吉地「舒适感」的方差分解:景观效应与接地效应各占几成,以及为什么「几成」这个数字不可信

Written by

in

Language

Evidence level: B

把”吉地舒服”拆成景观和接地两条通路之后,下一个问题听起来只是个除法:各占几成?但这道除法的分子和分母,都会随你选用哪种分解方法、信度多高、以及把谁当成原因而改变——所以”接地只占一成”这句话,几乎从来不是一个科学结论,而是一个模型设定的副产品。

正文

假设你已经做完那个关键实验:2×2 交叉设计——同一批被试轮流进入”景观好+真接地””景观好+假接地””景观差+真接地””景观差+假接地”四种条件,每次驻留测 HRV、唾液皮质醇、主观舒适度。数据到手,第一反应是跑多水平回归:舒适度 = 景观 + 接地 + 景观×接地 + 随机效应,然后看两个系数谁的标准化 β 大。这在统计上没错,在解释上几乎一定翻车。原因有五个。

第一,”方差分解”不是一种算法,是一族算法,给出的”占比”并不相同。 part R²(把某项拿掉、看解释方差掉了多少)算的是唯一贡献;结构系数换算的 inclusive R² 算的是总贡献(含共享部分);Shapley 值分解、相对权重法、优势分析、LMG 法各有一套分配共享方差的规则。同一份数据,这些方法排序有时一致、数值却可能差出十几个百分点(PMID 42375422)。”景观六成、接地四成”这句话,换个分解法就可能变成”七成对三成”——数据一个字节没动。

第二,真正的黑洞是”共享方差”,它既不属于景观也不属于接地。 现实中的吉地几乎必然”景观好”和”能接地”同时成立,因为有共同成因:选了一块好地方。所以两者协方差不是噪声,而是主体。拆这一块的工具叫共同性分析(commonality analysis),把总解释方差切成”唯一属于 A””唯一属于 B””A 与 B 共享”三块(PMID 40889234、31891314)。风水语境里共享块最有意思:它是两条通路共同产生、无法归给任何单一通路的效应,也是”气”能被操作化的上界——能被拆开的部分,已经分别是”景观”和”接地”,不再是”气”了。

第三,测量误差会系统性压低某条通路的占比,而且压向零。 景观测量相对容易:照片评分、绿视率、NDVI、可见天空比例,客观可重复。接地测量则粗糙得多——赤脚还是穿鞋、接地垫通不通、鞋底电阻多大、地面湿度多少,常被压缩成一个二分变量。测量误差在回归里造成回归稀释(regression dilution)偏倚:信度越低,系数被低估越厉害(PMID 37336529)。后果很具体:“景观贡献大于接地”可能完全来自两条通路的测量精度差异,而不是真实的效应差异。 要排除这个可能,必须报告接地变量的重测信度(ICC),并用回归校准(regression calibration)把衰减的系数修正回来。

第四,最容易被忽略的一层:把”结果”当成并列的”原因”,整个分解就失去意义。 景观和接地为什么共线?因为都是”选了一块好地方”的下游产物。若把接地当成与景观并列的独立预测项,等于假设两者是平行输入,而接地其实可能是选址的结果之一(土质好、植被密、人愿意赤脚)。这点在比较生物学里被明确讨论过:预测项与响应变量的共线性本身不是偏倚主因,模型设定与真实因果结构错配才是(PMID 35784454)。翻译成风水语言:如果接地本身是”好地方”的副产品,问”舒适感里有多少是接地给的”,等于问”发烧里有多少是体温计给的”。

第五,”几成”这个数字对样本结构的敏感度远超直觉。 多水平模型的 R² 有多个版本(边际、条件、层级专属),估计精度强烈依赖簇的数量:层 2 预测项越多,需要的簇数越多(PMID 41848595、28904005、34113487)。风水研究的现实是地点只有十几个、观测却上千,这种”簇少观测多”的结构会让层 2 方差成分极不稳定——你算出的”接地占比”,可能有一半是抽样噪声。

那正确做法是什么?

