Evidence level: B
上一篇文章给出了一个漂亮公式:观测效应量 = 真值 × ICC × sin(π/2·κ)。但”漂亮”不等于”正确”——这条曲线的真实形状,需要用蒙特卡洛模拟在三种候选模型之间做裁决。
正文
为什么曲线的形状很重要
本系列上一篇文章(#351)提出:当三质分型的信度只有 kappa≈0.28 时,组间效应量会被衰减到真值的一半以下,并给出了 sin(π/2·κ) 的解析公式。但这里有一个隐藏的假设:衰减的形状是”正弦”的。如果真实曲线是线性的、二次的,或者被体质亚型的非对称错分扭曲了,那么基于 sin 模型算出来的所有校正因子、最小信度阈值、样本量放大系数都会整体偏离。
曲线的形状不是数学洁癖,它直接决定三件事:(1) kappa=0.4 时效应量还剩多少——sin 模型算 0.59,线性模型算 0.4,差了将近一半;(2) 需要把分型信度提高到多少才够——若真实曲线更陡,要求的 kappa 阈值更高;(3) 三质(瓦塔/皮塔/卡法)错分敏感度不对称时,曲线会被往哪个方向扭曲——这决定”哪种体质的研究结论最不可信”。
三条候选曲线背后的数学直觉:sin 模型(PMID 41169915,2025)来自”可靠性悖论”系列:当连续潜变量叠加分类噪声、再被离散成标签时,组间差异的衰减因子恰好是 sin(π/2·κ) 与 ICC 的复合。它在 κ=0(纯随机)处衰减到 0,在 κ=1(完美)处为 1,中间是光滑的 S 形——描述”标签噪声逐渐吃掉信号”的平滑过程。
线性模型是经典流行病学教材的默认直觉:衰减因子 ≈ kappa(ARE=kappa 的渐近结果,PMID 7569489,1995)。错分多少,效应就按比例掉多少。它简单,但在小 kappa 区间比 sin 更悲观(kappa=0.3 时线性给 0.3,sin 给 0.45)。
二次模型则是另一种合理假设:衰减在低信度区加速恶化——错分超过阈值后,不仅稀释信号,还把两组分布真正搅混。
这三条曲线在 kappa=0 和 kappa=1 两端重合,但在中间区间(恰是三质研究的现实区间 0.2-0.6)分道扬镳。这就是为什么必须用数据裁决,而不是凭直觉选一条。
非对称错分:比曲线形状更隐蔽的扭曲源
三质不是二分类,而是三分类——这意味着错分不是单一的概率,而是一个 3×3 错分矩阵,而且这个矩阵几乎一定不对称:皮塔型容易被误判为瓦塔型(瘦、活跃),卡法型容易被误判为混合型(体态中庸),而瓦塔型相对容易被正确识别。三质问卷的实证数据支持这种不对称——三项自评问卷在离散分型上的重测信度”高度变异”,卡法内部一致性尤其差(PMID 29089187,2017)。
非对称错分的后果是:“非差分错分总是把效应拉向零”这条二分类世界的铁律,在多分类世界里失效。Dosemeci 等人在 1990 年就给出反例:对多分类暴露变量,非差分错分可以让中间类别估计偏离零值,甚至把真实的剂量-反应趋势扭成反向(PMID 2403115)。模拟研究进一步量化:差分错分造成的偏倚中约 70% 是正向的(夸大效应),30% 是负向的(PMID 26083507,2015);错分方向完全取决于敏感度/特异度组合与患病率(PMID 871121,1977 的经典结论)。对序数/粗略化混杂变量的校正结果则介于粗估与真值之间,方向不可预设(PMID 24014285,2013)。
翻译成三质研究的语言:如果瓦塔组被误分得多、卡法组被误分得少,那么”瓦塔 vs 卡法”的组间比较不仅被稀释,还可能被整体抬高或压低——而 sin 模型完全没有考虑这一点。这正是 #351 把蒙特卡洛错分注入作为方法学核心的原因:注入必须是”矩阵级”的,而不是”单一错分率”的。
蒙特卡洛裁决设计:三模型对决
把”形状之争”变成可执行的实验,设计分四步:
(1) 真值已知的数据生成。设定真效应量 d 从 0.2 到 1.0 扫描(覆盖小/中/大效应),三质真标签按 30%/40%/30% 分布(接近问卷常模)。
(2) 错分矩阵注入。对称场景:单一错分率 p 从 0 到 0.45 梯度扫描(kappa 近似 1-2p);非对称场景:按三质亚型分别设定错分率,如瓦塔 0.15、皮塔 0.35、卡法 0.30,构造真实的 3×3 矩阵。
