Evidence level: B
三个无量纲指标对”大凶/中凶/小凶”的判别力,看起来是一场选秀。其实它是一场被样本量、度量方式和测量噪声三重夹击的考试。
正文
前面几步已经做完三件事:形煞与风压异常带的重合度算过了,几何线索的暗示效应分离过了,三类异常带谁在背锅也拆过了。现在轮到最直白的一步——三类形煞各有最贴切的无量纲指标:天斩煞对风速放大比 K,尖角煞对极值负压系数 Cp,min,反弓煞对压力脉动均方根 RMS。那就把三个指标和风水师的凶应评级摆在一起,看谁最配”大凶”。
听起来像选秀:三个选手,同一批评委,谁相关性高谁赢。这个类比第一步就错了,而且一错,后面四步会连着错。
一、三个人考的不是同一张卷子。 K 是气流被挤得多快,Cp,min 是某点被吸得多狠,RMS 是压力抖得多厉害——三种物理过程,三条致害路径。穹顶实测显示,同一几何下外压峰值在背风区对高跨比和湍流强度并不敏感,但内压以负压为主、内外净负压峰值被局部推高(PMID 34576493)。一栋房子至少有两套压力,指标不指定测点就没有意义。更麻烦的是 Cp,min 不是直接测的,是从有限长度的压力时程里外推的,阈值取在哪会实打实改变数值(PMID 33642655)。三个指标各带各的外推误差,”谁最准”就没有共同分母。
二、样本一切三份,功效先崩。 检验只能分层做:天斩只用 K,尖角只用 Cp,min,反弓只用 RMS。假设总共 300 例、每类 100、凶级三档均分,每类里的”大凶”只剩三十来个;要在这规模上把 AUC=0.7 和 0.5 分开,功效已在边缘。再劈一半做跨地区验证,每格十几个,算出的置信区间宽到互相重叠,排序基本由抽样噪声决定。
这就是第一个陷阱:“哪个指标最配大凶”的阴性答案,可能不是被数据否定的,而是被样本量否定的。 两者必须分开说,否则一次功效不足的研究会被读成”风压与凶应无关”,整条线被误杀。回归里”每个变量需要多少事件”的老规矩(PMID 12025266)在这里照样适用,只是事件数已被三层分类切过一刀。
三、凶级是序数,不能拿二分类 AUC 硬套。 常见做法是把评级压成”大凶 vs 非大凶”,可这一压就把”中凶”和”小凶”合并了——而这两档恰恰是实务里”要不要动”的分界线,等于抹掉信息量最大的那道界。序数结局有自己的度量:多分类判别指数 PDI(PMID 33866577)、序数 c-index。一致性用加权 kappa 时必须写明线性还是二次权重:二次权重对”差一档”罚得轻、对”差两档”罚得重(PMID 25516203),正对应”大凶判成小凶”远比”判成中凶”严重。权重方案不是技术细节,是价值判断,必须在方法里先写死。 同一份数据还能给出三个不同的判别力,取决于协变量是当干扰、当对象还是当增量来源(PMID 20046933)。
四、联合模型的”涨分”可能是重复计数。 医学上这套逻辑成立过:三个炎症标志物单独 AUC 为 0.766、0.716、0.768,联合到 0.894(PMID 41929599),涨 0.13,因为三者走不同通路。风压三指标不是这样——K 和 RMS 由同一湍流场驱动,尖角处 Cp,min 与 RMS 也来自同一层分离剪切。先做共线性诊断,再跑嵌套模型。判据可定硬一点:若联合模型相对最强单指标的 AUC 增量小于 0.03 且似然比检验不显著,则三指标实为一个维度,应合并成”风压强度综合因子”再谈判别,而不是摆三个数假装有三个证据。
五、评委那边的噪声会把阴性结果放大。 风水师评级的一致性其实不差:108 条禁忌、11 位学者加 67 位从业者,跨性别、职业、学历、流派均无显著差异(PMID 37089290)。但同一研究承认重测信度不理想——隔半年再问,判断会漂。这是测量误差,会系统性压低关联:若评级可靠性 0.6、指标 0.9,真实相关 0.5 会被观测成 0.5×√(0.6×0.9)≈0.37,正好掉到门槛下(误差规划见 PMID 33410311)。必须做衰减校正并报两套数:校正后由不显著变显著,原阴性结论作废;校正后仍不显著,结论才站得住。
六、换个城市,排名会不会翻。 三指标都是几何与来流的函数,换城市等于换粗糙度、风气候、建筑密度。缩尺实验还有门槛:室外流动要进入雷诺数无关区间,临界值约 27000,判据是相对变化率 ≤5%(PMID 34659649)。不达门槛,系数换个尺度就不成立。所以跨地区验证不是把两地数据凑一起算总 AUC,而是先确认两地都过了门槛。若 A 城 Cp,min 最强、B 城 RMS 最强,那就不存在普适的”最配大凶的指标”,只有”在什么风气候下哪个最配”。
所以这篇真正的产出不是谁赢,是四道门:得分层后功效够不够、序数度量用对没有、联合增量是不是共线造的、评级噪声校没校正。四门有一门没过,答案就不可信。这也顺带说明了风水量化为何长期停在”建档”进不到”判别”——不缺数据,K、Cp,min、RMS 都是现成标准量;缺的是这套门。没有门,谁都能拿三百例样本宣布自己赢了。