能用设计分离的,不要用统计分离。 2×2 交叉设计的全部意义就在于把两条通路在设计层面正交化,而不是指望回归去解一个共线方程;同时必须报告携带效应(carry-over)检验与洗脱期设定(PMID 38563531、25115725)。

报告三个数,而不是一个数:唯一贡献(part R²)+ 总贡献(inclusive R²)+ 共享贡献。三者一起看才有信息量——如果共享块很大,说明两条通路在你的数据里根本分不开,此时任何”谁占几成”的说法都应当被拒绝。

占比只能用来排序,不能用来分成。 风水实践真正要回答的不是”各占几成”,而是”如果只能改一件事,改哪一件”——这对应的是可干预性与成本,不是方差占比。一块地的景观改不动,接地却可能只需换一双鞋。而且证据强度本身就不对称:接地文献多为小样本先导研究(12 人睡眠皮质醇试点,PMID 15650465)、非随机分组的生理实验(PMID 21469913)或叙述性综述(PMID 31831261、25848315);森林环境则已有系统综述与 Meta 分析支撑(唾液皮质醇 MD = −0.05 μg/dl,PMID 31001682;收缩压 −3.44 mmHg、舒张压 −3.07 mmHg,PMID 36612777)。这意味着在方差分解里,接地那一项的置信区间一定会比景观宽得多。

最后回到那个最容易被误读的数字上。方差分解给出的从来不是”真相的份额表”,而是”在你这个模型、这个测量精度、这个因果假设下,方差能怎么被切开”。换一个模型设定,切法就变。所以看到”景观占七成、接地占三成”这类说法时,该问的不是”真的吗”,而是:共线程度是多少?信度是多少?谁被当成了原因? 这三个问题答不上来,那个百分比就只是一个好看的空壳。

证据与局限

已验证:方差分解存在多种估计量且结果不同(PMID 42375422);多层模型 R² 的估计精度依赖簇数(PMID 41848595);测量误差导致回归稀释偏倚、可用回归校准修正(PMID 37336529);共线性不是偏倚主因、模型因果结构错配才是(PMID 35784454);森林环境的生理效应有系统综述/Meta 分析支撑(PMID 31001682、36612777)。

部分验证:接地的生理效应——有多项研究与综述报告正向结果,但样本普遍偏小、盲法与随机化描述不足,效应量与证据层级明显低于景观通路(PMID 31831261、21469913、15650465、25848315)。因此”两通路方差各占几成”这一问题,目前没有任何直接针对风水场景的实测数据,本篇给出的是方法学框架与可执行方案,不是实测占比。

待验证:(1)在 2×2 交叉设计数据上,part R²、inclusive R²、Shapley 分解三者对”景观/接地”占比的估计差异有多大?若排序不一致,则”占比”这一指标在该场景下应被弃用。(2)共同性分析中的”共享块”占比是否稳定大于两条通路的唯一贡献?若是,说明两条通路在本类数据中本质不可分。(3)把接地从二分变量升级为连续剂量变量(皮肤—地电位差 μV 级、接触电阻 kΩ 级)后,接地通路的 part R² 是否显著上升?若上升,则此前的”接地贡献低”应重新解释为测量假象。(4)在簇数 10–20、每簇观测 ≥500 的结构下,层 2 方差成分的 bootstrap 置信区间宽度是多少?宽度超过估计值本身则结论不可用。

相关阅读

证据等级: B
参考文献: PMID 42375422, PMID 40889234, PMID 31891314, PMID 37336529, PMID 35784454, PMID 41848595, PMID 28904005, PMID 34113487, PMID 31001682, PMID 36612777, PMID 38563531, PMID 25115725, PMID 15650465, PMID 21469913, PMID 31831261, PMID 25848315


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。

Variance Decomposition of Ji Di’s \”Sense of Comfort\”: The Proportions of Landscape and Grounding Effects, and Why the \”Several Parts\” Figure Is Unreliable

Evidence level: B

Decomposing “auspicious comfort” into two distinct pathways—landscape and grounding—makes the next question sound trivial: what proportion does each account for? However, the numerator and denominator of this division shift depending on the decomposition method selected, the reliability of measurements used, and which variable is designated as the cause. Therefore, the statement “grounding accounts for only one-tenth” is rarely a scientific conclusion but rather an artifact of model specification.

