判据相容区是空集的时候,这套流程把它判成了什么:把「无解」和「不合格」分开的那道判别器

Written by

in

Language

Evidence level: C

有些判据组合根本不可能同时满足——这是参数空间的形状问题,不是抽样噪声问题。扩样永远解不开一个空集,但现行流程不知道这件事。

一、这道题不用一个数据就能算完

上一轮(#1281)证明两条判据同轴:ICC_adj≥0.75 与 VB_adj/VB≥0.50 共用一个自由度,分界点 ρ=6/7。当时顺手提了一句——ρ<0.75 时那条线是负的。这句话比它听起来重要,因为「达标」在这里不是统计问题,是集合求交。

把两条判据写成同一个参数 s 的约束:甲给 s ≤ (4ρ−3)/ρ,乙给 s ≤ 0.50,丙给 s ≥ 0(份额非负)。交集上界 = min((4ρ−3)/ρ, 0.50),下界 = 0。于是当 ρ<0.75,(4ρ−3)/ρ<0,交集是空集。

空集意味着:不管 s 估得多准、n 加到多大,都没有能通过的点。问题不在「估得准不准」,在那个能通过的点根本不在参数空间里。这是恒等式不是近似——ρ=0.74 时上界 −0.054,ρ=0.62 时 −1.000。

二、可行域是三段,不是两段

多数人以为这条线只分可行/不可行。实际三段,第三段断点正是 6/7:

区段 ρ 范围 可行域宽度 w(ρ) 含义
空集区 ρ<0.75 0 任何 s 都不行
窄带区 0.75–6/7 (4ρ−3)/ρ,0 涨到 0.50 有解但窗口极窄
半空区 ρ>6/7 恒为 0.50 一半空间永远过不去

第三段最易漏。ρ=0.90 时上界不是甲给的 0.667,是乙的 0.50 在顶。即使信度极高,s>0.5 仍然通不过第二道闸——这是判据乙本来的苛刻(要求校正后至少保住一半真差),只在低 ρ 段被甲挡住看不见。

于是通过率有天花板:s 在 [0,1] 上均匀时,上界就是 w(ρ)。ρ=0.78 只有 15.4%,0.82 是 34.2%,0.86 才到 50%——再高的 ρ 也超不过去。

三、边界附近不可判,而且是渐近不可判

那就先估 ρ 看落在哪段。可 ρ 是估出来的,ICC 大样本标准误约 (1−ρ)√(2/n)。要判 ρ 在 0.75 哪侧,需 1.96·SE<|ρ−0.75|,反解得

n ≈ 2 × [(1−ρ)/(ρ−0.75)]²
真实 ρ 距 0.75 所需 n
0.80 0.050 123
0.78 0.030 413
0.74 0.010 5,194
0.745 0.005 19,984

n 与 Δ² 成反比,Δ→0 时 n→∞。边界附近不是「样本不够」,是「永远不够」——越接近 0.75,需要的 n 越没上限。对照计划 n=120,ρ=0.74 处缺口 43 倍,0.745 处 166 倍。这条盲带随 n 收缩但永不闭合。

四、现行流程在空集格里给了什么答案

按 #1281 那对判据模拟(每格 6 万次,ρ 与 s 各带估计噪声),并模拟现行流程:不过就按「偏差 vs 1.5 倍标准误」建议扩样或重做。在真空集格里,输出「重做/换样本」的比例:

ρ n=60 n=120 n=300
0.62 0.965 0.996 1.000
0.68 0.873 0.955 0.998
0.72 0.699 0.795 0.928
0.74 0.556 0.620 0.740

ρ≤0.68 几乎 100% 判成「样本不合格、重测重做」。而正确答案是「这个设计在这个人群上没有解」。代价可算:一轮重测(120 人×2 次、约两周)全白花,下一轮还是空集,还会再判一次重做。空集最恶劣之处就是它不因努力变好,只让人重复投入。

