Evidence level: B
风水师的”凶地”评级,如果只是断层、岩性、土壤氡、电导率这些物探数据的模糊转译,那么测得物探数据之后评级就该”下课”;如果评级在物探之外还贡献独立的预测信息,那它编码了科学尚未测量的维度。一句话:评级到底是不是物探数据的”多余复述”?
正文
上一篇文章《形峦凶地评级与断层距离/岩性/土壤氡/电导率的定量映射模型》回答了”凶地评级能不能被物探数据解释”——答案是能解释一部分,但解释到什么程度、解释完之后评级还剩多少”独有信息”,才是真正的问题。这正是信息增量(incremental value)检验的用武之地。
为什么这个问题重要。 风水师目测山水就能断吉凶,这个技能如果完全等价于”断层离得近、石头产氡多”,那它本质上是前科学的地质雷达——信息量已被现代物探数据完全覆盖,科学完全可以替代它。反过来,如果控制住全部物探变量之后,评级仍能预测健康相关结局(比如室内氡超标、地质风险带),那就说明传统经验编码了物探数据没测到的维度——可能是微气候、水文、景观结构,也可能是尚未识别的第三变量。两种结局都是知识,但走向完全不同。而风水实证的系统综述(PMID 37809664)明确指出:风水”对环境特征有效、对人类影响未定”,评级信度好(Cronbach’s α 0.76-0.99)但”评级×健康暴露实测”几乎空白——信息增量检验恰好补上这块拼图。
物探数据本身能预测什么。 现代氡潜力制图已经把”从地质推风险”做到了相当水平:德国用随机森林+空间交叉验证(40km×40km 空间分块留出)以 4448 个土壤气实测点制出全国地生氡潜力图,地质变量主导(PMID 33254926);韩国用 GMDH 元启发优化建模,测试集 AUC 达 0.844,信息增益比显示土壤性质与局部地质最关键(PMID 36608375);英格兰西南部用土壤地球化学+航空伽马能谱+地质预测室内氡,可解释约 40-47% 变异(PMID 27216317);挪威用基岩/漂砾地质分类预测氡潜力,地质类别解释约 40% 变异、kappa 0.37-0.56(PMID 27297055)。这些数字给出一个关键基准:纯物探模型已经能解释四成左右变异。凶地评级若要证明自己有增量价值,就必须越过这条线——在物探模型之上再提升判别力。
信息增量怎么测才严谨。 这里直接借用临床预测模型的方法学:嵌套模型比较,看加进评级后 ΔAUC(区分度增量)与净重分类指数(NRI,净重分类改善)。但方法学文献同时警告:ΔAUC 对增量不敏感(基线 AUC 高时好标记也加不了多少分),而 NRI 对风险切点选择高度敏感、且在小样本下假阳性率惊人(模拟显示训练集假阳性 63%,PMID 23179629 及 NCI 复评)。所以严谨做法是三个指标一起报:ΔAUC、类别 NRI(预设临床切点)、连续 NRI/IDI,并全部用自举置信区间,不单独依赖任何一个的 p 值。这正是”评级信息增量”检验的统计底座。
可执行设计。 沿用 #256 的数据基础(凶地评级数字化评分 × 断层距离+岩性+土壤氡+电导率实测),把分析升级为三步:第一步,只用物探特征建基线模型(XGBoost/随机森林,空间交叉验证防空间自相关,PMID 33254926 的做法);第二步,加入凶地评级作为额外特征建扩展模型;第三步,比较两模型的 ΔAUC 与 NRI,并检验”评级×物探”交互项。若 ΔAUC 与 NRI 均显著为正——评级携带独立预测信息,值得拆解其编码的未测维度;若评级系数在控制物探后被吸收殆尽——评级是物探信息的”经验转译”,信息学上可被替代,但作为低成本筛查工具仍有价值(目测不用打钻)。盲法必须贯穿:评级师只看地形照片不看测量结果,测量者不知评级结论,分析者盲态。
三个坑必须提前排。 一是共线性:评级与断层距离高度相关(#256 已提示),需要 VIF 诊断+去相关处理,否则增量会被高估;二是样本量-特征比:评级评分+四类物探变量在小样本下极易过拟合,需嵌套交叉验证并报特征重要性稳定性;三是跨地区泛化:单一县域训练出的”评级增量”可能只是当地地质的偶然拟合,必须用多地质背景地区做外部验证,检验评级增量方向是否一致。这也正是下一篇文章要深挖的方向。
证据与局限
已验证: 物探数据(断层距离/岩性/土壤氡/电导率)可预测氡风险,机器学习+空间交叉验证已成熟(PMID 33254926, 27216317, 27297055);室内氡-肺癌有明确剂量-反应(PMID 30535840, 8978406);风水评级有良好信度但”评级×健康暴露实测”空白(PMID 37809664);ΔAUC/NRI 增量检验方法学及其陷阱已有共识(PMID 23179629)。
部分验证: 断层邻近对土壤氡的影响强度已被 XGBoost+SHAP 量化,但相对贡献排序因地区而异(PMID 39481574)。
待验证: 凶地评级在控制全部物探变量后是否保留独立预测增量(ΔAUC/NRI 显著为正);评级增量的跨地区方向一致性;评级编码的”未测维度”具体是什么(微气候/水文/景观结构的候选)。
相关阅读
证据等级: B(物探-氡预测、氡-肺癌剂量反应、风水实证综述、增量检验方法学均有 PubMed 直接支撑;”凶地评级信息增量盲法检验”本身无先例,属研究设计提案)
参考文献: PMID 39481574, PMID 33254926, PMID 27216317, PMID 27297055, PMID 37809664, PMID 30535840, PMID 8978406, PMID 23179629
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。
Blind Validation of Information Increment in Feng Shui Risk Rating: Can Feng Shui Ratings Independently Predict Risk Beyond Geophysical Data?
Evidence level: B
If a Feng Shui master’s “hazardous site” (凶地) rating is merely a vague translation of geophysical data such as fault proximity, lithology, soil radon, and electrical conductivity, then the rating should be “dismissed” once geophysical data are measured; conversely, if the rating contributes independent predictive information beyond geophysical variables, it encodes dimensions yet to be measured by science. In short: is the rating merely a “redundant repetition” of geophysical data?