证据与局限
已验证: K、Cp,min、RMS 是风工程标准无量纲量,内外压双套、测点位置决定系数值(PMID 34576493);极值峰值须从有限时程外推、阈值选择有客观判据(PMID 33642655);缩尺实验存在雷诺数无关门槛,室外临界值约 27000、判据为相对变化率 ≤5%(PMID 34659649);风水评级内部一致性良好但重测信度不理想(PMID 37089290);序数结局有 PDI 等专用度量(PMID 33866577),加权 kappa 权重方案有明确语义(PMID 25516203);多标志物联合可显著提升判别(PMID 41929599);测量误差致关联系统性衰减(PMID 33410311)。
部分验证: “分层后功效不足”基于 300 例的推算假设,非实测样本;三指标实际共线性程度尚无风洞或实测直接报告,仅物理机制层面推断;回归样本量规则(PMID 12025266)原用于二分类结局,用于三档序数属外推。
待验证: ①三类形煞无量纲指标的实际相关矩阵与方差膨胀因子;②分层后各类形煞的真实可用案例数;③风水师形煞凶应评级任务的可靠性系数实测值;④跨地区(不同地表粗糙度、风气候)下指标判别排名的稳定性。
相关阅读
证据等级: B
参考文献: PMID 34576493, PMID 33642655, PMID 12025266, PMID 33866577, PMID 25516203, PMID 20046933, PMID 41929599, PMID 37089290, PMID 33410311, PMID 34659649
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。
Can we still determine who is most accurate after splitting the samples into three parts?
Evidence level: B
The discriminative power of three dimensionless indicators against “Great Misfortune/Medium Misfortune/Small Misfortune” appears to be a talent show. In reality, it is an examination subjected to triple pressure from sample size, measurement methodology, and measurement noise.
Main Text
The preceding steps have completed three tasks: we calculated the overlap between Form Sha and abnormal wind pressure zones, separated the suggestive effects of geometric cues, and decomposed which type of anomaly zone is taking the blame. Now comes the most straightforward step—each type of Form Sha has a most appropriate dimensionless indicator: the Velocity Amplification Ratio $K$ for Sky-Cutting Sha, the Minimum Coefficient of Pressure $C_{p,min}$ for Acute-Angle Sha, and the Root Mean Square (RMS) of pressure fluctuations for Reverse-Arrow Sha. We should then align these three indicators with the Feng Shui masters’ severity ratings to see which best matches “Great Misfortune.”
It sounds like a talent show: three contestants, one panel of critics; whoever has the highest correlation wins. This analogy is flawed from the first step, and once wrong, the subsequent four steps will also be wrong.