五、判别器一装上,就退化成「不知道」

加一步代数判定:先解交集,空集直接判「设计无解、换终点」。听着干净,实测同样有坑。用 ρ̂ 的 95%CI 三态判(上界<0.75 判空集 / 下界>0.75 判有解 / 否则未定位):

ρ n=120 判「无解」 判「未定位」
0.62 0.821 0.179
0.68 0.364 0.635
0.72 0.065 0.922
0.74 0.013 0.951

离断点越远越准,越近越只能说不知道。ρ=0.74、n=120 时 95% 落进「未定位」。这与第三节的 1/Δ² 是一体两面:判别器没消除盲带,只是把「误判成重做」换成了「承认判不了」。是进步,但不是终点——它把默认答案从「重做」换成显式的「不知道」,后者至少不让人白花钱。

六、两种错的代价不对称

「未定位」时按哪边倒?按空集处理(悲观):ρ∈(0.75,0.78) 的真可行设计被误弃,损失是少做一个本可做的课题。按有解处理(乐观):ρ<0.75 的真空集被送去扩样,损失是一轮全额重测加下一轮再来一次。

两者量级差很远——前者是一次决策失误,后者是自我重复的黑洞。所以悲观默认更合算,但必须明写这条默认,否则它又变成一个没人知道的隐含假设。

还要注意不能把 0.75 当普适常数:#1281 那对只是库内一例,换一对判据或换终点,断点就换成别的数,每对都得自己解一次交集。

七、该加的不是一个数字,是一道工序

所以缺的既不是更严的阈值也不是更大的样本,而是一个顺序:① 术前诊断(零成本纯代数)——把判据对写成同一组参数的约束式、解交集,空集直接判「设计无解」,不进判读、不进扩样;② 判交集宽度——非空也要看 w(ρ),w 很小时先报「即使有解,通过率天花板也只有 x%」,这数立项时该知道;③ 再谈样本量——前两步都过了,n 才有意义,否则是给数学上不可能的任务加预算;④ 未定位的默认值——明写「按空集处理」,并附 n≈2[(1−ρ)/Δ]² 让人自己判断值不值。

一句话:「数据不合格」和「设计无解」看着都是「没过」,但一个该加样本,一个加了也白加。分不清这两件,产线上每次「没过」都可能是不可回收的投入。

证据与局限

已验证:空集判据是纯代数——ρ<0.75 ⇒ (4ρ−3)/ρ<0 是恒等式,可任意精度验算;三段结构里第三段断点 6/7 由 min((4ρ−3)/ρ, 0.50) 的拐点直接给出;判定所需样本量 n≈2[(1−ρ)/Δ]² 只依赖 ICC 的标准误近似,可与任何 ICC 公式对接;空集格里现行流程的误判率由每格 6 万次模拟给出,四个 ρ 格 × 三个 n 格方向一致。

部分验证:模拟用的是「ρ 与 s 各自独立正态」的近似抽样,真实数据里两者强相关(s 与 ρ 由同一批方差分量算出),会让误判率的绝对值偏移;ICC 的标准误用了 (1−ρ)√(2/n) 这个常用近似,小样本、非正态、聚类结构下会低估(PMID 14673940 报告 ICC 自助区间在近边界处覆盖不足,正是这一点);「现行流程按 1.5 倍标准误分流」是我对产线做法的重建,不是逐字复制的实现。

未验证 / 该主题尚未进入 L1 对应网络:本文讨论的「判据相容区是否为空」在对应网络中既无节点也无边,属纯方法学层面,无法用任何易学—生理边锚定。另外,「未定位时按空集处理」这条默认规则本身没有经验校准——本文只论证了两种错的成本量级不对称,没有实际统计产线上「真可行被误弃」与「真空集被扩样」的发生频次,两条损失的实际比值仍是未知数。

