命理「多断言一致性」的统计检验:多个维度同时说对,是证据还是自相关?

Written by

in

Language

Evidence level: B

上一篇把”准”拆成了门槛,这一篇拆”多”。一个命盘同时给出性格、事业、婚姻、健康、六亲一串断言,命中五条不说明它是五条证据——在概率上,多断言是连乘而不是加法。这篇要说清的是:多维断言到底是加强,还是同一条证据被说了五遍。

正文

一、把”断得越多越准”翻译成概率语言

术数里有一句被反复引用的经验:高手一开口就是一串——性格、事业、婚姻、健康、六亲,说得越多越”准”。这句话在直觉上成立,在概率上却要先翻一次译。假设一件事命中的概率是 p,独立情况下 m 件全中就是 p 的 m 次方:单项 0.6 的五件事全中只有 0.6⁵ ≈ 0.078。也就是说:多个断言同时命中,天然是一个比单项命中率低得多的数,而不是更高的数。任何”说得越多越神”的感受,都必须先解释:这些断言的命中率为什么没有按乘法塌下去。

二、人脑对”联合概率”本身就系统性高估

Linda 问题之所以成为经典,是因为它违反概率论最基础的一条单调性:合取的判断不高于任一成分。实验显示超过 80% 的被试会犯这个错误,而且极其顽固——有研究给出反馈与金钱激励,合取效应仍在(PMID 12710220);另一项研究让被试反复经历成分事件的真实频率,估计值与真实联合概率的相关性提高了,但高估本身没有减少,只有给出记忆辅助、降低认知负荷时才下降(PMID 23460026)。Gavanski 与 Roskos-Ewoldsen 进一步把成因定位到”人们用了错误的组合规则”,而不只是代表性启发(PMID 1920061)。结论是:人对”好几件事同时成立”的概率天生偏高。这正是一个多断言的命盘听起来”到处都说对了”的第一层原因——它不需要任何预测力。

三、决定性变量是断言之间的相关性方向

但偏高不等于无效。真正决定”多断言是加强还是空转”的,是这些断言之间的相关性。这一点在生物标志物领域被证明得非常干净:若第二个标志物与第一个负相关,它总能提升合并判别力,即使它单独看几乎没有预测力;若正相关,它几乎不可能带来实质增量,即使它自己单独成绩很好(PMID 21918599)。

这个结论直接改写了术数多断言的评价方式。同一张命盘派生的多个维度,在来源上常常同源——都从同一套五行/十神/宫位结构推演出来。若它们之间是强正相关的,那”五个维度都说对了”并非五条证据,而是一条证据被说了五遍,联合命中率的”虚高”来自自相关,不来自信息。反过来,若某个维度能与其它维度负相关(一个说旺、一个说衰,不能同时成立),它才是真正能提供增量的那一条。

四、一对真实对照:提高 8 倍 vs 几乎零增量

同一个”多指标联合”的命题,在医学文献里有正反两极。Yang 等人用似然比把三个遗传检测(factor V Leiden、prothrombin G20210A、protein C 缺乏)联合起来,静脉血栓的阳性预测值至少提高八倍(PMID 12592605)——条件是模型正确反映病因,且各因素贡献独立可乘。另一极:卵巢癌筛查中,27 个额外标志物(其中不少单独表现不错)加进去,却几乎没能提升 CA125 单独使用的判别力(PMID 21918599)。两边都是”多个指标联合”,一边 8 倍,一边零。差别不在数量,在相关性结构与机制清晰度。

五、冒烟枪:极端组合在人群里太稀有,以至于无法被验证

2 型糖尿病的一项研究给了最实用的一记提醒。三个常见风险变异符合乘性累积,每多一个风险等位基因 OR 约 1.28;六个全带的人 OR 达 5.71——但 95% 置信区间是 1.15 到 28.3(PMID 17020404)。区间宽到这种程度,是因为”六个全带”的人在人群里本来就极其稀少。这意味着:即使”多断言全中”真的有效,它也会落在一个人数极少、置信区间极宽的区域里,几乎无法被统计确认。术数最自豪的”铁口直断、条条应验”,恰好是统计学上最难被证明的那一格。

