「差值区间跨了零」不是方法分歧,是检验力换了件衣服穿

Written by

in

Language

Evidence level: C

用「份额差的置信区间是否跨零」替代 Kendall W,看起来是换了一把更严的尺子;实际只换了读数方式,它量的仍是同一份信息,能不能分辨出来由一个与「W 对不对」无关的数决定。

一、这条替代路线在验什么

Kendall W 的毛病是:显著只说明排序不像随机,说不明名次可信。于是有了更聪明的替代——不排秩,直接算份额差。体位 28%、带位 27%,差 1 个百分点,给差值做 bootstrap 区间,看它跨不跨零。

逻辑看着无懈可击。但有个东西被塞进了判据却没写出来:跨不跨零说的不是两法对不对,是这个设计的分辨力够不够。这两件事在数学上是同一件事。

二、20% 的跨零率,就是 80% 的检验力

设差值真值 δ、标准误 SE,d/SE 是标准化效应量。区间跨零当且仅当 |d̂|<1.96SE,故

P(跨零) = Φ(1.96−d/SE) − Φ(−1.96−d/SE)

检验力为 Φ(d/SE−1.96)。两式对解:跨零率恰为 20% 时 d/SE = 1.96+0.842 = 2.802,检验力恰 80.00%。这是恒等式而非近似——跨零率 = 1 − 检验力。所以把判据写成「跨零率 >20% 即两法不一致」,等价于说「设计检验力不到 80%」。判据以为自己在讲分歧,实际在讲功效。

三、源设计的刻度:差 1 个百分点 vs 要 14 个百分点

体位 28% 对带位 27%,δ=0.01。三模态 n=5 的方差分量设计,单份额 bootstrap SE 保守取 0.05,则 d/SE=0.20,跨零率 94.5%,检验力 3.9%。

反解:让跨零率降到 20%,同 SE 下 δ 需 0.140,是实际的十四倍;把 SE 压到 0.03 也还要 δ=0.084。这一对 28/27 离「跨得出零」差一个数量级。

四、必然触发的判据,不构成证据

于是「W 显著而差值 CI 跨零的案例占比 >20%」这条判据,在源设计上会以接近 100% 触发而非 20%。原因不是 W 错了,是 d/SE 小于 1 时跨零率必然在 93% 以上——换 bootstrap、换解析式、换 2000 次重抽样,结论都一样。

它把「什么都没测出来」包装成「方法有矛盾」。拆包装只要一步:报跨零率的同时报出 d/SE。没有这个分母,跨零率是个无含义的百分比。

五、换的不是严度,是表达力

那这条路白走了?不。两法同数据、同方差分量、同重抽样;W 是秩统计量、CI 是连续量线性组合,但在「δ=0 与否」上它们是同一检验的两种表述。故 CI 不可能更严——检验力几乎相同,因为有效信息是同一份。

真正变的是信息类型:CI 给符号与量级,可预注册、可跨研究累加;W 只给一个换不回份额的 0–1 系数。这是表达力升级,值得做,但别当成分辨力升级。

六、整条排序可信还要再抬门槛

k 个模态要比 k−1 次。要求整条排序同时成立,两次区间须同时不跨零:联合覆盖率 k=3 时 0.95²=0.9025,k=5 掉到 0.8145。Bonferroni 把 z 抬到 k=3 的 2.241、k=5 的 2.498;同时满足「不假阳性」与「80% 检验力」所需 d/SE 升到 k=3 的 3.083、k=5 的 3.339。

越是把「替代 W」做认真,对数据要求越高。源设计连 2.802 都差十四倍——现实结论只有一个:先扩样本。

七、边界:差值正则,绝对份额不正则

差值 δ 在 ilr 坐标下是线性组合,正则、可用 bootstrap,这点优于 W。但一旦问「体位份额本身是否大于零」,就踏进参数空间边界:真值为 0 时估计量分布退化,百分位 bootstrap 会塌向观测值(重抽样生成不出比样本更小的方差)。文献记录明确:零边界处的方差估计会让其余参数区间一并失真(PMID 23670939);按广义异质性统计量反解的区间覆盖率对,但宽到没有信息(PMID 27604952)。同为「区间」,一个能报一个不能报,区别全在被测的量是否正则在空间内部。

证据与局限

已验证:「跨零率 = 1 − 检验力」是恒等式;跨零率 20% 精确对应 d/SE=2.802、检验力 80.00%,已用正态分位数反解核验(误差 <1e-6)。联合覆盖率 0.95^(k−1) 与 Bonferroni 门槛均为解析计算。组成数据走 ilr 有明确文献(PMID 27450923、25411321),固定/可变总和路径差异有仿真比较(PMID 40247167)。