(3) 拟合与模型比较。每个梯度点重复拟合组间比较模型(如 ANOVA 效应量),得到”错分率/错分矩阵 → 观测效应量”的经验曲线;再用 sin(π/2·κ)、线性、二次三模型分别拟合,AIC/BIC 比较——sin 胜出则公式可放心用;线性或二次胜出则需重写校正因子;非对称注入下三模型都拟合不佳,说明”单一形状”框架本身有问题。
(4) 敏感性分析。逐体质亚型改变错分矩阵元素(每次只动一个格子),观测曲线整体偏移的方向与幅度——定位”哪个格子的错分最危险”。
这个设计与 κ 系数稳健性的蒙特卡洛传统一脉相承——2022 年针对四种 κ 型系数的大规模模拟显示,系数稳健性随样本量、评分类别数变化,且在边缘频率不对称时表现迥异(PMID 35124830)——用模拟裁决理论假设,在信度方法学里是成熟做法。
与易学体系的呼应
这条”裁决曲线形状”的路线,本质上是易学”取象”的统计学版本:先承认模型只是象,再用数据校验象。三质、三因、中医体质这些分型系统,各自都带着自己的”错分矩阵”——皮塔型火象与中医阴虚质、卡法型水土象与痰湿质,跨体系翻译本身就是一层错分。古人用”望闻问切”四诊合参对冲单诊的不可靠,现代研究则用蒙特卡洛注入对冲单模型的不可靠。二者做的是同一件事:不信任单一观察,用冗余与交叉验证逼近真相。
证据与局限
已验证: 衰减公式 sin(π/2·κ) 有 2025 年方法论文献直接支撑(PMID 41169915);二分类错分 ARE=kappa 的渐近关系(PMID 7569489);多分类非差分错分可产生偏离零值的偏倚(PMID 2403115);差分错分偏倚方向取决于敏感度/特异度与患病率、约 70% 为正向(PMID 26083507,PMID 871121);κ 系数蒙特卡洛稳健性评估有成熟先例(PMID 35124830);三质问卷离散分型信度变异大、卡法内部一致性差(PMID 29089187)。
部分验证: 序数/粗略化混杂错分的衰减方向介于粗估与真值之间(PMID 24014285);多分类错分可用错分矩阵做后验校正(PMID 7625410);多分类混杂错分偏倚不一定向粗值方向(PMID 8433115)。这些结论来自邻近领域(流行病学/诊断学),尚未在三质×疗效场景直接复现。
待验证: 三质场景下”错分率→观测效应量”衰减曲线的实际形状是否严格符合 sin 模型;瓦塔/皮塔/卡法错分敏感度不对称时曲线被扭曲的方向与幅度;三模型 AIC/BIC 对决在真实三质数据上的结果;非对称错分矩阵下”单一形状”框架是否仍然成立。
相关阅读
证据等级: B(衰减公式、错分偏倚方向、κ 稳健性均有 PubMed 直接文献支撑;”三质场景三模型对决”本身无直接研究,属方法学设计与科学类比)
参考文献: PMID 41169915, PMID 7569489, PMID 2403115, PMID 26083507, PMID 871121, PMID 24014285, PMID 7625410, PMID 8433115, PMID 35124830, PMID 29089187
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 is the actual curve between error rate and effect size? — Shape determination of the three-quality research decay model
Evidence level: B
The previous article presented a beautiful formula: Observed Effect Size = True Value × ICC × sin(π/2·κ). But “beautiful” does not equal “correct”—the true shape of this curve requires adjudication via Monte Carlo simulation across three candidate models.