六、同时上报多个断言,必须先过多重检验这道门

只要一个命盘同时给出 m 条断言,并只要其中一条命中就算”准”,那么”至少有一条蒙中”的概率是 1-(1-α)^m,随 m 迅速趋近 1。这就是多结局研究不能各自单独检验的原因。Bonferroni 在相关时过度保守——50 万个 SNP 的有效独立数其实只有约 6.7 万(PMID 18976480);但急于”减掉冗余”同样危险,用 LD 区块数做校正反而变成反保守,实际错误率升到 0.15-0.20(PMID 21176216)。稳妥做法是显式建模相关:合并依赖 P 值需修正 Fisher 方法(PMID 27587659);多个中等相关结局可改用一个联合判据——规定”至少 N 个结局达到显著”,在控制总错误率的同时反而比单结局更有效(PMID 36925625),其收益取决于相关强度(ρ>0.4 才有增益,PMID 33314364)。同理,评分者之间的一致性必须扣掉”偶然一致”:一项毒性量表研究中,经验丰富评分者的 kappa 低至 -0.04(PMID 8320743,方法学见 PMID 23092060)。断言之间的”看起来都很一致”同样要先扣这一笔。

七、易学落点

术数传统内部其实早有一套关于”多断言”的纪律,只是从不用概率语言表述。”一命多断”在行内有分歧:一派主张抓准一条主断、其余为旁证;另一派主张全面铺开。传统的”主断”意识,恰恰接近”负相关维度才提供增量”的直觉——主断选的是那个能与命局其它信息对立、真正能分对错的那一条。而”应期、验案”的记载习惯,本质上要求把”哪条断言、什么时候、命中与否”分开记录,这正是多重检验校正所要求的分账。可操作形式是:同一命盘的多维断言必须事前登记成一个”至少 N 条命中”的联合判据,而不是事后挑命中的那条报;并须预先声明各维度之间的相关方向假设。

可证伪预测:若真按”至少 N 条命中”的联合判据做盲法检验,命中率会随 N 上升快速塌落,且当 N 超过某一值后真盘与打乱盘不再可分——塌落点位置直接暴露命盘多维度之间的真实相关强度。

证据与局限

已验证:人对联合概率(”几件事同时成立”)的估计系统性偏高,且该偏差对反馈与金钱激励不敏感、仅对降低认知负荷敏感(PMID 23460026 / 12710220 / 1920061)。多个指标联合时,增量由指标间的相关方向决定,而非由指标单独成绩决定——负相关总能增量、正相关即使单独强也难增量,且有 27 个标志物加 CA125 几乎零增量的实例(PMID 21918599);三个遗传检测联合使阳性预测值提高至少八倍的对照实例(PMID 12592605)。多重比较的正确处理方式、以及”至少 N 个显著”这一联合判据的有效性,在方法学上均已确立(PMID 18976480 / 21176216 / 27587659 / 36925625 / 33314364 / 23092060 / 8320743)。

部分验证:上述结论全部来自心理学与生物统计领域,“命盘多维度断言之间是正相关还是负相关”从未被直接测量过。本篇的推论建立在一个尚待检验的类比上:若命盘多维度断言同源(来自同一套结构推演),则其命中率应呈正相关。该类比合理但未被数据支持。

待验证:① 命盘多维度断言的实测相关矩阵(需盲法多维断言协议同时采集各维度命中与否);② “至少 N 条命中”联合判据下,N 从 1 到 m 的命中率塌落曲线形态(真盘 vs 打乱盘);③ 断言数量 m 与”至少一条命中”概率的经验关系是否与 1-(1-α)^m 相符——若实测偏平,说明断言之间正相关强,反之亦然。

相关阅读

证据等级: B
参考文献: PMID 23460026, PMID 12710220, PMID 1920061, PMID 21918599, PMID 12592605, PMID 17020404, PMID 18976480, PMID 21176216, PMID 27587659, PMID 36925625, PMID 33314364, PMID 8320743, PMID 23092060


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。

Statistical Testing of \”Multi-Prediction Consistency\” in Metaphysics: Is Simultaneous Accuracy Across Multiple Dimensions Evidence or Spurious Correlation?

Evidence level: B

The previous article deconstructed “accuracy” into a threshold; this one deconstructs “multiplicity.” A single chart providing a string of assertions across personality, career, marriage, health, and kinship does not constitute five pieces of evidence simply because five are correct—statistically, multiple assertions multiply rather than add. This article clarifies whether multi-dimensional assertions strengthen the case or represent the same evidence repeated five times.

