Evidence level: B
校正模型里混入一个”干预和结局共同产生的变量”(collider),比混入一个无关变量危险得多——它会在没有任何真实因果的地方凭空制造关联,让”干预有效/无效”的判断整体失真。
正文
上一轮我们处理了”校正过头”的问题:把中介变量当混杂变量控制,会把真实信号压向零。这一轮处理它的镜像问题——把 collider 变量当混杂变量放进模型。如果说中介误杀是”误杀忠臣”,collider 混入就是”引狼入室”:它不仅不帮忙去混杂,还会凭空造出虚假关联。
什么是 collider?流行病学的经典例子来自 Cole 等人的教科书式说明(PMID 19926667):流感和变质的三明治都能引起发烧,那么”发烧”就是这两个原因的共同效应。如果你只研究”发烧的人”,并问”吃三明治会不会得流感”,你会得出荒谬的结论——因为发烧人群里,没吃三明治的人几乎都是流感来的,三明治和流感就这样被”拧”出了虚假的负相关。这个共同效应变量(发烧)就是 collider:两个箭头同时指向它。对 collider 做任何形式的校正(分层、回归控制、按它筛选样本),都会制造出原本不存在的关联,或者扭曲原本真实的关联。这就是 collider-stratification bias(选择偏倚的一种)。
collider 和混杂的关键区别在于:混杂变量是”该调的”(控制掉它才能看到干净关联),collider 是”一调就坏”的。混杂像路上的岔口,你不堵它,信号会走岔;collider 像一块磁铁,你不碰它没事,一碰它反而把铁屑吸到一起。Luque-Fernandez 等人用钠摄入与血压的模拟数据演示了这一点(PMID 30561628):真实因果系数是 0.3,把尿蛋白(collider)加进回归后,估计值立刻偏离真实值——他们专门做了交互式网页应用让读者亲眼看到”加一个变量,结论反转”的过程。
GDV(气体放电可视化)研究为什么会撞上 collider?因为我们的分析对象是干预前后的状态测量。想象一个典型的场景:被试做某干预(比如观想、呼吸练习),我们采集干预后的 GDV 残差信号,想判断”干预是否改变了信号”。此时,干预后测得的任何状态变量(比如干预后的情绪评分、干预后的心率、甚至残差信号本身的一部分)都天然带有 collider 气质——它同时受”干预”和”被试的基线状态/未测混杂”影响。谁最危险?那些看起来”很有道理、必须控制”的变量:干预后情绪、干预后生理状态、被试自我报告的好转程度。研究者越想”控制全面”,越容易把这些变量塞进模型,结果就是 Griffith 等人在 COVID-19 研究里警告的场景(PMID 33184277):对检测人群(本身就是被选择出来的样本)的回归分析,会系统性扭曲风险因素的估计,甚至让真正的危险因素看起来”有保护作用”。GDV 研究的小样本 + 重复测量设计,恰恰是这种偏倚的高发地带。
那么”损伤到底多大”?这就是本篇要实测的东西。设计上沿用我们 GDV 系列的统一框架:梯度纳入 + 刻意污染对照。具体三步:第一,确定一组已知的 collider 结构变量(干预后测得的、受干预与未测混杂共同影响的变量,如干预后情绪评分),把它们作为”污染组”;第二,跑两条线的判别分析——”无 collider 对照组”(只用基线协变量校正)vs”collider 混入组”(把污染组变量加进校正模型),各自评估干预判别 AUC;第三,用配对设计的 AUC 差值置信区间方法(PMID 32390284)检验两组差异,并绘制”collider 伤害幅度—样本量”曲线——因为 collider 偏倚在小样本里常常被抽样噪声掩盖,样本量越大,伤害越可能被看清楚。
理论上,collider 伤害的量级取决于三个参数:collider 与干预的关联强度、collider 与结局的关联强度、以及未测混杂的强度(PMID 34939918 的系统综述给出了这一框架)。伤害的方向并不固定——它可能抬高 AUC(虚假的”干预有效”),也可能压低 AUC(把真实效果搅浑)。Tönnies 等人特别强调:collider 偏倚最阴险的地方是它无法靠”再加几个变量”补救,只能用因果图(DAG)预先排除。所以本篇的实测结论如果显示”混入 collider 后 AUC 变化超过预先设定的等价界”,就等于给 GDV 分析立下一条铁律:干预后测得的变量,一律不得进入校正模型,除非有先验因果图证明它不是 collider。
落到易学桥接的语境:GDV 实验想验证的是”调神/观想/环境布局与人体气场信号之间的真实关联”。如果校正模型里混入了干预后的状态变量,就可能把”静坐后放松感”这类主观回报(它既是干预的结果,又受期望效应影响)当成协变量,从而制造出虚假的”干预有效”或”干预无效”。这一条与 #838 的”校正深度曲线”互为镜像:一个管”调了太多该调的中介”,一个管”调了不该调的 collider”。两道闸门都守住,残差信号才能说自己是干净的。
证据与局限
已验证: collider 偏倚的机制与方向不定性已被模拟与实证反复确认(PMID 19926667 教学经典、PMID 30561628 钠-血压悖论演示、PMID 33184277 UK Biobank 实证);对 collider 做校正会系统性扭曲暴露-结局关联。
部分验证: collider 伤害幅度与三个强度参数(collider-暴露、collider-结局、未测混杂)的关系已有理论框架与模拟支撑(PMID 34939918),但在 GDV 这种小样本重复测量场景下的实际量级尚无直接实测——这正是本篇设计要补的空白。
待验证: GDV 场景中 collider 混入对干预判别 AUC 的伤害幅度是否超过预设等价界;伤害幅度-样本量曲线的形态;干预后情绪评分在真实 GDV 实验中是否构成可探测的 collider。
相关阅读
证据等级: B
参考文献: PMID 19926667, PMID 33184277, PMID 30561628, PMID 34939918, PMID 32390284
