.10,偏 0.10。

这是本轮最值得记住的一条:任何利用「次档结构」降方差的估计量,都把结论押在了次档构型这个从未被实测过的量上。 原始取最大之所以「粗糙却稳健」,正因为它只用最坏那一条,不假设其余两条怎么分配。

五、结论

那个「换平滑估计量、把 n 门槛从七千压下来」的假设,被证伪了——不是压了一档,是压根压不动。因为

  1. 三档时「取最大」与「绝对值求和的一半」是同一个数,非光滑这一层不存在;
  2. 平滑范数改的是被估的量而非估计误差,引入 8%–60% 乘性偏倚,方向与幅度都依赖未实测的次档构型;
  3. 偏倚在 n = 240 已可忽略,方差是门槛唯一来源,而方差由 1/(1 − p_e) 放大,与估计量形式无关。

报告规范不该改估计量,该改三件事:引用 κmax 必须同时报 n 与 p_e;高一致区(p_e > 0.9)别指望用更大 n 解决,那个区间可承载的精度本身就在 0.1 以上;以及——真需要 0.05 精度的研究,缺的不是更好的估计量,是放弃三档、改连续打分。

证据与局限

已验证(纯代数与蒙特卡洛,可复现):三档时 max|Δp| 与 ½Σ|Δp| 精确相等(50 万对随机边际,违反 0 次);q 范数替代的乘性偏倚由构型决定(q = 2 时 1.22–1.41 倍,q = 3 时 1.08–1.26 倍);原始取最大估计量的偏倚在 n = 240 降至 0.0007、n = 960 降至 −0.0001;κ̂max 半宽与 √n 成反比(960→3840 比值 2.014/1.811/1.588,随 m̄ 升高偏离 2);按 C × (39.2/(1−p_e))² 算出的 n 门槛 2846/5263/7519。

部分验证:平滑替代的方差收益(12%–16%)只在 n = 960、m̄ = 0.75–0.95 这一段网格上核过,未扫全参数域;「方差系数 C 随 m̄ 单调下降」这一观察来自 5 个 m̄ 取值 × 3 种构型,趋势清楚但未做函数形式拟合。

待验证:全部数值建立在参数化边际上,本组从未实测过真实三档裁决数据的 (m̄, TVD) 联合分布(这一条与前一篇同一缺口,仍未补);κ̂max 的区间用的是正态近似乘放大因子,未与 bootstrap 逐点对照;「改连续打分」这条建议的代价(连续打分的判读者间一致性本身更差,是另一条独立的文献)未评估。

L1 网络状态:| **E-060 标签前提** | 五脏辨证(作为分组标签) | 前提 | 标签信度 | 《难经》/ 辨证体系 | **KB-292(conflict):中医医师间诊断一致性在开放人群中「低(low)」,作者称为「令人担忧」(PMID 30383398)……** → **「用五脏标签做分组」这一步本身带噪声** | **冲突(未裁决)** | **conflict** |

该边是本主题在 L1 对应网络里唯一相关的边,且状态为 conflict。冲突双方:一侧是「医师间辨证一致性在开放人群中实测为低,作者明言令人担忧」(PMID 30383398),另一侧是「低值多由一致性统计量被误用造成,改用 Gwet AC2 类系数后结论并不如此悲观」(PMID 31730403)。本条边尚未裁决,因此本文所有关于「判读可靠性量化」的论断只能作方法学推演,不得读作「五脏辨证分组已被证明可用」或「已被证明不可用」。本文未对该边作单方结论。

证据等级: C
参考文献: PMID 36123642, PMID 41230191, PMID 24338600, PMID 8501467, PMID 8621993, PMID 30383398, PMID 31730403

相关阅读


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。

There is no \”take the maximum\” step in the three-tier judgment: that threshold of 7,000 sample size is not pushed up by the non-smooth top, but by 1/(1−p_e).

Evidence level: C

The previous article treated “TVD is the maximum of three marginal differences” as a suspicious source of precision loss and designed a smooth alternative to suppress the sample size threshold. Numerical evaluation showed the premise fails: in three tiers, “taking the maximum” is mathematically identical to “half the sum of absolute values”; they are the same number, so the non-smooth layer does not exist. Switching to any smooth norm cannot suppress the threshold because the bottleneck lies not in the estimator form, but in the division that amplifies error by five to twenty times.