Main Text
The previous article, “Quantitative Mapping Models for Feng Shui Hazard Ratings vs. Fault Distance/Lithology/Soil Radon/Electrical Conductivity,” addressed whether Feng Shui hazard ratings can be explained by geophysical data—the answer is yes, partially. The true question lies in the degree of explanation and, more importantly, how much “unique information” remains in the rating after accounting for geophysical variables. This is precisely where an information increment (incremental value) validation applies.
Why this question matters. A Feng Shui master can judge auspiciousness or inauspiciousness by visually inspecting landscapes. If this skill is entirely equivalent to “proximity to faults” or “high radon-producing rock,” it is essentially a pre-scientific geological radar, with its information volume fully covered by modern geophysical data, allowing science to replace it. Conversely, if controlling for all geophysical variables still allows the rating to predict health-related outcomes (e.g., indoor radon exceedance, geological hazard zones), it indicates that traditional experience encodes dimensions not measured by geophysical data—possibly microclimate, hydrology, or landscape structure, or an unidentified third variable. Both outcomes represent knowledge, but they lead in different directions. A systematic review of Feng Shui empirical evidence (PMID 37809664) explicitly states that Feng Shui is “effective for environmental characteristics but undetermined for human impacts,” with good rating reliability (Cronbach’s α 0.76–0.99) but a near-complete lack of “rating × health exposure measurement” data. An information increment validation恰好 fills this gap.
What geophysical data can predict. Modern radon potential mapping has achieved a considerable level in “inferring risk from geology”: Germany used random forests with spatial cross-validation (40km×40km spatial blocks held out) to create a national soil-gas radon potential map based on 4,448 measured soil gas points, with geological variables dominating (PMID 33254926); Korea used GMDH meta-heuristic optimization modeling, achieving a test set AUC of 0.844, with information gain ratios showing soil properties and local geology as most critical (PMID 36608375); Southwest England used soil geochemistry plus airborne gamma spectrometry plus geological prediction to model indoor radon, explaining approximately 40–47% of the variance (PMID 27216317); Norway used bedrock/glacial erratic geological classification to predict radon potential, with geological categories explaining ~40% of the variance and kappa values ranging from 0.37 to 0.56 (PMID 27297055). These figures provide a key baseline: pure geophysical models already explain roughly 40% of the variance. For a Feng Shui hazard rating to prove its own incremental value, it must surpass this threshold—further improving discriminative power on top of the geophysical model.