Main Text
Why the Shape of the Curve Matters
The previous article in this series (#351) proposed that when the reliability of three-body typology is only kappa≈0.28, between-group effect sizes are attenuated to below half the true value, providing the analytical formula sin(π/2·κ). However, there is a hidden assumption: the shape of attenuation is “sine-shaped”. If the true curve is linear, quadratic, or distorted by the asymmetry of misclassification across body-type subtypes, all correction factors, minimum reliability thresholds, and sample size inflation coefficients derived from the sin model will deviate systematically.
The shape of the curve is not a matter of mathematical pedantry; it directly determines three things: (1) how much effect size remains when kappa=0.4—the sin model calculates 0.59, while the linear model calculates 0.4, a difference of nearly half; (2) how high the classification reliability must be raised to be sufficient—if the true curve is steeper, a higher kappa threshold is required; (3) when misclassification sensitivity across the three bodies (Vata/Pitta/Kapha) is asymmetric, which direction the curve will be distorted—this determines “which body-type research conclusions are least trustworthy.”
The mathematical intuition behind the three candidate curves: the sin model (PMID 41169915, 2025) originates from the “Reliability Paradox” series: when continuous latent variables are叠加 with classification noise and then discretized into labels, the attenuation factor for between-group differences is precisely the composite of sin(π/2·κ) and ICC. It attenuates to 0 at κ=0 (pure random) and reaches 1 at κ=1 (perfect), with a smooth S-shape in between—describing the smooth process of “label noise gradually eating away the signal.”
The linear model is the default intuition in classic epidemiology textbooks: attenuation factor ≈ kappa (the asymptotic result ARE=kappa, PMID 7569489, 1995). Misclassification reduces the effect proportionally. It is simple but more pessimistic than the sin model in small kappa intervals (linear gives 0.3 at kappa=0.3, while sin gives 0.45).
The quadratic model represents another reasonable assumption: attenuation accelerates in low-reliability regions—once misclassification exceeds a threshold, it not only dilutes the signal but genuinely stirs up the distributions of the two groups.
These three curves coincide at both ends (kappa=0 and kappa=1) but diverge in the middle interval (which coincides with the realistic range for three-body research, 0.2–0.6). This is why data must adjudicate rather than intuition selecting a single curve.
Asymmetric Misclassification: A More Subtle Distortion Source Than Curve Shape
Three-body typology is not binary but trinary—meaning misclassification is not a single probability but a 3×3 confusion matrix, which is almost certainly asymmetric: Pitta types are easily misclassified as Vata types (thin, active), Kapha types as mixed types (average physique), while Vata types are relatively easier to identify correctly. Empirical data from three-body questionnaires support this asymmetry—the retest reliability of self-report questionnaires on discrete typologies shows “high variability,” with Kapha internal consistency being particularly poor (PMID 29089187, 2017).
The consequence of asymmetric misclassification is: “Non-differential misclassification always pulls effects toward zero,” a law in the binary world, fails in the multiclass world. Dosemeci et al. provided counter-examples as early as 1990: for multiclass exposure variables, non-differential misclassification can cause intermediate category estimates to deviate from zero, even twisting the true dose-response trend into a reverse direction (PMID 2403115). Simulation studies further quantify this: approximately 70% of bias from differential misclassification is positive (inflating effects), while 30% is negative (PMID 26083507, 2015); the direction of misclassification depends entirely on the combination of sensitivity/specificity and prevalence (PMID 871121, classic conclusion from 1977). Correction results for ordinal/ruggedized confounders lie between crude estimates and true values, with direction not preset (PMID 24014285, 2013).
Translated into the language of three-body research: if Vata groups are misclassified frequently while Kapha groups are misclassified less so, then the between-group comparison of “Vata vs. Kapha” is not only diluted but may also be overall inflated or deflated—while the sin model considers none of this. This is precisely why #351 injected Monte Carlo misclassification into the methodological core: the injection must be “matrix-level,” not based on a “single misclassification rate.”
Monte Carlo Adjudication Design: Three Models in Duel
Turning the “shape debate” into an executable experiment involves four steps:
(1) Data generation with known truth. Set true effect size d scanning from 0.2 to 1.0 (covering small/medium/large effects); distribute true three-body labels at 30%/40%/30% (close to questionnaire norms).
