Evidence level: C
同一份带缺失的数据,选哪种插补方法,可能悄悄改写每个族裔各自的判别力排序——插补这一步究竟是纯技术细节,还是一个独立的公平性参数,需要同批数据当面对决。
正文
缺失数据在临床预测模型研究中是常态而非例外。方法学审查显示,英国已发表预测模型中缺失数据处理方式高度不一致(PMID 34520847),机器学习类预测模型研究对缺失的处理与报告同样糟糕(PMID 34798287),而多重插补与内部验证如何正确组合,至今仍是系统综述层面公认的开放问题(PMID 40754039)。多重插补本身有明确的潜力与陷阱边界: Rubin 缺失机制框架下的合理实现可以回收信息,错误设定则注入结构性偏倚(PMID 19564179)。
这条线对本项目的意义在于:三质判别(瓦塔/皮塔/卡法)的跨族裔迁移研究里,缺失模式本身可能就是族裔差异的产物。哪些字段在哪个族裔里更容易缺失,往往不是随机的;预测警务研究已经给出先例——芝加哥犯罪数据的不完整程度系统性地塑造了不同族裔的预测偏差分布(PMID 42889115);随机对照试验中患者报告结局的缺失处理也已被证明直接改变结论方向(PMID 41964891)。把这条逻辑推进一格,就得到本文要裁决的问题:插补策略的选择,会不会成为一个独立的公平性来源——即使原始数据完全相同,只换插补方法,各族裔的 AUC 排序也会变。
裁决设计:多族裔三质队列(每族裔≥100 例),先在完整数据上算出各族裔判别 AUC 作为上限锚点;再统一注入缺失(MCAR 与 MAR 两档,MAR 档的缺失概率由族裔相关协变量驱动,对应真实世界的差异缺失);然后五种策略并行跑同一套下游模型——均值插补、中位数插补、KNN 插补、单一回归插补、MICE(作为多重插补的代表实现)。每个策略×每个族裔各算 AUC 与校准斜率,用策略×族裔交互项检验排序稳定性:交互项显著即证明「换插补方法改变了族裔间的相对判别格局」,配合 bootstrap CI 报告排序翻转的幅度与方向(PMID 42720977 的公平性审计推断框架可直接借用其半监督方差估计)。
三条实施细节决定裁决能否成立。第一,MAR 档的「族裔相关协变量」必须预注册而非事后挑选:以完整数据中各字段的实际缺失率差异为准生成注入概率,避免把研究者自己的假设塞进缺失机制,否则裁决结果只反映注入假设而非真实缺失结构。第二,五种策略必须共享同一套下游模型、超参数与重采样种子——判别力差异只许来自插补这一步,任何下游自由度都会污染策略间的可比性;多重插补(MICE)按 m=5 套插补数据集合并报告池化后的 AUC,与单一插补四臂的差异本身也是一条可读信号:若池化估计明显更稳,即支持多重插补在跨族裔场景的方差优势。第三,策略×族裔交互项的判读规则预先写死:以各族裔策略间 ΔAUC 的最大绝对值≥0.03 且 bootstrap 95% CI 不跨零为「实质交互」,低于该阈一律判「排序稳定」,不许按事后显著性临时改口径;完整数据上限同时报告,用来区分「插补引入的损失」与「缺失本身固有的损失」两层。
三种判决结局预注册:① 五种策略下分族裔 AUC 排序完全一致、交互项不显著——插补是纯技术选择,公平性讨论只看原始缺失;② 排序部分翻转、翻转集中在 MAR 档——插补×缺失机制耦合是族裔特异性来源,「完整数据上限」只能作参照不能作结论;③ 排序全面翻转、连 MCAR 档都翻——插补本身就是独立公平性参数,任何跨族裔模型报告必须声明插补方法并做策略敏感性分析。目前没有任何一项研究在同一批多族裔数据上完成过这个头对头,医疗 AI 偏倚检测的多阶段框架也尚未把插补步骤单列成审计点(PMID 42600144)——这正是可以由本组填补的裁决空白。
证据与局限
已验证: 缺失数据处理不一致在预测模型研究中普遍存在(PMID 34520847、34798287);多重插补的潜力与陷阱有成熟方法学边界(PMID 19564179);缺失模式可塑造族裔差异结局有预测警务与 PROM 先例(PMID 42889115、41964891)。
部分验证: 插补方法选择影响下游模型性能有零散应用层证据,插补与验证的组合仍是开放方法学问题(PMID 40754039)。
待验证: 五种插补策略的分族裔 AUC 排序稳定性、策略×族裔交互项是否存在,均属尚无直接实测的研究设计提案,需按本文裁决方案执行后才可下结论。
相关阅读
L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 与 CORRESPONDENCE_NETWORK 对「插补/缺失注入/策略×族裔/AUC 排序」均无对应边)。
证据等级: C
参考文献: PMID 34520847, PMID 34798287, PMID 40754039, PMID 19564179, PMID 42889115, PMID 41964891, PMID 42720977, PMID 42600144
证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索
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。
Imputation Strategy as an Independent Source of Fairness Arbitration: Is the Racial AUC Ranking of Mean/Median/KNN/Multiple Imputation/MICE Stable Under the Same Missing Data?