相关阅读

L1 网络状态:E-004|金|主|肺|《素问·阴阳应象大论》|现代证据「待补」|L1|open —— 该边仅为 L1 同构、状态 open,现代证据栏原文为「待补」;网络合计 L2 以上边数为 0。本文讨论的是判据相容区为空集时的流程误判,在对应网络中既无节点也无边(EDGE_TABLE 仅有「五行 50 篇 / 五脏 26 篇」等文章计数,无层级与状态判定),只能借 E-004 作话题锚点。按硬规则,凡 L1 标为「薄 / 仅文献 / 无 A 级证据」者最高只能标 C,故本文所有数字均为代数与模拟推演,不构成任何机制验证。

证据等级: C
参考文献: PMID 16317808, PMID 16927251, PMID 34433225, PMID 28850000, PMID 25043853, PMID 40551266, PMID 11746310, PMID 14673940


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。

When the Compatible Region of Criteria is Empty, This Workflow Misclassifies It: The Discriminator That Separates \”No Solution\” from \”Unqualified\

Evidence level: C

Some combinations of criteria are fundamentally impossible to satisfy simultaneously—this is a problem of the shape of the parameter space, not sampling noise. No matter how much we expand the sample, we cannot solve an empty set, yet current procedures remain unaware of this fact.

I. This Problem Can Be Solved with a Single Data Point

The previous round (#1281) demonstrated that two criteria are coaxial: ICC_adj ≥ 0.75 and VB_adj/VB ≥ 0.50 share one degree of freedom, with a boundary point at ρ = 6/7. At that time, it was briefly noted that when ρ < 0.75, the dividing line is negative. This detail is more significant than it sounds because “meeting the standard” here is not a statistical issue, but rather set intersection.

Expressing both criteria as constraints on the same parameter $s$: Criterion A gives $s \leq (4\rho – 3)/\rho$, Criterion B gives $s \leq 0.50$, and Criterion C gives $s \geq 0$ (since shares cannot be negative). The upper bound of the intersection is $\min((4\rho – 3)/\rho, 0.50)$ and the lower bound is $0$. Thus, when $\rho < 0.75$, $(4\rho – 3)/\rho < 0$, making the intersection an empty set.

An empty set means: regardless of how accurately $s$ is estimated or how large $n$ becomes, no point can pass. The problem lies not in “how accurate the estimate is,” but that the point capable of passing simply does not exist within the parameter space. This is an identity, not an approximation—when $\rho = 0.74$, the upper bound is $-0.054$, and when $\rho = 0.62$, it is $-1.000$.

II. The Feasible Region Has Three Segments, Not Two

Most people assume this line only divides feasible from infeasible. In reality, there are three segments, with the third segment’s breakpoint precisely at $6/7$:

Segment ρ Range Feasible Region Width w(ρ) Meaning
Empty Set Region ρ < 0.75 0 No $s$ works
Narrow Band Region 0.75–6/7 $(4\rho – 3)/\rho$ grows from 0 to 0.50 Solutions exist but the window is extremely narrow
Half-Empty Region ρ > 6/7 Constantly 0.50 Half the space is permanently unreachable

The third segment is most easily overlooked. When $\rho = 0.90$, the upper bound is not A’s $0.667$, but B’s $0.50$ at the top. Even with extremely high reliability, $s > 0.5$ still fails the second gate—this is Criterion B’s inherent strictness (requiring at least half of the true difference to be preserved after correction), which is only obscured by A’s blockage at low $\rho$.

Consequently, the pass rate has a ceiling: when $s$ is uniformly distributed on $[0,1]$, the upper bound is simply $w(\rho)$. At $\rho = 0.78$, the pass rate is only $15.4\%$; at $\rho = 0.82$, it is $34.2\%$; and only at $\rho = 0.86$ does it reach $50\%$. Higher $\rho$ values cannot overcome this limit.