Main Text

Assume you have completed the critical experiment: a 2×2 crossover design where the same participants rotate through four conditions—”good landscape + real grounding,” “good landscape + false grounding,” “poor landscape + real grounding,” and “poor landscape + false grounding”—measuring HRV, salivary cortisol, and subjective comfort during each stay. Once data are obtained, the first instinct is often to run a multilevel regression: Comfort = Landscape + Grounding + Landscape×Grounding + Random Effects, then compare the standardized β coefficients. While statistically valid, this approach almost certainly fails in interpretation for five reasons.

First, “variance decomposition” is not a single algorithm but a family of algorithms yielding different proportions. Part $R^2$ (removing a variable to see how much explained variance disappears) calculates unique contribution; structural coefficient conversion for inclusive $R^2$ calculates total contribution (including shared parts); Shapley value decomposition, relative weighting methods, advantage analysis, and LMG methods each follow distinct rules for allocating shared variance. With the same dataset, these methods may rank similarly yet differ by tens of percentage points (PMID 42375422). The claim “landscape accounts for sixty percent, grounding forty percent” could shift to “seventy percent versus thirty percent” simply by changing the decomposition method, without altering a single byte of data.

Second, the true black hole is “shared variance,” which belongs to neither landscape nor grounding. Auspicious locations almost invariably satisfy both “good landscape” and “effective grounding” simultaneously due to a common cause: selecting a good site. Thus, their covariance is not noise but the subject itself. The tool for this is commonality analysis, which slices total explained variance into three blocks: “unique to A,” “unique to B,” and “shared by A and B” (PMID 40889234, 31891314). In Feng Shui contexts, the shared block is most interesting: it represents an effect jointly produced by two pathways and unassignable to a single pathway, serving as the upper bound for operationalizing “Qi”—the portion that can be decomposed is already categorized as “landscape” or “grounding,” no longer “Qi.”

Third, measurement error systematically depresses the proportion attributed to a specific pathway, pushing it toward zero. Landscape measurement is relatively straightforward: photo ratings, green view index, NDVI, visible sky proportion; these are objective and repeatable. Grounding measurement is far rougher—barefoot vs. shod, grounding pad continuity, sole resistance, ground humidity—often compressed into a binary variable. Measurement error in regression causes regression dilution bias: lower reliability leads to greater underestimation of coefficients (PMID 37336529). The consequence is concrete: “landscape contribution exceeds grounding” may stem entirely from differences in measurement precision between the two pathways, not from true effect differences. To exclude this possibility, one must report the test-retest reliability (ICC) of the grounding variable and apply regression calibration to correct the attenuated coefficients.

Fourth, the most overlooked layer: treating the “outcome” as a parallel “cause,” rendering the decomposition meaningless. Why are landscape and grounding collinear? Because they are both downstream products of “selecting a good site.” Treating grounding as an independent predictor parallel to landscape assumes they are parallel inputs, whereas grounding might actually be a result of site selection (good soil, dense vegetation, willingness to go barefoot). This is explicitly discussed in comparative biology: collinearity between predictors and response variables is not the primary source of bias; misalignment between model specification and true causal structure is (PMID 35784454). Translated into Feng Shui language: if grounding itself is a byproduct of a “good place,” asking “how much of comfort comes from grounding” is equivalent to asking “how much of the fever is caused by the thermometer.”

Fifth, the “proportion” figure is far more sensitive to sample structure than intuition suggests. Multilevel model $R^2$ has multiple versions (marginal, conditional, level-specific), and estimation precision heavily depends on the number of clusters: more level-2 predictors require more clusters (PMID 41848595, 28904005, 34113487). The reality of Feng Shui research is having only a dozen or so locations with thousands of observations; this “few clusters, many observations” structure makes level-2 variance components extremely unstable—the calculated “grounding proportion” might be half sampling noise.

So what is the correct approach?

