判读遍数不齐时,例数口径还稳不稳:不平衡重测设计的覆盖率与宽度代价

Written by

in

Language

Evidence level: B

真实判读流程里每例的重复遍数几乎必然不齐,此前全部结论都建立在「每例遍数相同」的平衡设计上——这篇要把不齐这一层拆开,看例数口径到底还站不站得住。

正文

前面几篇裁决(保守自由度、标定误差侵蚀、窗长依赖)有一个共同的隐形前提:平衡设计——每一例判读的重复遍数完全相同。平衡设计下「每例取均值再算例数间方差」的口径干净利落:各例均值的方差相等,df=n−1 的 t 区间精确成立。但真实判读现场不是这样:漏读、补读、判读者中途离场、仪器抽风,每例的遍数几乎必然从 2 到 10 随机浮动。遍数不齐时,例数口径还稳不稳?

先把问题的数学骨架说清。每例重复 m 次、取均值后,均值的方差是 σ²/m。平衡设计里 m 对所有例相同,各例均值方差一样,例数间的方差差异纯粹来自被试本身;遍数不齐后,均值方差被 m 劈开——遍数多的例均值更稳、遍数少的例均值更飘。这时例数口径的 df=n−1 t 区间把所有例当成同方差处理,是个近似。近似什么时候还行、什么时候崩?这正是覆盖率研究要回答的。

文献对这类问题的回答分两层。第一层:不平衡本身会不会毁掉 t 类口径。对聚类数据的仿真显示,聚类规模不齐会实实在在影响聚类层面 t 检验的覆盖率表现,且样本越小、不齐越狠,影响越明显(PMID 19672840)——这意味着「不齐但继续用朴素口径」并非天然安全,代价随不平衡程度和样本量缩放。第二层:更合理的口径是什么。两条路现成:一是加权口径,按各例均值方差的倒数(正比于遍数)加权,让遍数多的例多出力;二是线性混合模型,把原始逐遍数据直接喂进去,遍数不齐天然被模型吸收,信度估计、标准误和置信区间都能给出(PMID 27681820)。混合模型在缺失/不平衡数据下的无偏性已有系统仿真背书(PMID 32033617)。而重测信度估计本身对样本量的敏感性也已被方法学文献明确警示(PMID 40119099),遍数不齐相当于在「有效信息量」上再砍一刀。

那么判据怎么定?仿真格子:n∈{10,20,30} × 遍数 2–10 的离散均匀分布随机指派 × ρ∈{0.1,0.3},三种口径同场竞技——例数口径、遍数加权口径、混合模型口径——比覆盖率与区间宽度。若例数口径覆盖率仍 ≥0.94 且相对加权口径宽度代价 <10%,则起步规则免改:不平衡场景直接沿用例数口径,工程上省掉「逐例记遍数再加权」的复杂度;若覆盖率 <0.93,则不齐场景必须升级为加权或混合模型口径,且样本量申报单必须加「遍数完整性」字段,让每例几遍从隐性损耗变成显性申报项。

而且不齐的方式本身有讲究,至少能拆出三档形态。第一档是均匀浮动:每例遍数从 2 到 10 完全随机,缺失与被试无关——这是仿真里的最温和设定,例数口径大概率只是精度略降。第二档是与被试相关的浮动:状态波动大的例更容易被补读、状态稳的例一遍过,遍数与真值挂钩,此时「每例取均值」会系统性偏向遍数多的那类被试,误差不再是噪声而是偏差,t 区间怎么算都救不回来。第三档是与结果相关的浮动:判读结果可疑的例被反复补测,遍数本身携带结果信息,这是最恶劣的一档,混合模型的无偏性保证也只在缺失随机的假设下才成立。文献里混合模型的无偏结论正是在「缺失完全随机或随机」的条件下证的(PMID 32033617),落到判读现场就是:遍数完整性不仅要申报,还得申报「为什么漏」。仿真第一轮只打第一档,先把不齐本身的影响从偏差里剥出来;二、三档是后续轮次的活,不能一口吃。