III. Near the Boundary, Judgment Is Impossible—and Asymptotically Impossible

Thus, we must first estimate $\rho$ to determine which segment it falls into. However, $\rho$ is itself estimated, with ICC’s large-sample standard error approximately $(1-\rho)\sqrt{2/n}$. To judge whether $\rho$ lies above or below $0.75$, we need $1.96 \cdot SE < |\rho – 0.75|$, which yields:

n ≈ 2 × [(1−ρ)/(ρ−0.75)]²
True ρ Distance from 0.75 Required n
0.80 0.050 123
0.78 0.030 413
0.74 0.010 5,194
0.745 0.005 19,984

$n$ is inversely proportional to $\Delta^2$; as $\Delta \to 0$, $n \to \infty$. Near the boundary, it is not “insufficient sample size,” but “never enough”—the closer to $0.75$, the higher the required $n$ becomes without limit. Compared to a planned $n = 120$, the gap at $\rho = 0.74$ is $43\times$, and at $\rho = 0.745$ it is $166\times$. This blind zone shrinks with $n$ but never closes.

IV. What Answer Does the Current Procedure Give Within the Empty Set?

Simulating according to #1281’s pair of criteria (60,000 iterations per cell, with estimation noise for both $\rho$ and $s$), and simulating the current procedure: simply follow the recommendation of “expand sample or redo if deviation vs 1.5× standard error.” Within the vacuum set, the proportion of outputs labeled “redo/change sample” is:

ρ n=60 n=120 n=300
0.62 0.965 0.996 1.000
0.68 0.873 0.955 0.998
0.72 0.699 0.795 0.928
0.74 0.556 0.620 0.740

When $\rho \leq 0.68$, nearly $100\%$ are judged as “sample ineligible, retest and redo.” Yet the correct answer is “this design has no solution for this population.” The cost can be calculated: one full round of retesting (120 people × 2 times, about two weeks) is entirely wasted, and the next round will still be an empty set, requiring another judgment of redo. The worst aspect of the empty set is that it does not improve with effort; it only causes repeated investment.

V. Once the Discriminator Is Added, It Degenerates into “Unknown”

Add an algebraic step: first solve the intersection; if it is empty, directly judge “design has no solution, change endpoint.” This sounds clean, but in practice, it still has pitfalls. Using $\hat{\rho}$’s 95% confidence interval for three-state judgment (upper bound < 0.75 → empty set; lower bound > 0.75 → solution exists; otherwise, undetermined):

ρ n=120 judged “No Solution” Judged “Undetermined”
0.62 0.821 0.179
0.68 0.364 0.635
0.72 0.065 0.922
0.74 0.013 0.951

The farther from the breakpoint, the more accurate; the closer, the only option is “unknown.” When $\rho = 0.74$ and $n = 120$, $95\%$ fall into “undetermined.” This corresponds with Section III’s $1/\Delta^2$: the discriminator does not eliminate the blind zone; it merely replaces “misjudged as redo” with “admitted inability to judge.” It is progress, but not the endpoint—it changes the default answer from “redo” to explicit “unknown,” which at least prevents wasting money.

VI. The Costs of Two Types of Errors Are Asymmetric

When “undetermined,” which side do we default to? If we treat it pessimistically (as empty set): $\rho \in (0.75, 0.78)$’s truly feasible designs are wrongly discarded, losing the opportunity to conduct research that could have been done. If we treat it optimistically (as having solution): $\rho < 0.75$’s vacuum sets are sent for sample expansion, losing a full round of retesting plus another attempt.

The magnitude difference is huge—the former is a single decision error, while the latter is a self-repeating black hole. Therefore, pessimistic default is more cost-effective, but this default must be explicitly stated; otherwise, it becomes another hidden assumption no one knows about.

Also note that $0.75$ should not be treated as a universal constant: the pair in #1281 is just one example within the library. Changing to another pair of criteria or endpoint shifts the breakpoint to a different number, requiring fresh intersection solving for each pair.

