「控制断层距离后评级还剩多少」:共线的时候,减法是三选一,而数据不告诉你选哪个

Written by

in

Language

Evidence level: B

当两个变量相关到 r=0.9 时,「控制掉一个、看另一个还剩多少」不再是一个动作,而是三个互不相同的动作;而选择哪一个,数据本身没有发言权。

正文

一个很自然的问法:凶地评级和岩性、电导率这些物探变量,到底是不是同一件事?做法也顺手——把断层距离当协变量控制掉,再看评级和岩性的系数还显不显著。这一步看似只是「去掉一个干扰项」,实际已做了三个假设:断层在因果上位于上游、两个效应可线性相加、公共解释的那块方差可被切干净。三层没有一层白送。

第一层,减法本身有三条路,给三个数。残差化(评级先对断层距离回归、取残差再对岩性回归)、联合模型(同时放进方程)、分层比较(先只放断层距离,再放岩性看 ΔR²)——三条路在完全共线时给出三个不同答案,而且都不算错。因为 R² 可分解成「独有 A + 独有 B + 共同」三块,25495950 用模拟数据把这套分解(commonality analysis)做了一遍,指出共同方差既不能归给 A 也不能归给 B,只能单列。所以「控制后还剩多少」的答案,取决于你把共同那块判给了谁。

第二层,VIF 是放大器指标,不是信息量指标。有人用 20 名射手做多重目标追踪,年龄与训练年限 r=0.928、VIF=11.7,于是两个追踪指标变成了「显著的偏相关(r≈0.49,p<0.03)」,而训练年限的系数翻了号(B=-2.12,尽管零阶相关是 +0.65);换成只放年龄的模型(VIF<2.5),这些效应全部退到 p>0.13(42807114)。这就是共线性下的典型产物:一个被放大的标准误,加一个被误读成机制的反号系数。VIF 高只说明「这个系数不好读」,不说明「这个变量没有信息」。

第三层,「共同方差」不是零也不是对半分。25495950 的结论是共同部分往往占大头,此时「独有贡献」数学上接近 0 是必然的,不是「岩性没用」。39280972 给了个约定式解法:按「平均共享方差」把总 R² 拆成可相加的个体 R²,在 GAM 里对付 concurvity。但那是约定不是发现——它让数字能加总,没让不可归属变得可归属。

第四层,残差化在空间数据上更糟。取残差后拿它当因变量会产生生成回归量问题;更要紧的是空间结构:40705491 证明自变量与空间相关残差相关时,标准空间回归有偏、区间覆盖率不达标,须用两阶段(去空间趋势后再估参数)才拿得到名义覆盖率。断层距离和岩性都是空间强结构变量,残差也带着空间结构,于是「残差里的独立信号」很容易是没去干净的时空趋势。

第五层,非线性与交互是默认,不是例外。断层距离对室内氡的效应是阈值型的:小于 150 m 的住宅有 22.8% 超过 200 Bq/m³,150–700 m 降到 15.2%;而只在富铀基岩上算,这两个数是 30.7% 和 17.5%(26630035)。注意后一对数字差距更大——说明断层距离与岩性之间有交互,不是可加的两项。把断层距离当线性协变量控制掉,等于把交互项硬塞进主效应,残差里剩下的「岩性信号」必然是脏的。30649165 给了同构证据:连续协变量的非线性被线性模型忽略时,估计会产生实质偏倚。

第六层,「不显著」不等于「没有独立贡献」。共线性抬高标准误、压低功效,于是「控制后 p>0.05」在几十个点的样本里几乎必然发生,无论真值如何。42807114 那篇自己算了:要检出 f²=0.15 需 N≈68,实际样本是 20。这是设计的算术,不是证据。要判「没有独立贡献」,须做等效检验(先定最小有意义增量,如 ΔR²=0.02,再看区间是否整个落在界内),而不是看 p 值。另,把断层距离切成「近/远」两档是最常见的自伤——41802432 指出任意切点会造成信息损失与功效下降。