How to rigorously measure information increment. Here we directly borrow methodological approaches from clinical prediction modeling: nested model comparison, examining the ΔAUC (increment in discrimination) and Net Reclassification Index (NRI, net reclassification improvement). However, methodological literature simultaneously warns that ΔAUC is insensitive to increments (little additional points can be gained even with good markers when baseline AUC is high), while NRI is highly sensitive to risk cutoff selection and exhibits alarming false-positive rates in small samples (simulation showed a 63% false-positive rate in the training set, PMID 23179629 and NCI re-evaluation). Therefore, the rigorous approach is to report all three indicators together: ΔAUC, categorical NRI (using preset clinical cutoffs), and continuous NRI/IDI, all accompanied by bootstrap confidence intervals, without relying solely on the p-value of any single metric. This constitutes the statistical foundation for validating “rating information increment.”
Executable design. Building upon the data foundation from #256 (digitalized Feng Shui hazard ratings × measured fault distance + lithology + soil radon + electrical conductivity), the analysis upgrades to three steps: Step 1, build a baseline model using only geophysical features (XGBoost/random forest, with spatial cross-validation to prevent spatial autocorrelation, following PMID 33254926); Step 2, add the Feng Shui hazard rating as an additional feature to build an extended model; Step 3, compare the two models for ΔAUC and NRI, and test the “rating × geophysical” interaction term. If both ΔAUC and NRI are significantly positive—the rating carries independent predictive information, warranting an analysis of the unmeasured dimensions it encodes; if the rating coefficient is absorbed to near zero after controlling for geophysical variables—the rating is an “empirical translation” of geophysical information, theoretically replaceable from an information science perspective, yet still valuable as a low-cost screening tool (visual inspection avoids drilling). Blinding must be maintained throughout: raters view only topographic photos without seeing measurement results, measurers are unaware of rating conclusions, and analysts remain blinded.
Three pitfalls to address in advance. First, collinearity: the rating is highly correlated with fault distance (#256 already noted), requiring VIF diagnosis and decorrelation handling, otherwise increments may be overestimated; Second, sample-to-feature ratio: adding rating scores plus four geophysical variables poses a high risk of overfitting in small samples, necessitating nested cross-validation and reporting feature importance stability; Third, cross-regional generalization: “rating increment” derived from training in a single county may merely be an accidental fit to local geology, requiring external validation across multiple geological backgrounds to verify whether the direction of the rating increment remains consistent. This is precisely the direction for deep exploration in the next article.
Evidence and Limitations
Verified: Geophysical data (fault distance/lithology/soil radon/electrical conductivity) can predict radon risk; machine learning plus spatial cross-validation is mature (PMID 33254926, 27216317, 27297055); indoor radon-lung cancer has a clear dose-response relationship (PMID 30535840, 8978406); Feng Shui ratings show good reliability but lack “rating × health exposure measurement” data (PMID 37809664); methodological consensus exists for ΔAUC/NRI increment validation and its pitfalls (PMID 23179629).
Partially Verified: The intensity of fault proximity effects on soil radon has been quantified using XGBoost+SHAP, though relative contribution rankings vary by region (PMID 39481574).
To be Verified: Whether Feng Shui hazard ratings retain independent predictive increments after controlling for all geophysical variables (significant positive ΔAUC/NRI); cross-regional consistency of rating increment direction; specific nature of the “unmeasured dimensions” encoded by ratings (candidates include microclimate/hydrology/landscape structure).
Related Reading
- Quantitative Mapping Models for Feng Shui Hazard Ratings vs. Fault Distance/Lithology/Soil Radon/Electrical Conductivity
- Double-blind Feng Shui Rating × Paired Indoor Radon Measurement Head-to-Head Validation
- Research Pool: Environment-Life Frequency Interface
Evidence Level: B (Geophysical-radon prediction, radon-lung cancer dose-response, Feng Shui empirical review, and increment validation methodology all have direct PubMed support; “blind validation of Feng Shui hazard rating information increment” itself has no precedent and constitutes a research design proposal)
References: PMID 39481574, PMID 33254926, PMID 27216317, PMID 27297055, PMID 37809664, PMID 30535840, PMID 8978406, PMID 23179629
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