VII. What Is Missing Is Not a Number, But a Process

Therefore, what is lacking is neither stricter thresholds nor larger samples, but an ordered sequence: ① preoperative diagnosis (zero-cost pure algebra)—rewrite the criteria pair as constraints on the same set of parameters, solve the intersection, and if empty, directly judge “design has no solution,” avoiding interpretation and expansion; ② judge intersection width—even if non-empty, check $w(\rho)$; when $w$ is very small, first report “even if solutions exist, the pass rate ceiling is only $x\%$,” which should be known during project planning; ③ re-discuss sample size—only after the first two steps are passed does $n$ make sense, otherwise it is adding budget to mathematically impossible tasks; ④ default value for undetermined—explicitly state “treat as empty set,” and include $n \approx 2[(1-\rho)/\Delta]^2$ so users can judge whether it is worth pursuing.

In one sentence: “Data ineligible” and “design has no solution” both look like “failed,” but one should expand samples, while the other will never succeed even with expansion. Without distinguishing these two, every “failed” outcome on the production line may represent unrecoverable investment.

Evidence and Limitations

Verified: The empty set criterion is purely algebraic—$\rho < 0.75 \Rightarrow (4\rho – 3)/\rho < 0$ is an identity verifiable to arbitrary precision; the third segment’s breakpoint at $6/7$ comes directly from the inflection point of $\min((4\rho – 3)/\rho, 0.50)$; the required sample size $n \approx 2[(1-\rho)/\Delta]^2$ depends only on ICC’s standard error approximation and can interface with any ICC formula; the misjudgment rate within the empty set for current procedures is given by simulation with 60,000 iterations per cell, with consistent direction across four $\rho$ cells and three $n$ cells.

Partially Verified: The simulation uses the approximation of “ρ and s independently normally distributed,” whereas in real data they are strongly correlated (since $s$ and $\rho$ are calculated from the same variance components), causing absolute value shifts in misjudgment rates; ICC’s standard error uses the common approximation $(1-\rho)\sqrt{2/n}$, which underestimates for small samples, non-normal distributions, and clustered structures (PMID 14673940 reports ICC bootstrap intervals have insufficient coverage near boundaries, precisely this point); “current procedure divides by 1.5× standard error” is my reconstruction of the production line practice, not a literal copy of implementation.

Not Verified / This topic has not yet entered the L1 corresponding network: The question of “whether the compatible region of criteria is empty” has neither nodes nor edges in the corresponding network, belonging purely to methodological level, and cannot be anchored by any Yijing-physiological edge. Furthermore, the default rule “treat undetermined as empty set” itself lacks empirical calibration—this paper only demonstrates the asymmetry in magnitude of cost between two types of errors, without actual statistical production line frequencies of “truly feasible wrongly discarded” versus “vacuum sets expanded,” so the actual ratio of these two losses remains unknown.

Related Reading

L1 Network Status: E-004 | Metal | Principal | Lung | Su Wen · Yin-Yang Ying Xiang Da Lun | Modern evidence “pending” | L1 | open — this edge is only L1 isomorphic, status open; the modern evidence column originally reads “pending”; total edges above L2 in the network is 0. This paper discusses procedural misjudgment when the compatible region of criteria is empty, which has neither nodes nor edges in the corresponding network (EDGE_TABLE contains only article counts such as “Five Phases 50 articles / Five Organs 26 articles,” without hierarchical and status determination), so we can only use E-004 as a topic anchor. According to hard rules, any L1 marked as “thin / only literature / no A-level evidence” can be marked at most C; therefore, all numbers in this paper are algebraic and simulated deduction, constituting no mechanism verification.

Evidence Level: C
References: PMID 16317808, PMID 16927251, PMID 34433225, PMID 28850000, PMID 25043853, PMID 40551266, PMID 11746310, PMID 14673940


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 *