三质判定的「过度校准」风险裁决:联合分层会不会把真实的跨族裔体质分布差异一起抹掉?

Written by

in

Language

Evidence level: B

校正不足会留下族裔偏倚,校正过头会把真实的跨族裔体质分布差异当噪声抹掉。判别力与差异保留度之间存在一条倒 U 曲线——峰值点与”关联跌破点”之间的区间,才是安全校正窗。

正文

上一篇文章算清了”联合分层要多少人”。这一篇要问一个方向相反、同样致命的问题:就算样本量够,联合分层校正会不会校正过头,把真实的跨族裔三质分布差异也当成偏差一并抹掉?

分层校正在两个方向上都会出错:校正不足——族裔偏倚残留,某族裔被系统性误判;校正过头——把”真实的跨族裔体质分布差异”当噪声消除,三质标签与族裔的真实关联被人为清零。传统三质分型里,这个”真实差异”是有据可查的:一项 3416 人、三中心的研究发现,主导 prakriti 与出生地显著相关(χ², P<0.01),湿润沿海地区卡法 prakriti 占 43.3%,干燥地区瓦塔占 52.3%(PMID 25336848);同一批样本的全基因组分析也显示三质与 52 个 SNP 显著关联,且能不受祖先成分干扰地把个体分类(Nature srep15786)。如果这些差异是真的,那么”把族裔差异校正掉”就等于把一部分真实体质信号当垃圾扔了。

流行病学早就给这个错误起了名字:过度校正偏倚(overadjustment bias)——控制了处在暴露到结局因果路径上的中介变量(PMID 19525685)。它和混杂偏倚的因果结构完全不同:混杂是”控制了才干净”,过度校正是”控制了反而脏”。一篇 2022 年的方法学入门明确给出三步:先讲清因果问题、再画因果图(DAG)、最后按图选变量,目标是”最大化混杂校正、最小化中介校正”(PMID 35662623)。一篇 2024 年的综述更指出,健康不平等研究严重忽视过度校正——84 篇综述里只有 2% 把它纳入偏倚评估(PMID 38129958)。问题不在”要不要校族裔”,而在”族裔在这个因果图里到底是什么角色”。

关键事实:族裔变量捕捉的往往不是生物学,而是数据质量。2024 年 PNAS 一项结直肠癌研究显示,种族调整能提升黑人群体 AUC(P=0.006),机制却是”修正了关键特征的数据质量差异”——家族史在白人群体能预测风险、在黑人群体不能,因该特征数据质量更差(PMID 39136986)。这意味着同一份校正,可能同时在做两件事:修正真实的测量偏倚(该做),和抹掉真实的体质分布差异(不该做)。而机器学习理论给出了更硬的约束:当两组结局率不同时,”校准一致”(calibration)与”错误率均等”(equalized odds)在数学上不可同时满足(PMID 32821854)。联合分层本质上是把校准做到组间一致——但代价恰恰可能是把真实的组间差异”洗平”。

那”真实的跨族裔体质分布差异”到底存不存在?两派都有据。反方:eGFR 与肺功能的种族校正被证明缺乏生物学基础,引证链追溯回去只是一堆薄弱数据(PMID 38447335),更多是历史包袱而非科学。正方:Borrell 等 2021 年在 NEJM 指出,把白人推导的方程直接套到黑人身上会带来过度诊断与不必要的随访,”忽略长期公认的正常生理功能差异”本身即是一种歧视。折中立场是:先证明差异是生物学的、还是环境/测量造成的,再决定校不校——这正是三质研究该做的事。

把这些拼起来,就得到一条可检验的 倒 U 曲线。横轴是校正程度(0=不校正 → 0.5=族裔单层 → 1=联合分层族裔内标准化);两条纵轴:判别 AUC 与”三质标签-族裔关联保留度”(Cramér’s V 或互信息)。可检验的预测是:判别 AUC 随校正程度先升后降——校正不足时族裔偏倚压着 AUC,校正到位时达峰,继续校正则把真实体质信号当偏差洗掉、AUC 回落;关联保留度随校正程度单调下降,过度校正时趋近于零,说明真实差异被清零。两条曲线的峰值点与”关联保留度跌破阈值点”之间的区间,就是”安全校正窗”——窗内是校正到位,窗外是校正过头。

还要防 Simpson 悖论:聚合层面的结论在分层后可能反转(PMID 22246415、PMID 23964259)。三质场景里,”总体看更公平”可能掩盖”分层看某族裔的真实体质信号被抹掉”,因此关联保留度必须在分层层面测量。

裁决实验设计:多族裔队列加三质金标准(信度须先量化);模拟注入已知效应量的”真实跨族裔三质分布差异”(如让某族裔卡法占比天然高 10 个百分点)——只有在已知真值的模拟数据上,才能分清”被抹掉的是真差异还是假偏差”;四档校正管线(无分层 / 族裔单层 / 联合分层 / 过度分层);双终点测量(判别 AUC 交叉验证值 + 关联保留度);拟合两条曲线定位峰值点与跌破点,输出安全校正窗。分层因素”该不该切”则用 Harris-Boyd / Lahti 准则裁定(PMID 2302771、PMID 11805016)——其核心是”分区要有生理学依据,且收益要大于样本被拆散的代价”,正对应本文的”真实差异 vs 噪声”之分。