第七层,特征重要性的排序在共线下不稳。41749845 在乳腺癌多组学数据上发现,不到 0.1% 的特征被移除,监督模型的 Top-20 排序就会翻盘;41616464 在 1055 种化学品的 QSAR 数据上给出更刺眼的对照:XGBoost 准确率最高(0.8569)但排序最不稳定,无监督方法准确率略低(0.819–0.849)却完全稳定。机制在 19460890 与 20187966 里写得很清楚:Gini 重要性在相关预测变量下有偏,置换重要性在零假设下无偏、但在相关存在时会把重要性摊给整组。所以「岩性的 SHAP 值高于电导率」这类话,在共线设计里不能当结论。

这一层正好照出传统那套做法的结构。形峦派讲「龙、穴、砂、水、向」,从不把这五样当五个可互相控制的自变量——口诀「寻龙先看祖与宗,次看过峡与束气」讲的是先后与层级,不是并列与权重。龙定骨、砂定势、水定界,谁在上游谁在下游,古人心里有序次;现代统计把这种次序叫「因果层级」,把缺了次序的并列变量叫「共线」。所以这个问法古人不会提:他先问「这一处的龙是谁、砂是谁」,定了主从,再谈分量。分寸在于:那是可辩论的假设(各派对「龙」的定义都不同),不是已验证的因果图;把它当先验写进 DAG 再用数据检验,才是能往前走的用法。

怎么办,四件事。不做残差化,做联合模型加 commonality 分解,报「独有评级 / 独有岩性 / 共同」三块数。断层距离按非线性进模型,并显式放入「断层距离 × 岩性」交互项。用空间块交叉验证、必要时用两阶段去空间趋势(40705491 的 DSR 思路),否则残差里的空间结构会把独立贡献喂大。特征重要性只报重抽样下的排序一致性,不报单次 SHAP 排名。

证据与局限

已验证:共线性会使系数标准误膨胀、系数可反号,这是统计学的确定结论,42807114 提供了一个 r=0.928、VIF=11.7 的完整现场记录。共同方差在方差分解中不可归属,25495950 与 39280972 从两个方向确认。相关预测变量下特征重要性排序不稳定,41749845、41616464、19460890、20187966 四篇独立复现。

部分验证:断层距离对氡的阈值型效应有实测支撑(26630035,640 个地下室测点),但 150 m 这个切点是数据驱动的、不是物理常数;「断层距离与岩性存在交互」是从两组分层数字(22.8%/15.2% 与 30.7%/17.5%)反推的,原研究没有报告交互项检验。空间两阶段估计(40705491)是在方法学模拟中验证的,尚未在氡—地质场景里落地。

待验证:commonality 分解在「评级—物探」这个具体变量组上的实际拆分比例,目前没有任何实测;「等效检验界值」取 0.02 还是 0.05,需要按决策后果定,本文给的是示例值而非标准。

相关阅读

证据等级: B
参考文献: PMID 26630035, PMID 36497982, PMID 25495950, PMID 39280972, PMID 41749845, PMID 41616464, PMID 19460890, PMID 20187966, PMID 42807114, PMID 40705491, PMID 30649165, PMID 41802432, PMID 36323796


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 Rating Remains After Controlling for Fault Distance: When Collinearity Occurs, Subtraction Offers Three Choices But Data Reveals None

Evidence level: B

When two variables correlate at $r=0.9$, the question of “controlling for one to see how much remains with the other” ceases to be a single action and becomes three mutually distinct actions; furthermore, the data itself holds no authority on which choice to make.