部分验证:「跨零率不构成分歧证据」依赖 d/SE 可估,而 SE 由 bootstrap 给出,小样本与边界附近偏窄,d/SE 会偏大、跨零率被低估。此层无法用解析消掉,只能并列报告解析式、jackknife、参数 bootstrap 三种 SE(PMID 24605969 记录小评判者数下各法覆盖率差异明显)。多重比较下区间覆盖不足有直接记录(PMID 2049512)。「重叠区间不等于不显著」在独立两组已有量化规则(PMID 18991332、11800251、30119088),但配对/相关结构下的对应规则未建立——本篇用近似,不是定理。

待验证:① 源设计真实 SE 没估过;② 「W 显著而 CI 跨零」在真实数据中占比多少,没跑过;③ 若把 d/SE≥2.8 设为前置闸门,会切掉多少本可做的分析,成本没算过。

L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 三域节点表(易学/中医侧、物理侧、生理/体验侧)登记的全是信号名——物理侧仅有「相位 / 节律 —— 周期性耦合的关键」,生理侧仅有「呼吸率 / 呼吸模式」「气感报告」等观测项,没有任何一栏登记「统计判定框架」这一层(区间覆盖率、检验力、联合覆盖、标准化效应量 d/SE 均无处挂载);边表 3.1–3.5 止于 E-050(「六十甲子纳音|对应|60 个频率」,层级 L1、状态 open,现代证据栏原文「尚无任何可测来源」),无以「判定框架 / 检验力 / 覆盖率」为源或目标的边。

相关阅读

证据等级: C
参考文献: PMID 27450923, PMID 25411321, PMID 40247167, PMID 23670939, PMID 27604952, PMID 2049512, PMID 18991332, PMID 11800251, PMID 30119088, PMID 24605969

证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


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。

The gap interval crossing zero is not methodological divergence but the validation force changing clothes.

Evidence level: C

Using “whether the confidence interval of the difference crosses zero” instead of Kendall’s W appears to be swapping for a stricter ruler; in reality, it only changes the reading method. It still measures the same information, and whether one can distinguish it depends on a number unrelated to “whether W is correct.”

I. What This Alternative Route Validates

The flaw in Kendall’s W is that significance only indicates the ranking is not random; it does not clarify how credible the ranks are. Thus, a smarter alternative emerged: do not rank, but directly calculate the difference in shares. With body posture at 28% and arm position at 27%, the difference is 1 percentage point. Perform a bootstrap interval on this difference and check if it crosses zero.

The logic appears unassailable. But something was embedded in the criterion without being written out: crossing zero does not indicate whether the two methods are right or wrong; it indicates whether the design’s sensitivity is sufficient. These two things are mathematically the same.

II. A 20% Crossing Rate Equals 80% Power

Let the true difference be $\delta$, the standard error $SE$, and $d/SE$ the standardized effect size. The interval crosses zero if and only if $|\hat{d}| < 1.96 SE$, hence

$$
P(\text{crossing}) = \Phi(1.96 – d/SE) – \Phi(-1.96 – d/SE)
$$

Power is $\Phi(d/SE – 1.96)$. Solving both equations: the crossing rate is exactly 20% when $d/SE = 1.96 + 0.842 = \mathbf{2.802}$, and power is exactly 80.00%. This is an identity, not an approximation—crossing rate = 1 − power. Thus, writing the criterion as “crossing rate >20% implies methods disagree” is equivalent to saying “design power is less than 80%.” The criterion thinks it is talking about discrepancy, but actually talks about power.

III. Scale of the Source Design: Difference of 1 vs. Required 14 Percentage Points

Body posture at 28% versus arm position at 27%, $\delta=0.01$. In a three-modal design with $n=5$, the variance component for a single share bootstrap SE is conservatively taken as 0.05, so $d/SE = \mathbf{0.20}$, crossing rate 94.5%, and power 3.9%.

Working backwards: to bring the crossing rate down to 20%, under the same SE, $\delta$ must be 0.140, fourteen times the actual value; even pressing SE to 0.03 still requires $\delta=0.084$. This pair of 28/27 is off by an order of magnitude from “crossing zero.”

IV. A Criterion That Always Triggers Does Not Constitute Evidence

Thus, the criterion “proportion of cases where W is significant but difference CI crosses zero >20%” will trigger at nearly 100% rather than 20% under the source design. The reason is not that W is wrong, but that when $d/SE < 1$, the crossing rate is inevitably above 93%—regardless of swapping bootstrap, analytical formulas, or resampling 2000 times.