Body

Let us return to the problem site. The previous article (#1395) presented a set of unsettling numbers: the ceiling of consistency, κmax, is determined by only two quantities—the Total Variation Distance (TVD) and the expected agreement rate p_e—with the closed-form expression κmax = (1 − TVD − p_e) / (1 − p_e). It also calculated that the TVD estimator systematically overestimates in large samples, and this overestimation is amplified by 1/(1 − p_e) when entering κmax—tenfold when p_e = 0.90, twentyfold when p_e = 0.95. The conclusion follows: to bring the 95% interval of κ̂max down to half a criterion bandwidth (0.05), studies with m̄ = 0.90 require 7,000 cases, while the median sample size in consistency studies is only 71 (PMID 36123642).

A natural question arises: is the TVD estimator itself mischosen? Its definition—”take the maximum of three marginal differences”—involves a non-smooth operation, creating sharp points where two differences are equal. Literature on such functionals reaches a consistent conclusion: interpolated estimators for non-smooth functionals systematically bias away from smooth regions and fail to achieve minimax rates, requiring polynomial approximation for correction. Thus, replacing the maximum with a smooth family like Σ |Δp|^q should reduce positive bias and lower required sample sizes. The reasoning chain appears complete, supported by existing literature—but the premise collapses immediately.

I. In three tiers, “taking the maximum” and “half the sum of absolute values” are identical numbers

In three-tier interpretation, p and q each have only three components; their respective sums equal 1, so the sum of the three differences is always zero—the total positive mass equals the total negative mass. The Total Variation Distance (TVD) is the sum of positive parts, which equals half the sum of absolute values. Under this constraint, the largest difference must exactly equal the sum of positive parts: it consumes all available positive mass, leaving only half the negative mass for the other two differences.

This is not an approximation. In a Monte Carlo test with 500,000 random three-tier marginals, violations occurred zero times; the maximum deviation was 3.3×10⁻¹⁶, purely floating-point error. Therefore, “precision loss due to non-smoothness” has no corresponding mathematical object in the three-tier case: both formulations yield the same random variable.

This identity holds only for three tiers. With four or more tiers, they diverge—as soon as there are two positive and two negative differences, the maximum is less than half the sum of absolute values. Yet mainstream consistency studies focus on two-tier and three-tier designs; the divergence does not occur in the cases under discussion.

II. Smooth alternatives do not “reduce bias”; they “replace the quantity being estimated”

For the same true TVD = 0.10, constructed under three sub-tier configurations, the multiplicative bias depends entirely on the configuration: q = 1.5 yields 1.43× and 1.59×; q = 2 yields 1.22× and 1.41×; q = 3 yields 1.08× and 1.26×. That is, the smooth-norm estimator yields a quantity that depends on “how the sub-tier differences are distributed”—which is unknown in real data. Using it as a TVD substitute means replacing a well-defined quantity with an ill-defined one, accepting multiplicative errors of 8% to 60%.

Monte Carlo simulations confirm this. With m̄ = 0.75, true TVD = 0.06, n = 960: the original “take maximum” estimator has zero bias and standard deviation 0.0186; the L2-type (divided by √2 calibration) has bias −0.0070 and SD 0.0160; the softmax-smoothed version has bias +0.0087 and SD 0.0157. Smooth norms do offer 12%–16% variance reduction, but at the cost of introducing a systematic bias that is zero in the original estimator. Trading a constant bias (not shrinking with n) for a variance that shrinks with n is disadvantageous when the goal is “interval width ≤ 0.05”.

III. The true bottleneck is the division

With m̄ = 0.75 and TVD = 0.10, the bias of the original estimator vanishes monotonically with n:

n Bias SD κ̂max half-width
60 +0.0129 0.0591 0.679
240 +0.0007 0.0351 0.400
960 −0.0001 0.0182 0.206
3840 0.0000 0.0091 0.104
7680 0.0000 0.0064 0.074

Bias becomes negligible at n = 240, yet the κ̂max half-width remains 0.074 at n = 7680, still above 0.05. Bias is not the bottleneck; variance is; and variance cannot be reduced because it is limited by the information capacity of counting across tiers in 200 cases. The inverse proportionality of half-width to √n confirms this: increasing from 960 to 3840 yields ratios of 2.014 (m̄=0.75), 1.811 (m̄=0.88), and 1.588 (m̄=0.95). Deviation from 2 grows as p_e approaches 1, because the amplification factor itself fluctuates, increasing its share in the variance.

Computing n × SD²: m̄ = 0.60 gives ~0.42, 0.75 gives ~0.42, 0.90 gives ~0.25, and 0.95 gives ~0.18. It is not constant; it decreases with m̄—because in high-agreement regions differences are small, so variance is dominated by noise floor. The proper threshold formula is therefore n ≈ C × (39.2/(1 − p_e))², where C is the variance coefficient varying with m̄. Plugging in real calculations: 2,846 for m̄=0.75, 5,263 for m̄=0.90, and 7,519 for m̄=0.95. The previous article’s 3,840 / 7,680 / 15,360 are slightly high but consistent in order of magnitude (it used bootstrap; this paper uses normal approximation multiplied by the amplification factor). Both approaches point to the same conclusion: 7,000 samples is indeed required.

