煞是一样的,灾不一样:把「体质差异」放进交互项之前

Written by

in

Language

Evidence level: B

同一间屋子、同一个位置,有人难受有人没事——这个观察可能是真的,但”谁敏感”这件事能被测到什么程度,决定了它能不能被当成结论。

正文

风水与命理里有一句几乎无法反驳的话:煞是一样的,灾不一样。同一栋楼、同一个朝向、同一间卧室,有人睡得香,有人整夜翻。传统给的解释是个体化分层——命卦、八字、体质。现代流行病学给这类说法的名字叫效应修饰:暴露的效应量随第三个变量变化。名字对得上,但”对得上”和”测得出来”是两个问题。本篇只处理后者。

第一,交互项是一个乘积,不是加强版的主效应。

模型写成 E[Y]=β₀+β₁D+β₂S+β₃(D×S),β₃ 检验的是两条斜率之差,而每条斜率自己已经带着误差。这意味着 β₃ 的标准误几乎总是大于 β₁ 或 β₂,即使真存在效应修饰。样本量上的代价是硬的:检出交互通常要数倍于检出主效应的样本,而且不同算法在同一情景下算出来的数还能差出很多(PMID 10206617)。这不是保守,是算术。

第二,分层用的那把尺,得先问它准不准。

MSSQ(晕动易感性问卷)是这条线最常用的分层工具。作者称其重测信度可假定优于 0.8,但预测实验室晕动耐受的有效性只有 r≈0.45(PMID 10052582)——解释约两成方差。用一把只解释两成方差的尺去分组,组内混着的错分,比组间那点差别还大。

第三,错分打在交互项上的方向,和打在主效应上不一样。

非差异性错分通常把乘法交互的估计拉向零(看起来”没有修饰”),但对加法交互,它可能把估计推向远离零的方向(PMID 10613335)。同一个数据集,两种尺度,两个相反的偏倚方向。这就是为什么”我们没发现敏感度修饰效应”这句话本身信息量很低——它可能只是尺子钝。

第四,二分还是连续,会改变结论本身。

把 MSSQ 切成”易感/非易感”最省事,也最伤:切点任意、组内异质、高分区样本天然稀疏。所谓”两层斜率不同”,其实是”组均值差”的再包装。更稳的做法是让敏感度留在连续尺度上直接估 D×S 的斜率——但这又绕回第一层的功效问题。两条路都不轻松,选哪条得写清。

第五,问卷测的可能不是前庭。

MSSQ 与过去 12 个月非运动性恶心呕吐的相关是 r=0.3,其中偏头痛是最大贡献者(PMID 10052582)。这说明它测的更像是”最终共同呕吐通路的兴奋性”,一个比前庭宽得多的东西。另一条路给出更窄的答案:前庭感知阈值在个体内高度一致、个体间差异显著,作者直接称其为可作生物标志的先天特质(PMID 38658516);在驾驶模拟数据上,两个个体参数就能把晕动反应拟合改善 1.7 倍(PMID 40589464)。能用两个数描述的个体差异,不必用一张回忆问卷去代理。

第六,症状这一端也不老实。

sopite 综合征(嗜睡、倦怠、易怒)在定义上就不特异,2014 年的修订版专门强调它”症状非特异、与个体健康状态和运动刺激的存在都有关”(PMID 24919391)。而完全不可感知的低频正弦运动确实能调制皮肤交感神经活动、降低皮肤血流(PMID 32965160)——生理端有信号,主观端却是一团模糊词。拿模糊词做结局,交互项只会更模糊。

第七,判决。

要做这件事,五条硬要求:① 样本量按交互算,不按主效应算;② 敏感度留在连续尺度,预注册切点只进敏感性分析;③ 同时报乘法与加法两个尺度(含 RERI),并写明哪个是主判据(PMID 38562994);④ 分层工具优先用实测阈值而非回忆问卷;⑤ 无重复验证前,结论只能写”探索性”。第五层还留了个更狠的问题:如果阈值真是稳定特质,它该被当作协变量校正掉,而不是当作修饰因子去切人群——这两条路对同一份数据会给出不同结论,而选哪条本身没有数据能替你决定。

