Evidence level: B
核心命题:没有预注册门槛的准确率,不是一个测量结果,而是一个可以随时被重新定义的形容词。
正文
任何一场关于”命理准不准”的争论,吵到最后都会停在同一句话上:那你说,多少算准?这句话一旦答不上来,前面测的所有数字都白测——因为一份数据可以同时”够准”和”不够准”,取决于你此刻用哪条线去量它。
医学预测研究早就撞上过这堵墙。Rousson 把”准确度”和”效用”明确拆开:一个 AUC 更高的模型,净获益可能反而更差(PMID 21696604)。Steyerberg 那组睾丸癌标志物的例子更直白——AUC 和 R² 都推荐加进 LDH 和 AFP,可换用净获益、并只在一个”临床站得住的 20% 风险阈值”下看,唯一被推荐的反而是 HCG(PMID 21726217)。门槛一换,”哪个指标有用”的答案就换人。
那门槛从哪来?Vickers 的决策曲线分析给出答案:最优切点不是数据里读出来的,而是由”假阴”与”假阳”的代价比推出来的(PMID 17099194)。他 2021 年又把这层说破:其实很难知道 AUC 要多高才算”够高”,多少校准偏差才算”不合格”(PMID 33676020)。同一个道理在统计检验端也成立——最优 α 取决于先验真效应比例、效应量、样本量和两类错误的相对代价,0.05 在有些领域接近最优,在另一些领域差得很远(PMID 30601826)。阈值必须有理由,不能有习惯。
不把门槛写死会怎样?答案叫”多重宇宙”。Steegen 提出的 multiverse 分析证明:同一条数据、同一组合理决策的不同组合,结论可以天差地别(PMID 27694465)。Rengasamy 用 NHANES 两万五千多人做过实测——炎症标志物与抑郁的关系,9 个常见分析决策展开成五万八千多种组合后,同一条关系的”显著率”只有 46.1%(白细胞)与 43.8%(CRP),效应量随决策变动可达中到大(PMID 37336935)。翻译过来就是:门槛不写死,你总能从那 46% 里挑一条出来报。预注册的全部作用,就是把”事前预测”和”事后附会”分成两笔账(PMID 29531091)。
还有一栏必须从门槛方程里永久移出去:“感觉准”。元认知研究显示,主观准确感与客观表现是两条可分离的线——元认知准确度能与任务表现解离,且个体间稳定(PMID 22492751);有些条件下提高客观表现反而降低元认知敏感度(PMID 26791233);而客观表现最差的人,往往是自评最高的人(PMID 10626367)。所以”我感觉断得挺准”只能是被测量的因变量,不能是判据本身。
值得安慰的是,医学界也不假装门槛有唯一真值。常用的 AUC 分级表(0.5–0.6 失败、0.6–0.7 差、0.7–0.8 尚可、0.8–0.9 好、0.9–1.0 优秀)本质是约定而非推导;量表的最小重要差异(MID)也随方法浮动——同一量表用不同算法,阈值能从 16 到 21(PMID 37415100),半标准差法与传统锚定法之间可以差一倍(PMID 33964390)。报告规范因此专门设了一条:切点必须说明理由,并区分”预设”与”探索性”(PMID 26510957 STARD 第 12a 条;参见 PMID 25561516 / PMID 25560730 TRIPOD)。
把这些接起来,本篇要给的是一套可预注册的门槛三件套 + 一条停止规则:
- 主判据只留一个,事前定死(命中率或判别力 AUC 二选一),其余全部降为次要指标;
- 门槛值 = 数字 + 理由,必须写明”我为什么更怕这种错”(即假阴/假阳代价比从哪来);
- 敏感性带:宽松、中档、严格三档门槛同时报告,并附全谱分析,不允许事后选档;
- 停止规则:事先写明”到什么程度即宣布未过门槛”,此后不得追加上下文解释。
易学这边其实一直有门槛,只是它是师门口径——”十断九中””铁口直断””慎言””戒多断”。有意思的是,”慎断、不断”这一档本身就是代价比意识:断错有代价,所以用”少断”去管理假阳性成本。本篇要做的,不是给术数另设一套标准,而是把它内部那套定性门槛,翻译成可以事前写下、写下就不许再动的操作定义。而且要说清一件事:设门槛这件事本身不偏向任何一方——门槛设高,术数吃亏;门槛设低,随机也能过。公平的做法不是让检验者挑一档,而是三档同报,让读者看它落在哪一档。
由此得到一个可证伪的预测:若门槛真的事前写死并三档齐报,结果会呈现典型的”门槛依赖”形态——宽松档连随机臂都能过,严格档命理与随机都过不了,真正的信息只可能出现在中间档。所以”只报一档门槛”本身就是一种被选出来的结果。
一句”多少算准”,问的其实不是命理,问的是我们愿意为哪一种错买单。
证据与局限
已验证: 判别力与净获益可分离,且增量指标的推荐结论随决策阈值翻转(PMID 21696604、PMID 21726217、PMID 17099194、PMID 33676020);最优 α 依赖两类错误的代价权衡,不存在普适阈值(PMID 30601826);不预注册时,同一条关系的”显著率”与效应量剧烈依赖分析决策(multiverse,PMID 27694465;NHANES 实测显著率 46.1%/43.8%,PMID 37336935);主观准确感可与客观命中解离(PMID 22492751、PMID 26791233、PMID 10626367);报告规范已强制区分预设与探索性切点(PMID 26510957、PMID 25561516、PMID 25560730)。
部分验证: AUC 分级表与最小重要差异这类”阈值约定”是学界公认但明确承认”略随意”的口径,且 MID 会随估计方法浮动(PMID 37415100、PMID 33964390)——”阈值有带宽”已是共识,但”带宽该多宽”尚无统一标准;本篇的三件套属方法学移植,尚无双盲协议下的实测。
待验证: ① 命理盲法协议下三档门槛(宽松/中档/严格)的具体取值、顺序关系与其对应的代价比尚无任何预注册估计;② “门槛依赖度”(同数据在宽松档与严格档下结论翻转的比例)在命理任务上是否达到医学预测模型的量级;③ 从业者的主观”感觉准”评分能否预测其客观元认知敏感度——若不能,”感觉准”须永久移出门槛方程。
相关阅读
证据等级: B
参考文献: PMID 26510957, PMID 25561516, PMID 25560730, PMID 33676020, PMID 17099194, PMID 21696604, PMID 29531091, PMID 30601826, PMID 27694465, PMID 37336935, PMID 37415100, PMID 33964390, PMID 21726217, PMID 22492751, PMID 26791233, PMID 10626367
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 Much Is \”Zhun\” Anyway? Why the Judgment Threshold for Metaphysical Verification Must Be Fixed in Advance
Evidence level: B
Core Proposition: An accuracy rate without pre-registered thresholds is not a measurement result; it is an adjective that can be redefined at any time.
Main Text