工程含义很直接:分层校正不是”越彻底越公平”。 先用 DAG 明确族裔的角色(混杂?中介?还是真实体质分布的载体?),再决定校到什么程度。如果族裔既是”数据质量差异”的代理(该校)又是”真实体质分布差异”的载体(不该校),那么正确的做法不是二选一,而是把两者拆开:用数据质量指标(缺失率、信号质量)替代族裔做校正变量,把族裔的”生物学/分布”部分保留下来。做不到这一步,联合分层就是在用”公平”的名义制造一种新的信息损失。

证据与局限

已验证: 过度校正偏倚的定义与因果结构、及其与混杂偏倚的区别(PMID 19525685);避免过度校正的三步变量选择法(PMID 35662623);健康不平等研究中过度校正被严重忽视的实证(PMID 38129958);prakriti 与出生地显著相关(PMID 25336848);族裔调整可修正数据质量差异从而提升某族裔 AUC(PMID 39136986);校准与均等化几率不可兼得(PMID 32821854);eGFR/肺功能种族校正缺乏生物学基础(PMID 38447335);Simpson 悖论在临床与心理学研究中真实存在(PMID 22246415、PMID 23964259);参考区间分层的客观准则(PMID 2302771、PMID 11805016)。

部分验证: “跨族裔三质分布差异中有一部分是真实的体质分布差异”有三质流行病学(出生地关联)与基因组学(SNP 关联)的间接支持,但”该保留多少、保留到什么粒度”尚无量化框架。

待验证: “校正程度-判别AUC”倒 U 曲线在三质场景中的真实峰值位置;”三质标签-族裔关联保留度”的阈值该定在哪;用数据质量变量替代族裔做校正能否同时保住公平性与真实分布差异;分层层面的关联保留度与总体指标是否会出现 Simpson 反转。

相关阅读

证据等级: B
参考文献: PMID 19525685, PMID 35662623, PMID 38129958, PMID 25336848, PMID 39136986, PMID 32821854, PMID 38447335, PMID 22246415, PMID 23964259, PMID 2302771, PMID 11805016


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。

Risk of Overfitting in Three-Stage Classification: Will Joint Stratification Erase Genuine Cross-Racial Phenotypic Distribution Differences?

Evidence level: B

Insufficient correction leaves ethnic bias, while excessive correction treats genuine cross-ethnic tridosha distribution differences as noise to be removed. There is an inverted U-shaped curve between discrimination power and difference retention—the interval between the peak and the “association drop-off point” constitutes the safe calibration window.

Main Text

The previous article calculated “how many samples are needed for joint stratification.” This one asks a directionally opposite yet equally critical question: Even with sufficient sample size, will joint stratification over-correct, treating genuine cross-ethnic tridosha distribution differences as bias to be eliminated?

Stratification correction can err in two directions: insufficient correction—residual ethnic bias where one ethnicity is systematically misclassified; and excessive correction—treating “genuine cross-ethnic tridosha distribution differences” as noise to be eliminated, artificially zeroing out the association between tridosha labels and ethnicity. In traditional tridosha typology, this “genuine difference” is empirically supported: a study of 3,416 individuals across three centers found that prakriti is significantly associated with birthplace ($\chi^2$, $P<0.01$), with Kapha prakriti accounting for 43.3% in humid coastal regions and Vata for 52.3% in dry regions (PMID 25336848); the same cohort’s whole-genome analysis showed significant associations with 52 SNPs and successfully classified individuals independent of ancestral components (Nature srep15786). If these differences are genuine, then “correcting for ethnicity” equates to discarding a portion of true tridosha signals as garbage.

Epidemiology has already named this error: overadjustment bias—controlling for mediators situated on the causal pathway from exposure to outcome (PMID 19525685). Its causal structure differs fundamentally from confounding bias: confounding is “cleaned by controlling,” whereas overadjustment is “made dirtier by controlling.” A 2022 methodological primer explicitly outlines a three-step process: clarify the causal problem, draw a Directed Acyclic Graph (DAG), and select variables based on the graph, aiming to “maximize confounding control while minimizing mediator control” (PMID 35662623). A 2024 review further notes that health inequality research severely neglects overadjustment—only 2% of 84 reviews included it in bias assessment (PMID 38129958). The issue is not “whether to adjust for ethnicity,” but “what role ethnicity plays in this causal graph.”

Key fact: Ethnicity variables often capture data quality rather than biology. A 2024 PNAS study on colorectal cancer showed that race adjustment improved the AUC for Black populations ($P=0.006$), driven by “correcting data quality disparities in key features”—family history predicts risk in White populations but not Black populations due to poorer data quality for that feature (PMID 39136986). This implies a single adjustment performs two simultaneous actions: correcting genuine measurement bias (appropriate) and erasing genuine tridosha distribution differences (inappropriate). Machine learning theory provides a stricter constraint: when outcome rates differ between groups, “calibration consistency” and “equalized odds” are mathematically incompatible (PMID 32821854). Joint stratification essentially forces calibration consistency across groups—the cost of which may be “flattening” genuine inter-group differences.