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 Does Discriminative Ability Suffer When Collider Variables Are Mixed into the GDV Correction Model?
Evidence level: B
Introducing a “variable generated by both the intervention and the outcome” (collider) into a correction model is far more dangerous than introducing an irrelevant variable—it artificially generates associations in the absence of true causality, causing systematic distortion in judgments regarding whether an intervention is effective or not.
Main Text
In the previous iteration, we addressed the issue of “over-adjustment”: treating mediators as confounders suppresses true signals toward zero. This iteration addresses its mirror problem—treating a collider variable as a confounder in the model. If mediator misadjustment is “killing the loyal minister,” collider inclusion is “inviting wolves into the house”: it does not merely fail to assist in confounding control; it artificially creates spurious associations.
What is a collider? A classic example from epidemiology comes from Cole et al.’s textbook explanation (PMID 19926667): both the flu and spoiled sandwiches can cause fever, making “fever” a common effect of these two causes. If you only study “people with fever” and ask whether eating sandwiches causes the flu, you arrive at an absurd conclusion—because among the fever population, those who did not eat sandwiches are almost certainly cases of the flu, creating a spurious negative correlation between sandwiches and the flu. This common effect variable (fever) is the collider: two arrows point simultaneously toward it. Any form of adjustment for a collider (stratification, regression control, or sample selection based on it) generates associations that do not exist or distorts true associations. This is known as collider-stratification bias (a type of selection bias).
The key distinction between a collider and a confounder lies in that confounders should be adjusted for (controlling them reveals the clean association), whereas colliders break when adjusted. A confounder is like a fork in the road; if you do not block it, the signal takes a wrong turn. A collider is like a magnet; if you do not touch it, nothing happens, but touching it causes iron filings to clump together. Luque-Fernandez et al. demonstrated this using simulated data on sodium intake and blood pressure (PMID 30561628): the true causal coefficient is 0.3, but adding urine protein (a collider) to the regression immediately deviates the estimate from the true value—they specifically built an interactive web application for readers to witness firsthand how “adding a variable reverses the conclusion.”