IV. A “smarter estimator” that has been falsified

Since the three differences have a fixed structural relationship, theoretically one could construct an optimal estimator via inverse-variance weighting. Standard deviation does drop to 0.31–0.47 of the original, but bias explodes: −0.07 to −0.17, and does not vanish across any n. This occurs because the structural assumption holds only for specific configurations; once the true configuration is “main tier difference zero, all divergence between two sub-tiers,” inverse-variance weighting faithfully estimates zero, while true TVD is 0.10, yielding a bias of 0.10.

This is the most important point to remember: any estimator that reduces variance by exploiting “sub-tier structure” bets its conclusion on the sub-tier configuration—a quantity never empirically measured. The original “take maximum” is “rough yet robust” precisely because it uses only the worst-case difference, making no assumptions about how the other two are distributed.

V. Conclusion

The hypothesis that “replacing the smooth estimator will bring the n threshold down from 7,000” has been falsified—not by reducing one tier, but because it cannot be reduced at all. Because

  1. In three tiers, “taking the maximum” and “half the sum of absolute values” are identical; the non-smooth layer does not exist;
  2. Smooth norms change the quantity being estimated, not the estimation error, introducing multiplicative biases of 8%–60%, with direction and magnitude depending on the unmeasured sub-tier configuration;
  3. Bias becomes negligible at n = 240; variance is the sole source of the threshold, and it is amplified by 1/(1 − p_e), independent of estimator form.

Reporting standards should not change the estimator; they should address three things: κmax reporting must always include both n and p_e; high-agreement regions (p_e > 0.9) should not expect larger n to solve the problem, because that region itself can carry precision above 0.1; and—studies truly needing 0.05 precision lack not better estimators, but the decision to abandon three-tier scoring in favor of continuous scoring.

Evidence and Limitations

Verified (pure algebra and Monte Carlo, reproducible): In three tiers, max|Δp| equals ½Σ|Δp| exactly (500,000 random marginals, zero violations); multiplicative bias of q-norm substitution depends on configuration (q = 2: 1.22–1.41×; q = 3: 1.08–1.26×); original “take maximum” estimator bias drops to 0.0007 at n = 240 and −0.0001 at n = 960; κ̂max half-width is inversely proportional to √n (960→3840 ratios: 2.014/1.811/1.588, deviating from 2 as m̄ increases); computed n thresholds via C × (39.2/(1−p_e))²: 2,846/5,263/7,519.

Partially verified: Variance reduction of smooth substitution (12%–16%) was checked only on the grid n = 960, m̄ = 0.75–0.95, not the full parameter space; the observation that “variance coefficient C monotonically decreases with m̄” comes from 5 m̄ values × 3 configurations—trend clear but no functional fit performed.

To be verified: All numerical results rely on parameterized marginals; this group has never empirically measured the joint distribution of (m̄, TVD) in real three-tier adjudication data (this gap remains unfilled, same as the previous article); κ̂max intervals use normal approximation multiplied by amplification factor, not pointwise comparison with bootstrap; the cost of “switching to continuous scoring” (consistency among raters is itself worse, per independent literature) has not been evaluated.

L1 Network State: | **E-060 Label Premise** | Five-organ syndrome differentiation (as grouping label) | Premise | Label reliability | 《Nan Jing》/ syndrome-differentiation system | **KB-292 (conflict): Inter-physician diagnostic consistency in open populations is "low (low)", described as "concerning" (PMID 30383398)…** → **"Using five-organ labels for grouping introduces noise"** | **Conflict (unresolved)** | **conflict** |

This edge is the only relevant one in the L1 corresponding network for this topic, and its state is conflict. The two sides of conflict: one states “inter-physician diagnostic consistency measured in open populations is low, explicitly described as concerning” (PMID 30383398); the other states “low values stem from misapplication of consistency statistics; using Gwet AC2-type coefficients yields a less pessimistic conclusion” (PMID 31730403). This edge remains unresolved; therefore, all claims in this paper regarding “quantifying adjudication reliability” are methodological deductions and must not be read as “five-organ syndrome grouping has been proven usable” or “proven unusable.” This paper does not issue a unilateral conclusion on this edge.

Evidence Level: C
References: PMID 36123642, PMID 41230191, PMID 24338600, PMID 8501467, PMID 8621993, PMID 30383398, PMID 31730403

Further Reading


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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *