Evidence level: B
一个变量在控住两个物探变量之后”还有增量”,这句话要成立,得先过四道关:增量指标本身会不会被骗、控制变量是不是躺在评级的下游、跨地区方向不一致算不算失败、以及评级有没有一把跨地区共用的尺子。
正文
“信息增量”在统计上不是一个量,而是一族量:判别力增量、再分类改善、综合判别改善、临床净收益。它们对同一份数据可以给出互相矛盾的答案。Kerr 的系统评述指出,再分类改善的类别边界是人为选的,对模型校准高度敏感,而且在新变量完全无用时期望也不为零(PMID 24240655)。Pepe 更早就说明,优势比与曲线下面积之间是非线性关系:一个优势比 3 的强危险因素,加进基线面积已经 0.80 的模型,可能只把面积推到 0.82(PMID 15105181)。反过来,一个优势比只有 1.2 的弱变量,在低基线模型里能推出更大的增量。也就是说,”增量有多大”不说明变量有多强,只说明它与现有模型重叠多少。Steyerberg 的框架把判别、校准、再分类、净收益拆成四件事,明确说不可互相替代(PMID 20010215);决策曲线给的是”在哪个阈值区间净收益变正”(PMID 17099194),它不产出单一数字,也就没那么容易被挑出一个好看的值。
第二关更要命:把电导率和室内氡放进模型当控制,隐含假设是它们在因果图上位于评级的上游或旁路。可如果”凶地”在传统判据里本就包含潮湿、背阴、通风不畅,那么室内氡与湿度就不是中立的物探代理,它们身上带着评级的成分。此时控制它们等于控制了处理的一部分,Schisterman 把这叫过度校正(PMID 19525685);若控制变量同时受评级与另一个共同原因影响,它就是个对撞点,控制它反而会凭空造出关联(PMID 29040562)。判定办法很朴素:在测氡之前,把”该宅是否符合潮湿与通风类传统判据”单独编码,看它与氡的相关有多强。相关明显,氡就只能进敏感性分析,不能进主模型。
第三关是”复制”这个词本身。最常见的错误做法是每个地区单独检验、数几个显著。交互检验的功效本来就低,检测一个交互所需的样本量常是检测主效应的数倍(PMID 17578572),六个地区各自做交互检验,有一个显著已属侥幸。更稳的判据是符号:预设交互项方向,只看六个地区同向的有几个。若真实无交互,六地全部同向的概率是 (1/2)⁵,约 3.1%——同样样本下比逐个检验稳得多。但它有前提:方向必须在看数据之前写死,否则事后从六个方向里挑一个,3.1% 会变成 100%。子群分析的系统评估反复印证这一点,试验方案里预设的子群与论文里报告的子群经常对不上(PMID 25030633)。
第四关最容易被跳过:评级可能根本没有跨地区共用的刻度。盲法评级时,专家在一个地区内同时看二十处宅,他实际做的是地区内的相对排序,不是往一把全国通用的尺子上放。后果有两个:模型里放地区固定效应,评级的组间变异被抽走,只剩组内那一小段;不放,评级就部分充当地区身份的代理,与地质背景共线,”信息增量”测到的其实是地区差异。同时,评级者间一致性本身随边际分布浮动——一致性系数的分母含期望一致率,同样的判断行为在两个吉凶比例不同的地区会算出不同数值(分级基准见 PMID 843571)。所以”两地都是 0.5″并不意味着评级质量相同。
判决是分三层报,不要合成一个分数。前置门槛:先报评级者间一致性与吉凶比例,跨地区只比”是否落在同一档”,不比数值。主判据:预设唯一主模型,报判别力增量配自助法区间与全流程置换零分布(重跑整个挑选过程),并报决策曲线净收益,不报再分类改善。复制判据:报六地区交互符号一致率(单侧符号检验),加个体资料合并的交互估计与异质性方差(PMID 28057641),样本量按外部验证要求反算(PMID 34031906)。
《管子·地员》讲”凡草土之道,各有谷造,或高或下,各有草土”——同一套判据,在高处和低处给出不同的具体答案;《葬书》讲”土欲细而坚,润而不泽”,判据是可以现场捏、现场看的物性,不是可以邮寄的分数。古典地理从来没有给过一把跨区域通用的吉凶刻度,”上吉、中吉、下凶”这类分级是后世为教学与交易硬化的。所以六个地区交互项方向不一致,在古人的模型里不是复制失败,而是模型的正常输出——它本来就说”因地而异”。
证据与局限
已验证: 室内氡与肺癌的剂量-反应关系在 13 项欧洲病例对照研究的个体资料合并中稳定成立(PMID 15613366);氡的空间分布受地质单元强烈约束,地质类别可作为空间建模的外部漂移项(PMID 17720284),国家级调查也显示室内氡的空间变异具有明确区域结构(PMID 30645746);季节性修正因子会系统性改变同一住宅的氡读数(PMID 30165499)。方法学侧,再分类改善指标的边界依赖与零假设下非零期望、判别力增量与优势比的非线性、交互检验的低功效、过度校正与对撞点偏倚均已确立(PMID 24240655、15105181、17578572、19525685、29040562)。
部分验证: “风水评级在控制物探变量后是否保留独立信息增量”这一问题本身尚无三变量同步盲法队列的直接答案;跨区域交互方向一致性也没有任何实测数据。本文给出的是判据与设计,不是结论。
待验证: ① 传统”凶地”判据与室内氡、湿度、通风的耦合强度有多大——这决定氡该进主模型还是只进敏感性分析;② 盲法评级在同一地区内的组内变异是否足以支撑任何增量估计;③ 跨地区交互符号一致率在真实队列中能否超过 3.1% 的零假设线。
相关阅读
- 风水吉凶评级、大地电导率与室内氡暴露:三变量联合盲法队列能证明什么
- 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
- 风水”吉地/凶地”与大地电导率分布的系统测量比对
- 研究池子:环境-生命频率接口
证据等级: B
参考文献: PMID 24240655, PMID 15105181, PMID 20010215, PMID 17099194, PMID 19525685, PMID 29040562, PMID 17578572, PMID 25030633, PMID 843571, PMID 28057641, PMID 34031906, PMID 15613366, PMID 17720284, PMID 30645746, PMID 30165499
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。
What Remains After Control: Information Increment in Feng Shui Ratings and Cross-Regional Replicability
Evidence level: B
For a variable to show “incremental value” after controlling for two geophysical variables, four hurdles must be cleared: whether the incremental metric itself is susceptible to bias, whether control variables lie downstream of the grading rather than upstream, whether inconsistent directions across regions constitute failure, and whether a common scale exists for cross-regional grading.
Main Text
“Information gain” is not a single quantity in statistics but a family of metrics: discrimination increment, reclassification improvement, overall discrimination improvement, and net benefit. They can yield contradictory answers for the same dataset. Kerr’s systematic review notes that reclassification improvement relies on arbitrarily chosen class boundaries, is highly sensitive to model calibration, and retains non-zero expectations even when new variables are completely useless (PMID 24240655). Pepe had earlier demonstrated that the relationship between odds ratios and the area under the curve is nonlinear: a strong risk factor with an odds ratio of 3 added to a baseline model with an area already at 0.80 might only push it to 0.82 (PMID 15105181). Conversely, a weak variable with an odds ratio of only 1.2 can yield a larger increment in a low-baseline model. Thus, “how large the increment is” does not indicate variable strength but rather the degree of overlap with the existing model. Steyerberg’s framework decomposes discrimination, calibration, reclassification, and net benefit into four distinct tasks, explicitly stating they are not interchangeable (PMID 20010215); decision curves provide the “threshold range where net benefit becomes positive” (PMID 17099194), do not produce a single number, and are therefore less susceptible to cherry-picking for aesthetic appeal.
The second hurdle is more critical: including indoor radon and electrical conductivity as controls implies the assumption that they lie upstream or on a parallel path in the causal diagram. However, if “adverse sites” (形煞) traditionally include criteria for dampness, lack of sunlight, and poor ventilation, then indoor radon and humidity are not neutral geophysical proxies but carry the imprint of the grading itself. Controlling for them effectively controls part of the treatment, a phenomenon Schisterman terms overcorrection (PMID 19525685); if a control variable is simultaneously influenced by the grading and another common cause, it acts as a collider, and controlling for it can artificially induce an association (PMID 29040562). The determination method is straightforward: before measuring radon, code separately whether the property conforms to traditional criteria for dampness and ventilation, then assess the strength of correlation with radon. If the correlation is strong, radon should be excluded from the main model and included only in sensitivity analyses.
The third hurdle concerns the word “replication” itself. A common error is to test each region separately and count significant findings. The power of interaction tests is inherently low; the sample size required to detect an interaction is often several times that needed for a main effect (PMID 17578572). Performing interaction tests across six regions means finding one significant result is merely luck. A more robust criterion is the sign: preset the direction of the interaction term and count how many of the six regions align. If there is no true interaction, the probability of all six regions showing the same direction is $(1/2)^5$, approximately 3.1%—significantly more stable than individual testing under the same sample size. However, this has a prerequisite: the direction must be fixed prior to data inspection; otherwise, selecting one direction from six post-hoc inflates the probability to 100%. Systematic evaluations of subgroup analyses repeatedly confirm this point, where subgroups pre-specified in trial protocols often do not match those reported in the final papers (PMID 25030633).