Main Text

A natural inquiry arises: Is the rating of a “bad site” identical to geological properties and geophysical variables such as electrical conductivity? The practical approach is straightforward: treat fault distance as a covariate to control for, then examine whether the coefficient linking rating to lithology remains significant. This step appears merely as “removing one confounder,” yet it implicitly assumes three conditions: that the fault lies upstream in causality, that the two effects are linearly additive, and that the shared variance can be cleanly partitioned. None of these layers is free of assumption.

First, subtraction itself offers three pathways yielding three numbers: residualization (regressing rating against fault distance first, taking the residuals, then regressing against lithology), joint modeling (simultaneously including both in the equation), or stratified comparison (first modeling only fault distance, then adding lithology to observe $\Delta R^2$). Under perfect collinearity, these three paths yield different answers, none of which are inherently “wrong.” Because $R^2$ decomposes into “Unique A,” “Unique B,” and “Shared” components, 25495950 utilized simulation data to perform a commonality analysis demonstrating that shared variance cannot be attributed solely to A or B but must be listed separately. Therefore, the answer to “how much remains after controlling” depends entirely on how one assigns the shared component.

Second, VIF is an amplification indicator, not an information quantity indicator. Consider a study with 20 archers tracking multiple targets: age and training duration correlate at $r=0.928$ with a VIF of 11.7. Consequently, two tracking indicators become “significant partial correlations ($r\approx0.49$, $p<0.03$),” while the coefficient for training duration flips sign ($B=-2.12$, despite a zero-order correlation of +0.65); conversely, in a model containing only age (VIF<2.5), these effects retreat to $p>0.13$ (42807114). This represents a typical artifact under collinearity: an inflated standard error paired with a misinterpreted reversed-sign coefficient. A high VIF merely indicates that “this coefficient is difficult to interpret,” not that “this variable lacks information.”

Third, “shared variance” is neither zero nor exactly half. 25495950 concludes that the shared component often dominates; thus, when “unique contribution” mathematically approaches zero, it is a necessity, not evidence that “lithology is useless.” 39280972 offers a conventional solution: decomposing total $R^2$ into additive individual $R^2$s based on “average shared variance” to handle concurvity in GAMs. However, this is convention rather than discovery—it allows numbers to sum but does not render the unattributable attributable.

Fourth, residualization performs worse with spatial data. Using residuals as the dependent variable introduces endogeneity issues; more critically, regarding spatial structure: 40705491 demonstrates that when independent variables correlate with spatial residuals, standard spatial regression yields biased estimates and fails to meet nominal coverage rates, requiring a two-stage approach (removing spatial trends before estimating parameters) to achieve nominal coverage. Since both fault distance and lithology are variables with strong spatial structure, their residuals also carry spatial structure; thus, the “independent signal” remaining in the residuals is likely contaminated by unremoved spatiotemporal trends.

Fifth, nonlinearity and interaction are defaults, not exceptions. The effect of fault distance on indoor radon is threshold-based: residences within 150 m show 22.8% exceeding 200 Bq/m³, dropping to 15.2% between 150–700 m; however, calculating only on uranium-rich bedrock yields 30.7% and 17.5%, respectively (26630035). Note the larger gap in the latter pair—indicating interaction between fault distance and lithology, not merely additive terms. Treating fault distance as a linear covariate to control for is equivalent to forcing an interaction term into the main effect; the remaining “lithology signal” in the residuals will inevitably be contaminated. 30649165 provides isomorphic evidence: ignoring nonlinearity in continuous covariates via linear models produces substantial bias in estimates.

Sixth, “non-significance” does not equate to “no independent contribution.” Collinearity inflates standard errors and reduces power, making “post-control $p>0.05$” nearly inevitable in samples of dozens of points regardless of the true value. 42807114 calculated this internally: detecting $f^2=0.15$ requires $N\approx68$, yet the actual sample size is 20. This is arithmetic of design, not evidence. To judge “no independent contribution,” one must perform an equivalence test (first defining a minimum meaningful increment, e.g., $\Delta R^2=0.02$, then checking if the confidence interval falls entirely within bounds), rather than relying on p-values. Furthermore, dichotomizing fault distance into “near/far” is a common self-harm—41802432 notes that arbitrary cutpoints cause information loss and reduced power.