易学落点。

“煞同而灾异”这个观察,古人是看对的。但古人从”不一样”直接推到”我知道你为什么不一样”,中间少了一步:先证明你分得开。命卦、八字、体质,本质上都是在给个体差异起名字;起名字不是测量。现代方法学最不客气的一句话是:想分层,得先有一把解释得动方差的尺,再有一份算得动交互的样本。两样都没有的时候,”某人命里怕这个煞”不是结论,是一个还没被检验的假设——而且是一个功效上很难检验的假设。

这不是把古法否掉,是把它从一个可以随口说的判断,还原成一个要花很大代价才能说的判断。

证据与局限

已验证:交互项检验的功效代价及不同算法间的分歧(PMID 10206617);MSSQ 的预测有效性 r≈0.45 及其与偏头痛等非运动性恶心的相关(PMID 10052582);非差异性错分对乘法/加法交互的相反偏倚方向(PMID 10613335);前庭感知阈值的个体内高一致性(PMID 38658516);两个个体参数可将晕动反应拟合改善 1.7 倍(PMID 40589464);不可感知低频运动对皮肤交感神经活动的调制(PMID 32965160)。

部分验证:把 MSSQ 分数当作稳定的”体质”代理——信度支持,效度只部分支持;sopite 症状的非特异性(PMID 24919391)意味着主观结局的测量误差方向不明,结论只能弱化。

待验证:4-16 Hz 建筑振动这条暴露上,”敏感度×暴露”的交互斜率能否检出、需要多大样本;实测前庭阈值与问卷分数在同一人群中的一致性;把阈值当协变量校正与当修饰因子分层,两种建模对结论的实际翻转幅度。这三项都还没有人做,本篇的判决是设计建议,不是实测结果。

相关阅读

证据等级: B
参考文献: PMID 10206617, PMID 10052582, PMID 10613335, PMID 38658516, PMID 40589464, PMID 24919391, PMID 32965160, PMID 38562994


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。

Same Cause, Different Effects: Including ‘Physiological Differences’ as an Interaction Term

Evidence level: B

In the same room, at the same location, one person feels ill while another remains unaffected—this observation may be true, but the degree to which “who is sensitive” can be measured determines whether it can stand as a conclusion.

Main Text

In Feng Shui and Yijing Studies, there is an almost irrefutable saying: Same Sha, Different Calamity. In the same building, same orientation, same bedroom, one sleeps soundly while another tosses and turns throughout the night. Traditional explanations attribute this to individualized stratification—Life Hexagrams, Bazi (Four Pillars), and constitution. Modern epidemiology names this phenomenon effect modification: the effect of exposure varies according to a third variable. The terminology aligns, but “aligning” and “measuring it out” are two distinct problems. This piece addresses only the latter.

First, an interaction term is a product, not an amplified main effect.

The model is written as $E[Y]=\beta_0+\beta_1D+\beta_2S+\beta_3(D\times S)$. $\beta_3$ tests the difference between two slopes, while each slope already carries its own error. This means the standard error of $\beta_3$ is almost always greater than that of $\beta_1$ or $\beta_2$, even if true effect modification exists. The cost in sample size is rigid: detecting an interaction typically requires several times the sample size needed to detect a main effect, and different algorithms can yield vastly different results in the same scenario (PMID 10206617). This is not conservatism; it is arithmetic.

Second, the ruler used for stratification must first be questioned for its accuracy.

The MSSQ (Motion Sickness Susceptibility Questionnaire) is the most common stratification tool in this line. Authors claim its test-retest reliability can be assumed to exceed 0.8, but its predictive validity for laboratory motion sickness tolerance is only $r\approx0.45$ (PMID 10052582)—accounting for roughly 20% of the variance. Using a ruler that explains only 20% of the variance to group subjects results in misclassification within groups that exceeds the differences between groups.

Third, the direction of misclassification applied to the interaction term differs from that applied to the main effect.

Non-differential misclassification typically pulls the estimate of a multiplicative interaction toward zero (appearing “no modification”), but for an additive interaction, it may push the estimate away from zero (PMID 10613335). The same dataset, two scales, two opposing bias directions. This is why the statement “we did not find a sensitivity modification effect” carries very low information—it may simply mean the ruler is dull.

Fourth, binary versus continuous scales alter the conclusion itself.