The fourth hurdle is most easily overlooked: the grading may lack a common scale across regions. In blinded grading, when experts evaluate twenty properties simultaneously within one region, they perform relative ranking within that region rather than placing them on a single national standard ruler. This leads to two consequences: first, including region fixed effects in the model removes between-group variance in grading, leaving only within-group variation; second, omitting them causes the grading to partially act as a proxy for regional identity, becoming collinear with geological background, meaning the measured “information gain” reflects regional differences rather than universal predictive power. Furthermore, inter-rater reliability fluctuates with marginal distributions—the denominator of the agreement coefficient contains the expected agreement rate, so identical judgment behaviors yield different values in regions with different proportions of good and bad sites (see grading benchmarks in PMID 843571). Therefore, “0.5 in both regions” does not imply equivalent grading quality.
Judgments should be reported in three layers without synthesizing a single score. Prerequisite thresholds: first report inter-rater reliability and the ratio of good to bad sites; compare across regions only on whether they fall into the same category, not on specific values. Primary criterion: preset a single main model, report discrimination increment with bootstrap confidence intervals and full permutation zero distributions (re-running the entire selection process), and report decision curve net benefit; do not report reclassification improvement. Replication criterion: report the consistency rate of interaction signs across six regions (one-sided sign test), add individual data pooled interaction estimates and heterogeneity variance (PMID 28057641), and calculate sample size requirements based on external validation standards (PMID 34031906).
The Guanzi · Diyuan states, “All grass and soil paths have their own valleys; some are high, some low, each with its own grass and soil”—the same set of criteria yields different specific answers at different elevations; the Zang Shu* states, “Soil should be fine and firm, moist but not shiny”—criteria are physical properties observable on-site, not scores that can be mailed. Classical geography never provided a cross-regional scale of auspicious or inauspicious ratings; categories like “upper auspicious, middle auspicious, lower inauspicious” were rigidified later for teaching and transaction purposes. Therefore, inconsistent interaction directions across six regions are not failures of replication in ancient models but normal outputs—the model itself states “varies by location.”
Evidence and Limitations
Verified: The dose-response relationship between indoor radon and lung cancer holds stably across individual data pooled from 13 European case-control studies (PMID 15613366); the spatial distribution of radon is strongly constrained by geological units, allowing geological categories to serve as external drift terms in spatial modeling (PMID 17720284), and national surveys confirm that indoor radon exhibits clear regional spatial structure (PMID 30645746); seasonal correction factors systematically alter radon readings for the same residence (PMID 30165499). Methodologically, the dependence of reclassification improvement boundaries on arbitrary choices, non-zero expectations under the null hypothesis, the nonlinear relationship between discrimination increment and odds ratios, low power of interaction tests, overcorrection, and collider bias are all established (PMID 24240655, 15105181, 17578572, 19525685, 29040562).
Partially Verified: There is currently no direct answer from a three-variable synchronous blinded cohort regarding whether “Feng Shui grading retains independent information gain after controlling for geophysical variables”; furthermore, no empirical data exists regarding cross-regional interaction direction consistency. This paper provides criteria and design, not conclusions.
To Be Verified: ① The coupling intensity between traditional “adverse site” criteria and indoor radon, humidity, and ventilation—this determines whether radon enters the main model or only sensitivity analysis; ② Whether within-group variation in blinded grading within a single region is sufficient to support any increment estimation; ③ Whether cross-regional interaction sign consistency rates can exceed the 3.1% null hypothesis line in real-world cohorts.
Related Reading
- Feng Shui Auspicious/Inauspicious Grading, Ground Electrical Conductivity, and Indoor Radon Exposure: What Can a Three-Variable Joint Blinded Cohort Prove
- Quantitative Mapping Models for Adverse Site Grading and Fault Distance, Lithology, Soil Radon, and Conductivity
- Systematic Measurement Comparison of Feng Shui “Auspicious/Inauspicious Sites” with Ground Electrical Conductivity Distribution
- Research Pool: Environment-Life Frequency Interface
Evidence Level: B
References: PMID 24240655, PMID 15105181, PMID 20010215, PMID 17099194, PMID 19525685, PMID 29040562, PMID 17578572, PMID 25030633, PMID 843571, PMID 28057641, PMID 34031906, PMID 15613366, PMID 17720284, PMID 30645746, PMID 30165499
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