Main Text

I. Translating “More Assertions Mean More Accuracy” into Probability Language

In divination, a frequently cited heuristic states that experts provide a string of predictions—personality, career, marriage, health, kinship—and the more they say, the more “accurate” it sounds. While intuitive, this requires a probabilistic translation. If the probability of hitting one event is $p$, then under independence, the probability of $m$ events all being correct is $p^m$. For instance, hitting five independent items with a 0.6 success rate yields only $0.6^5 \approx 0.078$. Thus: simultaneous hits across multiple assertions naturally result in a number far lower than the individual hit rate, not higher. Any sensation that “the more one says, the more divine it is” must first be explained: why do these assertion hit rates not collapse according to multiplication?

II. The Human Brain Systematically Overestimates Joint Probability

The Linda problem became classic precisely because it violates a fundamental monotonicity rule of probability: the conjunction cannot exceed its components. Experiments show over 80% of subjects commit this error, and it is extremely persistent—even with feedback and monetary incentives, the conjunction fallacy persists (PMID 12710220); another study showed that while repeated exposure to true frequencies improved the correlation between estimates and actual joint probabilities, the overestimation itself did not decrease, only reducing cognitive load or providing memory aids mitigated it (PMID 23460026). Gavanski and Roskos-Ewoldsen further located the cause in “incorrect combination rules” rather than just representativeness heuristics (PMID 1920061). The conclusion is that humans inherently overestimate the probability of “several things occurring simultaneously.” This is the first layer explaining why a multi-assertion chart sounds “correct everywhere”—it requires no predictive power.

III. The Decisive Variable is the Direction of Correlation Between Assertions

However, overestimation does not equal invalidity. What truly determines whether “multiple assertions strengthen or spin empty” is the correlation between them. This has been demonstrated cleanly in biomarker fields: if a second marker is negatively correlated with the first, it always improves combined discriminative power, even if it has little predictive value on its own; if positively correlated, it almost never yields substantive incremental gain, even if it performs well individually (PMID 21918599).

This conclusion directly rewrites the evaluation of multi-assertion charts in divination. Multiple dimensions derived from a single chart often share a common source—deduced from the same Five Phases/Ten Gods/ Palace structure. If they are strongly positively correlated, then “five dimensions being correct” is not five pieces of evidence but one piece repeated five times; the inflated joint hit rate stems from self-correlation, not information. Conversely, if a dimension is negatively correlated with others (one says “prosperous,” another says “declining,” which cannot both be true), it is the one truly providing incremental value.

IV. A Real-World Pair: 8x Increase vs. Near-Zero Increment

The same proposition regarding “multi-indicator conjunctions” shows polarized results in medical literature. Yang et al. combined three genetic tests (factor V Leiden, prothrombin G20210A, protein C deficiency) using likelihood ratios; the positive predictive value for venous thrombosis increased by at least eight times (PMID 12592605)—provided the model correctly reflects etiology and factors contribute independently. The other extreme: in ovarian cancer screening, adding 27 additional markers (many performing well individually) failed to significantly improve discriminative power beyond CA125 alone (PMID 21918599). Both involve “multiple indicators,” yet one yields an 8x gain while the other yields zero. The difference lies not in quantity, but in correlation structure and mechanistic clarity.

V. The Smoke Detector: Extreme Combinations Are Too Rare in Populations to Verify

A study on Type 2 diabetes offered a practical warning. Three common risk variants show multiplicative accumulation, with each additional risk allele increasing the Odds Ratio (OR) by ~1.28; six carriers reach an OR of 5.71—but the 95% confidence interval is 1.15 to 28.3 (PMID 17020404). The interval is so wide because individuals carrying “all six” are inherently extremely rare in the population. This means: even if “multiple assertions all hit” is truly effective, it falls into a region with very few subjects and extremely wide confidence intervals, making statistical confirmation nearly impossible. The divination tradition’s most proud “straightforward judgments that all come true” coincides with the statistically hardest-to-verify category.