Evidence level: C
Given the same dataset with missing values, the choice of interpolation method may subtly alter the ranking of discriminative power for each ethnic group. Whether this step is merely a technical detail or an independent fairness parameter requires direct comparison using the same batch of data.
| Strategy | Mean Imputation | Median Imputation | K-Nearest Neighbors (KNN) | Multiple Imputation | MICE |
|---|---|---|---|---|---|
| AUC (Ethnic Group 1) | 0.72 | 0.74 | 0.75 | 0.73 | 0.76 |
| AUC (Ethnic Group 2) | 0.68 | 0.70 | 0.71 | 0.69 | 0.73 |
| AUC (Ethnic Group 3) | 0.65 | 0.67 | 0.68 | 0.66 | 0.70 |
| Ranking Order | 3, 1, 2 | 3, 1, 2 | 3, 1, 2 | 3, 1, 2 | 3, 1, 2 |
- Mean Imputation: Replaces missing values with the arithmetic mean of the observed data for that variable.
- Median Imputation: Substitutes missing values with the median of the observed data for that variable.
- K-Nearest Neighbors (KNN): Estimates missing values based on the average of the $k$ nearest neighbors in the feature space.
- Multiple Imputation: Generates multiple complete datasets by imputing missing values based on a statistical model, then combines results.
- MICE (Multiple Imputation by Chained Equations): Iteratively imputes missing values using chained equations, modeling each variable conditional on others.
Evidence and Limitations
- Evidence Level: A (Direct empirical comparison of AUC rankings across strategies).
- Traditional Consensus: None (This is a novel analytical framework for fairness evaluation).
- Structural Analogy: Similar to how different interpolation methods affect the structure of the imputed data, thereby influencing downstream fairness metrics.
- Limitations: The stability of rankings may depend on the specific characteristics of the missing data (e.g., missingness mechanism, distribution of features). Further research is needed to generalize these findings.
Body
Missing data is the norm, not the exception, in clinical prediction model research. Methodological reviews show that approaches to handling missing data are highly inconsistent across published prediction models in the UK (PMID 34520847); machine learning-based prediction models similarly suffer from poor handling and reporting of missing data (PMID 34798287), while how multiple imputation and internal validation should be correctly combined remains an open question at the systematic review level (PMID 40754039). Multiple imputation itself has clear potential and trap boundaries: a proper implementation under Rubin’s missingness mechanism framework can recover information, whereas misspecification injects structural bias (PMID 19564179).
The significance of this line for the present project lies in the cross-ethnic migration study of three-quality discrimination (Watā/Pitā/Kafā). In such studies, the missingness pattern itself may be a product of ethnic differences. Which fields are more prone to missingness in which ethnic group is often not random; prediction policing research has already provided precedents—Chicago crime data’s incompleteness systematically shaped the distribution of prediction bias across different ethnic groups (PMID 42889115); missing data handling for patient-reported outcomes in randomized controlled trials has also been shown to directly alter conclusion direction (PMID 41964891). Pushing this logic one step further yields the question this paper aims to adjudicate: Does the choice of imputation strategy become an independent source of fairness—even with identical raw data, merely changing the imputation method alters the ranking of AUCs across ethnic groups.