It packages “nothing was measured” as “methods contradict each other.” Unpacking it takes one step: report the crossing rate together with $d/SE$. Without this denominator, the crossing rate is a meaningless percentage.

V. What Changed Is Not Stringency, But Expressiveness

So this route was in vain? No. Both methods use the same data, same variance components, same resampling; W is a rank statistic, CI is a linear combination of continuous quantities, but regarding “whether $\delta=0$,” they are two expressions of the same test. Hence CI cannot be stricter—power is almost identical because effective information is the same.

What truly changes is the type of information: CI provides sign and magnitude, allowing pre-registration and cross-study accumulation; W only provides a 0–1 coefficient that cannot be converted back to shares. This is an upgrade in expressiveness, worth doing, but do not mistake it for an upgrade in sensitivity.

VI. The Entire Ranking Credibility Must Raise the Threshold

For $k$ modalities, one must compare against $k-1$. Requiring the entire ranking to hold simultaneously means both intervals must not cross zero: joint coverage for $k=3$ is $0.95^2 = \mathbf{0.9025}$, dropping to 0.8145 for $k=5$. Bonferroni raises the $z$ threshold to 2.241 for $k=3$ and 2.498 for $k=5; simultaneously satisfying “no false positives” and “80% power” requires $d/SE$ to rise to 3.083 for $k=3$ and 3.339 for $k=5$.

The more seriously one treats “replacing W,” the higher the data requirements. The source design is off by fourteen times from 2.802—the real conclusion is only one: expand the sample size first.

VII. Boundary: Difference Is Regular, Absolute Share Is Not

The difference $\delta$ is a linear combination in ilr coordinates, regular and usable with bootstrap, an advantage over W. But once asking “whether the body posture share itself is greater than zero,” one steps into the parameter space boundary: when the true value is 0, the estimator distribution degenerates, and percentile bootstrap collapses toward observed values (resampling cannot generate variance smaller than the sample). Literature records clearly: variance estimation at the zero boundary causes distortion of all other parameter intervals (PMID 23670939); interval coverage solved via generalized heterogeneity statistics is wide but uninformative (PMID 27604952). Both are “intervals,” one can report, the other cannot; the difference lies entirely in whether the measured quantity is regular within the space.

Evidence and Limitations

Verified: “Crossing rate = 1 − power” is an identity; crossing rate 20% precisely corresponds to $d/SE=2.802$ and power 80.00%, verified via normal quantile inversion (error <1e-6). Joint coverage $0.95^{k-1}$ and Bonferroni thresholds are analytical calculations. Compositional data via ilr has clear literature (PMID 27450923, PMID 25411321); fixed/variable sum path differences have simulation comparison (PMID 40247167).

Partially Verified: “Crossing rate does not constitute discrepancy evidence” depends on $d/SE$ being estimable, while SE is given by bootstrap; small samples and boundaries yield narrow SE, causing $d/SE$ to be overestimated and crossing rate underestimated. This layer cannot be eliminated analytically; only reporting analytical, jackknife, and parametric bootstrap SEs together (PMID 24605969 records clear coverage differences under small judge counts). Multiple comparison leads to insufficient interval coverage (PMID 2049512). “Overlapping intervals do not mean non-significance” has quantified rules for independent groups (PMID 18991332, PMID 11800251, PMID 30119088), but no corresponding rules have been established for paired/correlated structures—this paper uses approximation, not theorem.

To Be Verified: ① the true SE of the source design has never been estimated; ② what proportion of real data shows “W significant but CI crosses zero” has not been run; ③ if setting $d/SE \geq 2.8$ as a prerequisite gate, how much analysis would be cut off and what cost has not been calculated.

L1 Network Status: This topic has not yet entered the L1 corresponding network. The three-domain node table in CORRESPONDENCE_NETWORK.md (Yijing Studies/TCM side, physics side, physiology/experience side) registers only signal names—physics side has only “phase / rhythm — key of periodic coupling,” physiology side has only “respiration rate / respiration pattern,” “qi sensation report” and other observational items, no column registers the layer of “statistical decision framework” (interval coverage, power, joint coverage, standardized effect size $d/SE$ have nowhere to attach); edge table 3.1–3.5 stops at E-050 (“60 Heavenly Stems and Earthly Branches Na Yin | Correspondence | 60 Frequencies”), level L1, status open, modern evidence column original text “no measurable source yet,” with no edges sourced or targeted by “decision framework / power / coverage.”

Related Reading

Evidence Level: C
References: PMID 27450923, PMID 25411321, PMID 40247167, PMID 23670939, PMID 27604952, PMID 2049512, PMID 18991332, PMID 11800251, PMID 30119088, PMID 24605969

Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


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 *