Why did GDV (Gas Discharge Visualization) research encounter colliders? Because our analysis targets state measurements before and after intervention. Imagine a typical scenario: participants undergo an intervention (e.g., visualization, breathing practice), and we collect post-intervention GDV residual signals to determine whether the intervention altered the signal. At this stage, any state variable measured post-intervention (such as post-intervention mood scores, post-intervention heart rate, or even a portion of the residual signal itself) inherently possesses collider characteristics—it is simultaneously influenced by “the intervention” and “baseline states/unmeasured confounders.” Which variables are most dangerous? Those that appear “plausible and must be controlled”: post-intervention mood, post-intervention physiological states, or self-reported improvement ratings. The more researchers strive to “control comprehensively,” the more likely they are to include these variables in the model, resulting in the scenario warned by Griffith et al. in COVID-19 research (PMID 33184277): regression analysis on screened populations (samples already selected) systematically distorts risk factor estimates, even causing true risk factors to appear as having a “protective effect.” GDV research’s small sample size and repeated-measures design are precisely high-risk zones for such bias.
So, “how much is the damage?” This is what this paper empirically tests. Following our GDV series’ unified framework, the design employs gradual inclusion + deliberate contamination of controls. Specifically: First, identify a set of known collider structural variables (post-intervention measures influenced by both intervention and unmeasured confounding, such as post-intervention mood scores) and designate them as the “contamination group”; Second, run two lines of discriminant analysis—”No Collider Control Group” (using only baseline covariates for adjustment) vs. “Collider Inclusion Group” (adding contamination group variables to the correction model)—each evaluating intervention discrimination AUC; Third, use paired-design AUC difference confidence interval methods (PMID 32390284) to test group differences and plot a “Collider Damage Magnitude—Sample Size” curve—because collider bias is often masked by sampling noise in small samples, larger sample sizes make the damage more visible.
Theoretically, the magnitude of collider damage depends on three parameters: the strength of association between the collider and the intervention, the strength between the collider and the outcome, and the strength of unmeasured confounding (PMID 34939918’s systematic review provides this framework). The direction of damage is not fixed—it may inflate AUC (false “intervention effective”) or deflate AUC (muddling the true effect). Tönnies et al. particularly emphasize: the most insidious aspect of collider bias is that it cannot be remedied by “adding more variables,” but must be excluded a priori via causal diagrams (DAGs). Therefore, if our empirical results show that “AUC changes after collider inclusion exceed a pre-set equivalence bound,” it establishes an iron rule for GDV analysis: post-intervention variables must never enter correction models unless a prior causal diagram proves they are not colliders.
In the context of bridging to Yijing Studies: GDV experiments aim to verify “true associations between tuning the mind/visualization/environmental layout and human Qi-field signals.” If post-intervention state variables are included in the correction model, subjective returns like “relaxation after sitting still” (which is both an outcome of the intervention and affected by expectancy effects) might be treated as covariates, thereby manufacturing false judgments of “intervention effective” or “intervention ineffective.” This mirrors the “correction depth curve” in #838: one governs “adjusting too much of the legitimate mediators,” while the other governs “adjusting inappropriate colliders.” Both gates must be secured for residual signals to claim cleanliness.
Evidence and Limitations
Verified: The mechanisms and directionality of collider bias have been repeatedly confirmed through simulation and empirical studies (PMID 19926667 textbook classic, PMID 30561628 sodium-blood pressure paradox demonstration, PMID 33184277 UK Biobank empirical evidence); adjusting for colliders systematically distorts exposure-outcome associations.
Partially Verified: The relationship between collider damage magnitude and three intensity parameters (collider-exposure, collider-outcome, unmeasured confounding) has theoretical framework and simulation support (PMID 34939918), but actual magnitudes in GDV’s small-sample repeated-measures context lack direct empirical measurement—this is precisely the gap this paper aims to fill.
To Be Verified: Whether collider inclusion in GDV contexts causes intervention discrimination AUC damage exceeding the pre-set equivalence bound; the shape of the damage magnitude–sample size curve; whether post-intervention mood scores constitute a detectable collider in real GDV experiments.
Related Reading
- Quantification of Discriminatory Damage from “Over-adjustment” in GDV Correction Models
- “Correction Depth Dependency” Curve of Net GDV Residual Information Quantity
- Adjudication of “Unimodal vs Monotonic” Morphology in GDV Signal Decay Curves
- Research Pool: Spacetime Frequency Topology
Evidence Level: B
References: PMID 19926667, PMID 33184277, PMID 30561628, PMID 34939918, PMID 32390284
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