Cutting the MSSQ into “susceptible/non-susceptible” is the most convenient yet most damaging approach: arbitrary cut-points, heterogeneity within groups, and naturally sparse samples in high-score regions. The claim of “two different slopes” is merely a re-packaging of “group mean differences.” A more stable approach is to keep sensitivity on a continuous scale to directly estimate the $D\times S$ slope—but this circles back to the power issue raised in the first layer. Neither path is easy; whichever is chosen must be clearly stated.

Fifth, the questionnaire may not measure the vestibular system.

The correlation between MSSQ and non-exertional nausea/vomiting over the past 12 months is $r=0.3$, with migraine being the largest contributor (PMID 10052582). This suggests it measures more like “excitability of the final common vomiting pathway,” something far broader than the vestibular system. Another path offers a narrower answer: vestibular perception thresholds show high consistency within individuals and significant differences between individuals; authors directly refer to it as an innate trait suitable as a biological marker (PMID 38658516). In driving simulation data, two individual parameters can improve the fit of motion sickness responses by 1.7 times (PMID 40589464). Individual differences describable by two numbers need not be proxied by a single recall questionnaire.

Sixth, the symptom side is not straightforward either.

Sopite syndrome (sleepiness, fatigue, irritability) is inherently non-specific by definition; the 2014 revision specifically emphasizes its “non-specific symptoms, related to individual health status and the presence of motion stimulation” (PMID 24919391). Meanwhile, imperceptible low-frequency sinusoidal motion can indeed modulate cutaneous sympathetic nerve activity and reduce skin blood flow (PMID 32965160)—there is a signal on the physiological end, but the subjective end remains a blur of vague terms. Using vague terms as an outcome only makes the interaction term more ambiguous.

Seventh, the verdict.

To conduct this work, five hard requirements apply: ① Sample size calculated for interaction, not main effects; ② Keep sensitivity on a continuous scale, with pre-registered cut-points entering only sensitivity analyses; ③ Report both multiplicative and additive scales (including RERI), and explicitly state which is the primary criterion (PMID 38562994); ④ Prioritize measured thresholds over recall questionnaires for stratification tools; ⑤ Before repeated validation, conclusions can only be written as “exploratory.” Layer five also leaves a harsher question: if the threshold is truly a stable trait, should it be treated as a covariate to adjust for, or as a modifier factor to stratify the population? These two paths yield different conclusions on the same data, and no data exists to decide which path to take.

Yijing Studies Landing Point.

The observation that “Sha is the same but calamity differs” is correct in the eyes of ancients. However, ancients jumped directly from “different” to “I know why you are different,” skipping a crucial step: first prove you can distinguish them. Life Hexagrams, Bazi, and constitution are essentially just naming individual differences; naming is not measuring. The most unyielding statement from modern methodology is: to stratify, one must first have a ruler capable of explaining variance and a sample size capable of calculating interactions. When neither exists, “someone is afraid of this Sha in their fate” is not a conclusion but an untested hypothesis—one that is difficult to test regarding statistical power.

This does not negate ancient methods but restores them from casual judgments to conclusions requiring significant cost.

Evidence and Limitations

Verified: The power cost of interaction testing and divergence between different algorithms (PMID 10206617); MSSQ predictive validity $r\approx0.45$ and its correlation with non-exertional nausea/migraine (PMID 10052582); opposite bias directions of non-differential misclassification on multiplicative/additive interactions (PMID 10613335); high within-individual consistency of vestibular perception thresholds (PMID 38658516); two individual parameters improving motion sickness response fit by 1.7 times (PMID 40589464); modulation of cutaneous sympathetic nerve activity by imperceptible low-frequency motion (PMID 32965160).

Partially Verified: Treating MSSQ scores as a stable proxy for “constitution”—reliability supported, validity only partially supported; the non-specificity of sopite symptoms (PMID 24919391) implies unclear measurement error direction for subjective outcomes, weakening conclusions.

To Be Verified: Whether the “sensitivity × exposure” interaction slope can be detected on the 4–16 Hz building vibration exposure and what sample size is required; consistency between measured vestibular thresholds and questionnaire scores in the same population; the actual magnitude of conclusion reversal when treating thresholds as covariates versus stratifying as modifier factors. None of these have been done yet; the verdict in this piece is a design recommendation, not an empirical result.

Related Reading

Evidence Level: B
References: PMID 10206617, PMID 10052582, PMID 10613335, PMID 38658516, PMID 40589464, PMID 24919391, PMID 32965160, PMID 38562994


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 *