Any debate regarding the “accuracy” of metaphysics eventually converges on a single question: So, what counts as accurate? If this cannot be answered, all preceding numerical measurements are rendered meaningless—because a single dataset can simultaneously be “accurate enough” and “not accurate enough,” depending on which line you draw to measure it at that moment.
Medical prediction research has long encountered this wall. Rousson explicitly separates “accuracy” from “utility”: a model with a higher AUC may actually yield lower net benefit (PMID 21696604). Steyerberg’s example regarding testicular cancer markers is even more direct—both AUC and R² recommend adding LDH and AFP, yet switching to net benefit and evaluating only at a “clinically viable 20% risk threshold” results in HCG being the sole recommended marker (PMID 21726217). Changing the threshold changes who is deemed to have a “useful” indicator.
Where does this threshold originate? Vickers’ decision curve analysis provides the answer: the optimal cutoff is not read directly from the data but is derived from the ratio of false negatives to false positives (PMID 17099194). He further clarified this in 2021: it is difficult to know how high an AUC must be considered “sufficiently high,” or how much calibration deviation constitutes “inadequate” (PMID 33676020). The same logic holds in statistical testing—the optimal $\alpha$ depends on the prior proportion of true effects, effect size, sample size, and the relative costs of Type I and Type II errors; 0.05 may be near-optimal in some fields but far off in others (PMID 30601826). Thresholds must have a rationale, not just habit.
What happens if the threshold is not fixed? The answer is “multiple universes.” Steegen’s multiverse analysis demonstrates that the same dataset, combined with different reasonable decision sets, can yield conclusions that differ drastically (PMID 27694465). Rengasamy conducted a real-world test using over 25,000 participants from NHANES—examining the relationship between inflammatory markers and depression. When nine common analytical decisions expand into over 58,000 combinations, the “significance rate” of the same relationship drops to 46.1% (for white blood cells) and 43.8% (for CRP), with effect sizes varying from moderate to large depending on the decision (PMID 37336935). Translated: If the threshold is not fixed, you can always pick one of those 46% to report. The entire function of pre-registration is to separate “prior prediction” from “post-hoc rationalization” into two distinct accounts (PMID 29531091).
There is another category that must be permanently removed from the threshold equation: “The feeling of accuracy.” Meta-cognitive research shows that subjective accuracy perception and objective performance lie on two separable lines—meta-cognitive accuracy can dissociate from task performance and remains stable across individuals (PMID 22492751); under certain conditions, improving objective performance can actually reduce meta-cognitive sensitivity (PMID 26791233); and those with the poorest objective performance are often those who rate themselves highest (PMID 10626367). Therefore, “I feel like I’m calling it accurately” can only be a measured dependent variable, not the criterion itself.