Do “genuine cross-ethnic tridosha distribution differences” exist? Both sides have evidence. Counter-argument: race corrections for eGFR and lung function lack a biological basis, tracing back to a chain of weak data (PMID 38447335), representing historical baggage rather than science. Pro-argument: Borrell et al. (2021) in NEJM noted that applying equations derived from Whites to Blacks leads to overdiagnosis and unnecessary follow-up, where “ignoring long-standing recognized physiological functional differences” itself constitutes discrimination. The compromise stance is: first prove whether the difference is biological or environmental/measurement-related, then decide whether to adjust—which is precisely what tridosha research should do.

Synthesizing these points yields a testable inverted U-shaped curve. The x-axis represents correction intensity (0 = no adjustment → 0.5 = single-ethnicity stratification → 1 = joint stratification within ethnicity); the two y-axes are discrimination AUC and “tridosha label-ethnicity association retention” (Cramér’s V or mutual information). Testable predictions are: Discrimination AUC rises then falls with correction intensity—ethnic bias suppresses AUC at insufficient correction, peaks at adequate correction, then declines as true tridosha signals are washed out as bias; Association retention decreases monotonically with correction intensity, approaching zero at excessive correction, indicating genuine differences are nullified. The interval between the peak of the two curves and the “association retention drop-off threshold” constitutes the “safe calibration window”—within is adequate correction, outside is excessive.

Also guard against Simpson’s paradox: aggregate-level conclusions may reverse after stratification (PMID 22246415, PMID 23964259). In the tridosha context, “overall fairness” may mask “stratified erasure of genuine tridosha signals for specific ethnicities”; therefore, association retention must be measured at the stratified level.

Decision experiment design: multi-ethnic cohort with tridosha gold standard (reliability must be quantified first); simulate injection of known-effect-size “genuine cross-ethnic tridosha distribution differences” (e.g., naturally increasing Kapha proportion by 10 percentage points for one ethnicity)—only on simulated data with known truth can one distinguish “erased genuine difference vs. false bias”; four-tier correction pipelines (no stratification / single-ethnicity / joint stratification / excessive stratification); dual-endpoint measurement (cross-validated discrimination AUC + association retention); fit both curves to locate the peak and drop-off point, outputting the safe calibration window. Whether to split stratification factors should be adjudicated by Harris-Boyd / Lahti criteria (PMID 2302771, PMID 11805016)—their core principle is “partitioning must have physiological basis, and benefits must outweigh the cost of sample fragmentation,” directly corresponding to this article’s “genuine difference vs. noise” distinction.

Engineering implication is direct: Stratified correction is not “more thorough = more fair.” First, clarify ethnicity’s role in the DAG (confounder? mediator? or carrier of genuine tridosha distribution?). Then decide the degree of correction. If ethnicity acts as a proxy for “data quality disparity” (adjustable) yet also as a carrier of “genuine tridosha distribution differences” (not adjustable), the correct approach is not a binary choice but to decouple the two: replace ethnicity with data quality metrics (missingness rates, signal quality) as correction variables, preserving the “biological/distribution” portion of ethnicity. Failing this step, joint stratification manufactures new information loss under the guise of “fairness.”

Evidence and Limitations

Verified: Definition and causal structure of overadjustment bias, and its distinction from confounding bias (PMID 19525685); three-step variable selection method to avoid overadjustment (PMID 35662623); empirical evidence of severe neglect of overadjustment in health inequality research (PMID 38129958); significant association between prakriti and birthplace (PMID 25336848); race adjustment correcting data quality disparities to improve AUC for specific populations (PMID 39136986); incompatibility of calibration and equalized odds (PMID 32821854); lack of biological basis for race corrections in eGFR/lung function (PMID 38447335); real existence of Simpson’s paradox in clinical and psychological studies (PMID 22246415, PMID 23964259); objective criteria for stratified reference ranges (PMID 2302771, PMID 11805016).

Partially Verified: There is indirect support from tridosha epidemiology (birthplace association) and genomics (SNP association) that “a portion of cross-ethnic tridosha distribution differences is genuine,” but no quantitative framework exists for “how much to retain” or “at what granularity.”

To be Verified: The true peak position of the “correction intensity-discrimination AUC” inverted U-curve in tridosha contexts; the threshold for “tridosha label-ethnicity association retention”; whether substituting data quality variables for ethnicity preserves both fairness and genuine distribution differences; whether stratified-level association retention diverges from overall metrics via Simpson reversal.

Related Reading

Evidence Level: B
References: PMID 19525685, PMID 35662623, PMID 38129958, PMID 25336848, PMID 39136986, PMID 32821854, PMID 38447335, PMID 22246415, PMID 23964259, PMID 2302771, PMID 11805016


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 *