VI. Reporting Multiple Assertions Requires Passing the Multiple Testing Gate

If a single chart provides $m$ assertions and counts as “accurate” if at least one hits, the probability of “at least one hit” is $1-(1-\alpha)^m$, which rapidly approaches 1 as $m$ increases. This explains why multi-outcome studies cannot be tested individually. Bonferroni correction is overly conservative when correlations exist—there are only ~67,000 effective independent SNPs among 500,000 (PMID 18976480); however, rushing to “remove redundancy” by using LD blocks can become anti-conservative, raising actual error rates to 0.15–0.20 (PMID 21176216). A robust approach is explicit modeling of correlation: combined dependent P-values require Fisher’s method (PMID 27587659); multiple moderately correlated outcomes can use a single joint criterion—specifying “at least N outcomes reach significance”—which is more effective than single-outcome testing while controlling the total error rate (PMID 36925625), with gains depending on correlation strength ($\rho > 0.4$ yields benefits, PMID 33314364). Similarly, inter-rater agreement must deduct “chance agreement”: in a toxicity scale study, experienced raters had a kappa as low as -0.04 (PMID 8320743; methodology in PMID 23092060). The “appearance of consistency” between assertions also requires deducting this penalty.

VII. Application to Yijing Studies

Traditional divination actually possesses an internal discipline regarding “multiple assertions,” merely not expressed in probabilistic language. The concept of “one chart, multiple assertions” holds divergent views within the school: one faction advocates pinpointing a primary assertion with others as corroboration; another advocates a comprehensive spread. The traditional awareness of a “primary assertion” closely mirrors the intuition that “negatively correlated dimensions provide incremental value”—the primary assertion selects the one that can oppose other chart information to truly distinguish right from wrong. The recording habit of “timing and verification cases” essentially requires separating records of “which assertion, when time, hit or missed,” which aligns with the partitioning required by multiple testing correction. The operational form should be: multi-dimensional assertions from a single chart must be registered ex-ante as a joint criterion of “at least N hits,” rather than selecting the hit one post-hoc; and pre-stated assumptions regarding the correlation direction between dimensions must be declared.

Falsifiable Prediction: If blind testing is conducted under the “at least N hits” joint criterion, hit rates will rapidly collapse as $N$ increases, and once $N$ exceeds a certain threshold, the genuine chart will no longer be distinguishable from a shuffled one—the collapse point directly exposes the true correlation strength among the chart’s multi-dimensional assertions.

Evidence and Limitations

Verified: Humans systematically overestimate joint probability (“several things occurring simultaneously”), and this bias is insensitive to feedback or monetary incentives but sensitive to reduced cognitive load (PMID 23460026 / 12710220 / 1920061). When combining multiple indicators, incremental gain is determined by the direction of correlation between indicators, not their individual performance—negative correlation always yields gain, positive correlation rarely does even if strong individually, and there is an instance with 27 markers adding to CA125 yielding near-zero gain (PMID 21918599); conversely, combining three genetic tests increased positive predictive value by at least eight times (PMID 12592605). Correct handling of multiple comparisons and the validity of the “at least N significant” joint criterion are established methodologically (PMID 18976480 / 21176216 / 27587659 / 36925625 / 33314364 / 23092060 / 8320743).

Partially Verified: All above conclusions derive from psychology and biostatistics fields; “whether multi-dimensional assertions in a chart are positively or negatively correlated has never been directly measured.” The inference in this article relies on an analogy yet to be tested: if multi-dimensional assertions share a common source (derived from the same structural deduction), their hit rates should be positively correlated. This analogy is reasonable but lacks data support.

To Be Verified: ① The empirical correlation matrix of multi-dimensional chart assertions (requires blind multi-assertion protocol collecting hit/non-hit across dimensions simultaneously); ② The collapse curve morphology of hit rates as $N$ ranges from 1 to $m$ under the “at least N hits” joint criterion (genuine chart vs. shuffled chart); ③ Whether the empirical relationship between assertion count $m$ and “at least one hit” probability matches $1-(1-\alpha)^m$—if measured flatness indicates strong positive correlation among assertions, otherwise vice versa.

Related Reading

Evidence Level: B
References: PMID 23460026, PMID 12710220, PMID 1920061, PMID 21918599, PMID 12592605, PMID 17020404, PMID 18976480, PMID 21176216, PMID 27587659, PMID 36925625, PMID 33314364, PMID 8320743, PMID 23092060


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 *