To one’s relief, the medical community does not pretend there is a single true value for thresholds. Common AUC grading tables (0.5–0.6: fail, 0.6–0.7: poor, 0.7–0.8: fair, 0.8–0.9: good, 0.9–1.0: excellent) are essentially conventions rather than derivations; the Minimum Important Difference (MID) also fluctuates with methodology—using different algorithms on the same scale can shift thresholds from 16 to 21 (PMID 37415100), and the difference between the half-standard deviation method and traditional anchoring can be a factor of two (PMID 33964390). Reporting guidelines therefore specifically mandate stating the rationale for cutoffs and distinguishing between “prespecified” and “exploratory” (PMID 26510957 STARD Item 12a; see PMID 25561516 / PMID 25560730 TRIPOD).
Connecting these points, this article proposes a pre-registered threshold triad plus one stopping rule:
- Single Primary Criterion: Fix in advance (either hit rate or discriminative power AUC), demoting all others to secondary indicators;
- Threshold Value = Number + Rationale: Explicitly state “why I fear this specific error more” (i.e., derive the false negative/false positive cost ratio);
- Sensitivity Bands: Report three tiers of thresholds (loose, moderate, strict) simultaneously, accompanied by full-spectrum analysis, prohibiting post-hoc tier selection;
- Stopping Rule: Pre-state “at what point the threshold is declared not met,” after which no additional contextual explanation shall be added.
In Yijing Studies, thresholds have always existed, but they are gatekeeping criteria for masters—”nine out of ten hits,” “directly calling it without hesitation,” “caution in speech,” and “refraining from excessive calls.” Interestingly, the tier of “cautious calling, not calling at all” itself embodies a cost-benefit awareness: errors carry costs, so “calling less” manages false positive costs. What this article aims to do is not impose a new standard on divination but translate that internal qualitative threshold into an operational definition that can be written in advance and never altered. Furthermore, it must clarify one thing: Setting a threshold itself is not biased toward any side—setting it high disadvantages divination; setting it low allows randomness to pass. A fair approach is not for the tester to pick a tier, but to report all three simultaneously so readers can see where it falls.
This yields a falsifiable prediction: if thresholds are truly fixed in advance and all three tiers reported, results will exhibit a typical “threshold-dependency” morphology—the loose tier passes even the random arm, the strict tier fails for both divination and randomness, and true information can only appear in the middle tier. Therefore, “reporting only one threshold tier” is itself a selected outcome.
The question “How accurate must it be?” is not really about divination; it asks what kind of error we are willing to pay for.
Evidence and Limitations
Verified: Discriminative power and net benefit are separable, and recommendations for incremental indicators flip based on decision thresholds (PMID 21696604, PMID 21726217, PMID 17099194, PMID 33676020); optimal $\alpha$ depends on the trade-off between Type I and Type II error costs, with no universal threshold (PMID 30601826); without pre-registration, the “significance rate” and effect size of the same relationship depend drastically on analytical decisions (multiverse, PMID 27694465; NHANES observed significance rates 46.1%/43.8%, PMID 37336935); subjective accuracy perception can dissociate from objective hit rates (PMID 22492751, PMID 26791233, PMID 10626367); reporting guidelines now mandate distinguishing prespecified from exploratory cutoffs (PMID 26510957, PMID 25561516, PMID 25560730).
Partially Verified: AUC grading tables and Minimum Important Differences represent “threshold conventions” that are academically recognized but explicitly acknowledged as somewhat arbitrary, and MID fluctuates with estimation methods (PMID 37415100, PMID 33964390)—”threshold bandwidth” is consensus, but “how wide should the bandwidth be” lacks a unified standard; this article’s triad constitutes methodological transplantation without empirical validation under double-blind protocols.
To Be Verified: ① Specific values and ordering relationships for three-tier thresholds (loose/moderate/strict) in Yijing Studies, along with their corresponding cost ratios, have no pre-registered estimates; ② Whether the “degree of threshold dependency” (the proportion of conclusion flips for the same data between loose and strict tiers) reaches the magnitude observed in medical prediction models; ③ Whether practitioners’ subjective “feeling of accuracy” scores can predict their objective meta-cognitive sensitivity—if not, “feeling of accuracy” must be permanently removed from the threshold equation.
Related Reading
- Information Content Upper Limit of Birth Time: Quantifying the Gap Between Metaphysics and Statistics into Bits
- The Gap Between Metaphysical “Feeling of Accuracy” and “Actual Hit”: Quantifying the Barnum Baseline
- Blind Testing of Metaphysical Event Predictions: Writing Assertions and Time Windows in Advance
- Spatiotemporal Frequency Topology Research Pool
Evidence Level: B
References: PMID 26510957, PMID 25561516, PMID 25560730, PMID 33676020, PMID 17099194, PMID 21696604, PMID 29531091, PMID 30601826, PMID 27694465, PMID 37336935, PMID 37415100, PMID 33964390, PMID 21726217, PMID 22492751, PMID 26791233, PMID 10626367
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