(2) Injection of confusion matrices. Symmetric scenario: scan single misclassification rate p from 0 to 0.45 (kappa approximates 1-2p); asymmetric scenario: set misclassification rates separately for each body subtype, e.g., Vata 0.15, Pitta 0.35, Kapha 0.30, constructing the true 3×3 matrix.
(3) Fitting and model comparison. Repeat fitting of between-group comparison models (e.g., ANOVA effect sizes) at each gradient point to obtain the empirical curve “misclassification rate/matrix → observed effect size”; then fit using sin(π/2·κ), linear, and quadratic models respectively, comparing via AIC/BIC—if sin wins, the formula can be used with confidence; if linear or quadratic wins, correction factors must be rewritten; under asymmetric injection, if all three models fit poorly, it indicates the “single shape” framework itself is flawed.
(4) Sensitivity analysis. Alter confusion matrix elements for each body subtype individually (moving only one cell at a time), observing the direction and magnitude of overall curve shifts—locating “which cell’s misclassification is most dangerous.”
This design aligns with the Monte Carlo tradition for κ coefficient robustness—2022 large-scale simulations targeting four κ types showed that coefficient robustness varies with sample size and number of rating categories, performing differently under edge frequency asymmetry (PMID 35124830)—using simulation to adjudicate theoretical assumptions is a mature practice in reliability methodology.
Resonance with Yijing Systems
This “adjudicating curve shape” route is essentially the statistical version of Yijing’s “taking images” (取象): first acknowledging models are merely images, then verifying images with data. Typification systems like Three Bodies, Three Factors, and TCM constitutions each carry their own “confusion matrix”—Pitta fire image corresponds to TCM Yin-deficiency constitution, Kapha water-earth image corresponds to phlegm-dampness constitution; cross-system translation itself is a layer of misclassification. Ancient practitioners used the four diagnostic methods (inspection, listening/smelling, inquiry, palpation) to counter single-diagnosis unreliability; modern research uses Monte Carlo injection to counter single-model unreliability. Both accomplish the same thing: not trusting a single observation, using redundancy and cross-validation to approach truth.
Evidence and Limitations
Verified: The attenuation formula sin(π/2·κ) has direct support from 2025 methodological literature (PMID 41169915); the asymptotic relationship ARE=kappa for binary misclassification (PMID 7569489); multiclass non-differential misclassification can produce bias deviating from zero (PMID 2403115); the direction of differential misclassification bias depends on sensitivity/specificity and prevalence, with ~70% being positive (PMID 26083507, PMID 871121); mature precedents exist for Monte Carlo robustness assessment of κ coefficients (PMID 35124830); three-body questionnaire discrete typology reliability shows high variability, with Kapha internal consistency being poor (PMID 29089187).
Partially Verified: The attenuation direction for ordinal/ruggedized confounder misclassification lies between crude estimates and true values (PMID 24014285); multiclass misclassification can be post-hoc corrected using confusion matrices (PMID 7625410); multiclass confounder misclassification bias does not necessarily move toward crude values (PMID 8433115). These conclusions come from adjacent fields (epidemiology/diagnostics) and have not yet been directly replicated in the three-body × efficacy scenario.
To Be Verified: Whether the actual shape of the “misclassification rate → observed effect size” attenuation curve in the three-body scenario strictly conforms to the sin model; the direction and magnitude of curve distortion when Vata/Pitta/Kapha misclassification sensitivity is asymmetric; the results of the three-model AIC/BIC duel on real three-body data; whether the “single shape” framework remains valid under asymmetric confusion matrices.
Related Reading
- Insufficient Reliability of Body Typification Dilutes Efficacy Estimates — Monte Carlo Correction
- When Gold Standard Typification Itself is Unreliable, the叠加 of Effect Size Dilution and AUC Contamination
- Three-Body Data Quality Differences and Fairness Testing
- Research Pool: Biological Frequency Code
Evidence Level: B (Attenuation formula, misclassification bias direction, and κ robustness all have direct PubMed literature support; “three-body scenario three-model duel” itself has no direct research, belonging to methodological design and scientific analogy)
References: PMID 41169915, PMID 7569489, PMID 2403115, PMID 26083507, PMID 871121, PMID 24014285, PMID 7625410, PMID 8433115, PMID 35124830, PMID 29089187
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