Seventh, ranking feature importance is unstable under collinearity. 41749845 found in breast cancer multi-omics data that removing less than 0.1% of features can flip the Top-20 ranking in supervised models; 41616464 provided a more striking contrast in QSAR data for 1055 chemicals: XGBoost achieved the highest accuracy (0.8569) but showed the least stability in ranking, whereas unsupervised methods had slightly lower accuracy (0.819–0.849) yet were completely stable. The mechanism is clearly outlined in 19460890 and 20187966: Gini importance is biased under correlated predictors, while permutation importance is unbiased under the null hypothesis but distributes importance across the group when correlation exists. Thus, statements like “lithology’s SHAP value exceeds electrical conductivity” cannot be treated as conclusions in collinear designs.

This layer precisely mirrors the structural approach of traditional methodologies. The Xiang Luan school discusses “Dragon, Cave, Sand, Water, and Direction,” never treating these five as five independent variables mutually controllable. The mnemonic “Seek the Dragon first by looking at its ancestors and descendants, then examine the narrow passes and qi-binding” refers to sequence and hierarchy, not parallelism and weighting. The Dragon sets the bone, Sand sets the posture, Water sets the boundary; ancient masters understood the upstream-downstream order within their minds; modern statistics labels such ordering as “causal hierarchy,” while treating collinear variables lacking such order as “collinearity.” Therefore, this specific question would not arise in traditional thought: one first asks “Who is the Dragon here? Who is the Sand?” establishing primacy and subordination before discussing proportions. The nuance lies in the fact that this is a debatable hypothesis (different schools define “Dragon” differently), not a verified causal graph; incorporating it as a prior into a DAG and then testing with data is the only way to proceed forward.

What to do? Four things: Avoid residualization; use joint modeling with commonality decomposition, reporting the three components (“Unique Rating,” “Unique Lithology,” “Shared”). Model fault distance non-linearly and explicitly include an interaction term (“Fault Distance × Lithology”). Use spatial block cross-validation, employing the two-stage spatial trend removal (DSR approach from 40705491) when necessary, otherwise spatial structure in residuals will inflate independent contributions. Report feature importance only based on ranking consistency under resampling, not single-run SHAP rankings.

Evidence and Limitations

Verified: Collinearity causes coefficient standard errors to inflate and coefficients to flip signs; this is a definitive statistical conclusion, supported by 42807114 which provides a complete field record with $r=0.928$ and VIF=11.7. Shared variance is unattributable in variance decomposition, confirmed by 25495950 and 39280972 from two directions. Feature importance rankings are unstable under correlated predictors, independently replicated across four studies: 41749845, 41616464, 19460890, and 20187966.

Partially Verified: Threshold effects of fault distance on radon have empirical support (26630035, 640 basement measurement points), though the 150 m cutpoint is data-driven rather than a physical constant; the existence of interaction between “fault distance and lithology” is inferred from two stratified sets of numbers (22.8%/15.2% vs 30.7%/17.5%), but the original study did not report an interaction term test. The two-stage estimation method (40705491) was verified through methodological simulations and has not yet been implemented in the radon-geological scenario.

To Be Verified: The actual proportion of splitting via commonality decomposition for the specific variable group “Rating vs. Geophysical” currently lacks empirical data; whether the “equivalence test threshold” should be 0.02 or 0.05 requires determination based on decision consequences; the values provided in this text are examples rather than standards.

Related Reading

Evidence Level: B
References: PMID 26630035, PMID 36497982, PMID 25495950, PMID 39280972, PMID 41749845, PMID 41616464, PMID 19460890, PMID 20187966, PMID 42807114, PMID 40705491, PMID 30649165, PMID 41802432, PMID 36323796


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 *