1. They are not taking the same exam. $K$ measures how quickly air is squeezed; $C_{p,min}$ measures how intensely a point is sucked; RMS measures how violently pressure fluctuates—three distinct physical processes, three distinct pathways of harm. Dome measurements show that under the same geometry, external pressure peaks in the leeward zone are not sensitive to span-to-height ratios or turbulence intensity, whereas internal pressure is dominated by negative values and the net peak negative pressure is locally elevated (PMID 34576493). A building has at least two sets of pressures; without specifying the measurement point, the indicators are meaningless. Furthermore, $C_{p,min}$ is not measured directly but extrapolated from pressure time histories of finite length, and the choice of threshold will concretely alter the value (PMID 33642655). Each indicator carries its own extrapolation error; therefore, there is no common denominator for “who is most accurate.”
2. Splitting the sample three ways collapses statistical power first. Testing must be stratified: Sky-Cutting uses only $K$, Acute-Angle uses only $C_{p,min}$, and Reverse-Arrow uses only RMS. Assuming a total of 300 cases, 100 per category, and an equal split across three severity levels, the number of “Great Misfortune” cases within each category drops to only a few dozen; distinguishing between an AUC of 0.7 and 0.5 at this scale is already at the edge of feasibility. If we split again for cross-regional validation, with only a dozen per cell, the resulting confidence intervals will be wide enough to overlap, making the ranking determined largely by sampling noise.
This is the first trap: The negative answer to “which indicator best fits Great Misfortune” may not be rejected by the data, but rather by insufficient sample size. These two factors must be discussed separately; otherwise, a study with insufficient power will be misread as “wind pressure is unrelated to misfortune,” leading to the erroneous elimination of an entire line of inquiry. The established rule regarding the number of events required per variable in regression (PMID 12025266) still applies here, though the event count has been sliced by three layers of classification.
3. Severity levels are ordinal and cannot be forced into binary AUC. Common practice involves collapsing ratings into “Great Misfortune vs. Not Great Misfortune,” but this collapse merges “Medium Misfortune” and “Small Misfortune”—the very tiers that serve as the practical boundary for “whether to act,” thereby erasing the most informative cut. Ordinal outcomes have their own metrics: the Multi-class Discrimination Index (PDI) (PMID 33866577) or the ordinal c-index. When using weighted kappa for consistency, one must explicitly state whether linear or quadratic weighting is used: quadratic weighting penalizes a one-tier difference less severely than a two-tier difference (PMID 25510263), which corresponds precisely to the fact that misclassifying “Great Misfortune” as “Small Misfortune” is far more severe than misclassifying it as “Medium Misfortune.” Weighting schemes are not technical details but value judgments; they must be fixed in the methodology. The same dataset can yield three different discriminative powers depending on whether covariates are treated as confounders, subjects, or incremental sources (PMID 20046933).
4. The “score increase” in joint models may stem from double counting. This logic holds in medicine: three inflammatory markers have individual AUCs of 0.766, 0.716, and 0.768, but jointly reach 0.894 (PMID 41929599), an increase of 0.13, because they follow different pathways. The three wind pressure indicators do not operate this way—$K$ and RMS are driven by the same turbulent field, and $C_{p,min}$ at acute angles and RMS also originate from the same layer of separated shear. First conduct a collinearity diagnosis, then run nested models. The criterion should be strict: If the relative AUC increment of the joint model compared to the strongest single indicator is less than 0.03 and the likelihood ratio test is not significant, then the three indicators represent a single dimension and should be merged into a “Comprehensive Wind Pressure Intensity Factor” before re-evaluating severity, rather than presenting three numbers as if they constitute three independent pieces of evidence.