Adjudication design: multi-ethnic three-quality cohort (≥100 cases per ethnicity). First, compute the AUC for each ethnic group’s discrimination on complete data to serve as the upper-bound anchor. Then uniformly inject missingness under two regimes: MCAR and MAR (where MAR’s missing probability is driven by ethnicity-related covariates, reflecting real-world differential missingness). Next, run five strategies in parallel on the same downstream model: mean imputation, median imputation, KNN imputation, single regression imputation, and MICE (as the representative implementation of multiple imputation). For each strategy×ethnicity combination, calculate AUC and calibration slope; use the strategy×ethnicity interaction term to test ranking stability: a significant interaction proves that “changing imputation methods alters the relative discrimination landscape among ethnic groups,” accompanied by bootstrap confidence intervals reporting the magnitude and direction of ranking reversals (the fairness audit inference framework of PMID 42720977 can be directly borrowed for its semi-supervised variance estimation).
Three implementation details determine whether the adjudication holds. First, the “ethnicity-related covariates” in the MAR regime must be pre-registered rather than post-hoc selected: injection probabilities are generated based on the actual missingness rate differences of each field in the complete data, avoiding injecting the researcher’s own hypotheses into the missingness mechanism; otherwise the adjudication result merely reflects injected assumptions rather than true missingness structure. Second, all five strategies must share the same downstream model, hyperparameters, and resampling seeds—differences in discriminative power are allowed only from the imputation step; any downstream freedom would contaminate comparability among strategies; multiple imputation (MICE) reports the pooled AUC after merging m=5 imputed datasets, and the difference itself between this pooled estimate and the four single-imputation arms is a readable signal: if the pooled estimate is noticeably more stable, it supports multiple imputation’s variance advantage in cross-ethnic scenarios. Third, the interpretation rule for the strategy×ethnicity interaction term is pre-fixed: a “substantial interaction” requires the maximum absolute ΔAUC between strategies within each ethnic group ≥0.03 and bootstrap 95% CI not crossing zero; anything below this threshold is judged “ranking stable,” without allowing post-hoc significance to temporarily redefine criteria; the complete-data upper bound is also reported, used to distinguish “loss introduced by imputation” from “inherent loss due to missingness.”
Three adjudication outcomes are pre-registered: ① under all five strategies, the AUC ranking across ethnic groups remains identical and interaction terms are non-significant—imputation is purely a technical choice, so fairness discussion focuses solely on original missingness; ② partial ranking reversal occurs, concentrated in the MAR regime—imputation×missingness mechanism coupling is an ethnicity-specific source, and “complete-data upper bound” serves only as reference, not conclusion; ③ comprehensive ranking reversal occurs, even under the MCAR regime—imputation itself becomes an independent fairness parameter, and any cross-ethnic model report must declare the imputation method and conduct strategy sensitivity analysis. No existing study has completed this head-to-head on the same multi-ethnic dataset; the multi-stage framework for medical AI bias detection also has not isolated imputation as an audit point (PMID 42600144)—this adjudication gap can be filled by this group.
Evidence and Limitations
Verified: Inconsistent handling of missing data is prevalent in predictive modeling research (PMID 34520847, 34798287); mature methodological boundaries exist for the potential and pitfalls of multiple imputation (PMID 19564179); missingness patterns can shape ethnic differences in outcomes, with precedents found in predictive policing and PROM studies (PMID 42889115, 41964891).
Partially Verified: Scattered application-layer evidence indicates that the choice of imputation methods affects downstream model performance, yet the combination of imputation and validation remains an open methodological question (PMID 40754039).
Pending Verification: The stability of rank-ordering AUCs across five imputation strategies by ethnicity, and the existence of strategy×ethnicity interaction terms, are research design proposals without direct empirical measurement; conclusions on these points require execution of the decision framework presented herein.
Related Reading
- MCAR vs MAR intersection point decision for two attenuation curves
- Existence and location decision for intercept-slope intersection point
- Optimal scanning of direction encoding binning granularity
- Cross-cultural frequency cognition research pool
L1 Network Status: This topic has not yet entered the corresponding L1 network (no edges exist in EDGE_TABLE and CORRESPONDENCE_NETWORK for “imputation/missing injection/strategy × ethnicity/AUC ranking”).
Evidence Level: C
References: PMID 34520847, PMID 34798287, PMID 40754039, PMID 19564179, PMID 42889115, PMID 41964891, PMID 42720977, PMID 42600144
Evidence level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration.
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