Do not rely on statistical separation if design separation is possible. The entire significance of the 2×2 crossover design lies in orthogonalizing the two pathways at the design level rather than relying on regression to solve a collinear equation; simultaneously, carry-over effect tests and washout period settings must be reported (PMID 38563531, 25115725).

Report three numbers, not one: Unique contribution (part $R^2$) + Total contribution (inclusive $R^2$) + Shared contribution. Only viewing all three together provides information—if the shared block is large, it indicates that the two pathways are fundamentally inseparable in your data, at which point any claim of “who accounts for what percent” should be rejected.

Proportions are for ranking, not partitioning. The practical question in Feng Shui is not “what percent does each account for,” but “if only one change can be made, which one should be chosen?” This corresponds to interventability and cost, not variance proportion. A site’s landscape may be unchangeable, yet grounding might only require changing footwear. Furthermore, evidence strength is inherently asymmetric: grounding literature consists mostly of small-sample pilot studies (12-person sleep cortisol trial, PMID 15650465), non-randomized physiological experiments (PMID 21469913), or narrative reviews (PMID 31831261, 25848315); forest environments already have systematic reviews and meta-analyses supporting them (salivary cortisol MD = −0.05 μg/dl, PMID 31001682; systolic BP −3.44 mmHg, diastolic BP −3.07 mmHg, PMID 36612777). This means in variance decomposition, the confidence interval for the grounding term will be far wider than that for landscape.

Finally, return to the most easily misinterpreted number. Variance decomposition never provides a “table of truth shares”; it shows how variance can be sliced under your specific model, measurement precision, and causal assumptions. Change the model specification, and the slicing changes. Therefore, when encountering statements like “landscape accounts for seventy percent, grounding thirty percent,” do not ask “is this true?” but rather: What is the degree of collinearity? What is the reliability? Who was designated as the cause? If these three questions cannot be answered, that percentage is merely an attractive empty shell.

Evidence and Limitations

Verified: Variance decomposition yields multiple estimators with different results (PMID 42375422); multilevel model $R^2$ estimation precision depends on cluster count (PMID 41848595); measurement error causes regression dilution bias correctable via regression calibration (PMID 37336529); collinearity is not the primary bias source, but misalignment of model causal structure is (PMID 35784454); physiological effects in forest environments are supported by systematic reviews/meta-analyses (PMID 31001682, 36612777).

Partially Verified: Physiological effects of grounding—multiple studies and reviews report positive results, but samples are generally small, blinding and randomization descriptions are insufficient, and effect sizes/evidence levels are clearly lower than the landscape pathway (PMID 31831261, 21469913, 15650465, 25848315). Therefore, the question of “what proportion does each pathway account for” currently has no direct empirical data specifically for the Feng Shui scenario; this paper provides a methodological framework and executable protocol, not an empirical proportion.

To Be Verified: (1) In 2×2 crossover data, what is the estimation difference between part $R^2$, inclusive $R^2$, and Shapley decomposition regarding “landscape/grounding” proportions? If rankings differ, this metric should be abandoned in this context. (2) Is the proportion of the “shared block” in commonality analysis consistently greater than the unique contributions of both pathways? If so, the two pathways are essentially inseparable in this data type. (3) After upgrading grounding from a binary variable to a continuous dose variable (skin-ground potential difference in μV range, contact resistance in kΩ range), does the part $R^2$ for the grounding pathway increase significantly? If so, prior claims of “low grounding contribution” should be reinterpreted as measurement artifacts. (4) In a structure with 10–20 clusters and ≥500 observations per cluster, what is the bootstrap confidence interval width for level-2 variance components? If the width exceeds the estimate itself, the conclusion is unusable.

Related Reading

Evidence Level: B
References: PMID 42375422, PMID 40889234, PMID 31891314, PMID 37336529, PMID 35784454, PMID 41848595, PMID 28904005, PMID 34113487, PMID 31001682, PMID 36612777, PMID 38563531, PMID 25115725, PMID 15650465, PMID 21469913, PMID 31831261, PMID 25848315


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 *