5. Noise on the panel side amplifies negative results. The consistency among Feng Shui masters is actually not poor: across 108 prohibitions, with 11 scholars and 67 practitioners, there were no significant differences across gender, profession, education level, or school (PMID 37089290). However, the same study admits that retest reliability is not ideal—judgments drift if asked again after half a year. This is measurement error, which systematically suppresses associations: if rating reliability is 0.6 and indicator reliability is 0.9, a true correlation of 0.5 would be observed as $0.5 \times \sqrt{0.6 \times 0.9} \approx 0.37$, dropping exactly below the threshold (error planning see PMID 33410311). One must perform attenuation correction and report two sets of numbers: If the corrected result becomes significant while the original was not, the original negative conclusion is void; if it remains non-significant after correction, the conclusion stands.
6. Would rankings flip in a different city? All three indicators are functions of geometry and inflow; changing cities equals changing roughness, wind climate, and building density. Scaling experiments have thresholds: outdoor flow must enter the Reynolds number-independent regime, with a critical value around 27,000, judged by a relative change rate $\leq 5\%$ (PMID 34659649). Without meeting this threshold, coefficients do not hold across scales. Therefore, cross-regional validation is not simply pooling data from two cities to calculate a total AUC, but first confirming that both regions meet the threshold. If $C_{p,min}$ is strongest in City A and RMS is strongest in City B, then there is no universal “indicator best fitting Great Misfortune,” only “which indicator fits best under what wind climates.”
Therefore, the true output of this paper is not who wins, but four gates: whether power is sufficient after stratification, whether ordinal metrics are applied correctly, whether joint increments are artifacts of collinearity, and whether rating noise has been corrected. If any one gate is not passed, the answer is unreliable. This also explains why Feng Shui quantification has long stalled at “filing” rather than advancing to “discrimination”—data is not lacking; $K$, $C_{p,min}$, and RMS are all standard dimensionless quantities. What is lacking is this set of gates. Without them, anyone can claim victory with 300 cases.
Evidence and Limitations
Verified: $K$, $C_{p,min}$, and RMS are standard dimensionless quantities in wind engineering; internal and external pressure sets and measurement point locations determine coefficient values (PMID 34576493); peak extrema must be extrapolated from finite time histories, and threshold selection has objective criteria (PMID 33642655); scaling experiments have a Reynolds number-independent threshold, with an outdoor critical value around 27,000 and a criterion of relative change rate $\leq 5\%$ (PMID 34659649); internal consistency among Feng Shui ratings is good but retest reliability is not ideal (PMID 37089290); ordinal outcomes have dedicated metrics such as PDI (PMID 33866577), and weighting schemes for weighted kappa have clear semantics (PMID 25510263); multi-marker combinations can significantly improve discrimination (PMID 41929599); measurement error causes systematic attenuation of associations (PMID 33410311).
Partially Verified: “Insufficient power after stratification” is based on hypothetical assumptions using 300 cases, not actual samples; the actual degree of collinearity among the three indicators has not been directly reported by wind tunnel or field measurements, inferred only from physical mechanisms; regression sample size rules (PMID 12025266) originally apply to binary outcomes and applying them to three-tier ordinal outcomes constitutes extrapolation.
To Be Verified: ① The actual correlation matrix and variance inflation factors for the dimensionless indicators of the three Form Sha types; ② The true number of usable cases per Form Sha type after stratification; ③ The measured reliability coefficient for the Feng Shui masters’ task of rating Form Sha severity; ④ The stability of indicator ranking for discrimination across regions (different surface roughness, wind climates).
Related Reading
- Quantitative Filing of Velocity Amplification Ratios and Minimum Coefficients of Pressure for Sky-Cutting, Acute-Angle, and Reverse-Arrow Sha
- After Painting Photos Gray and White Blocks: Can Feng Shui Masters Still Identify Misfortune?
- Contribution Proportions of Three Wind Pressure Anomaly Types and Permutation Tests
- Research Pool: Environment-Life Frequency Interface
Evidence Level: B
References: PMID 34576493, PMID 33642655, PMID 12025266, PMID 33866577, PMID 25510263, PMID 20046933, PMID 41929599, PMID 37089290, PMID 33410311, PMID 34659649
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