易学判读语境里,这件事比看起来更根本:脉象、得气、辨证标签这些重测对象,其判读遍数本来就受现场条件摆布。若例数口径在不齐场景下失守,那么此前所有「例数口径安全区」的结论都要加一条脚注——只在平衡设计下成立。

证据与局限

已验证: 遍数/聚类规模不齐会损害朴素 t 类口径的覆盖率(PMID 19672840 仿真);混合模型对不平衡重复测量数据给出无偏的信度与区间估计(PMID 32033617);重测信度估计对样本量敏感(PMID 40119099)。

部分验证: 加权口径与混合模型作为不齐场景的替代方案有直接方法学文献(PMID 27681820),但「判读遍数 2–10 随机浮动 + 三口径头对头」这一具体比较属本项目仿真设计,结果待出。

待验证: 覆盖率 ≥0.94 与宽度代价 <10% 的判据本身;遍数不齐与判读者间信度的交互(漏读是否集中在某类例上,造成非随机缺失);遍数完整性字段在样本量申报流程中的落地成本。

相关阅读

证据等级: B
参考文献: PMID 19672840, PMID 27681820, PMID 32033617, PMID 40119099
L1 网络状态:该主题(遍数不齐 × 例数口径稳健性)尚未进入 L1 对应网络;最近邻为 E-060 标签信度边,当前状态为「冲突(未裁决,且新增两处单位错配/精度区间错配)」(PMID 30383398 / 31730403:中医医师间诊断一致性在开放人群中「低」;偶然校正一致性统计常被误用、建议 Gwet AC2)。

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。

Coverage and Width Trade-offs in Unbalanced Repeated Measures Designs When the Number of Judgments is Uneven

Evidence level: B

In real-world adjudication workflows, the number of repetitions per case is almost inevitably unequal. All prior conclusions rested on the assumption of balanced design—where every case undergoes the exact same number of repetitions. This paper disentangles this assumption to examine whether the case-count metric remains valid when repetition counts vary.

Main Body

Previous adjudications (conservative degrees of freedom, calibration error erosion, window-length dependence) shared a hidden premise: balanced design—every case undergoes the exact same number of repetitions. Under balanced design, the metric “mean each case, then variance across cases” is clean and straightforward: variances of means are equal, so the t-interval with df=n−1 holds precisely. But real adjudication scenes are not like this: missed readings, supplementary readings, adjudicators leaving mid-session, instrument glitches—every case’s repetition count floats randomly between 2 and 10. When repetition counts are unequal, does the case-count metric still hold?

Let’s clarify the mathematical skeleton of the problem. Each case repeated m times, mean taken; variance of the mean is σ²/m. In balanced design, m is identical for all cases, so variances of means are equal; variance differences across cases stem purely from subjects. Once repetition counts become unequal, variance of means is split by m—cases with more repetitions have more stable means; cases with fewer repetitions have more volatile means. At this point, the df=n−1 t-interval treats all cases as having equal variance—an approximation. When does this approximation hold, and when does it collapse? This is precisely what coverage studies must answer.

Literature answers this in two layers. First layer: Does imbalance itself destroy the t-class metric? Simulations on clustered data show that uneven cluster sizes genuinely affect coverage of t-tests at the cluster level, with smaller samples and greater imbalance causing more pronounced effects (PMID 19672840)—meaning “unequal but continuing with the naive case-count metric” is not inherently safe; costs scale with imbalance degree and sample size. Second layer: What is a more appropriate metric? Two paths are available: weighted metrics, weighting each case by the inverse of its mean variance (proportional to repetition count), letting cases with more repetitions contribute more; or linear mixed models, feeding raw per-repetition data directly, where unequal repetition counts are naturally absorbed by the model, yielding reliable estimates, standard errors, and confidence intervals (PMID 27681820). Mixed models have systematic simulation backing for unbiasedness under missing/unbalanced data (PMID 32033617). Meanwhile, retest reliability estimation itself is sensitive to sample size, as methodological literature explicitly warns (PMID 40119099)—unequal repetition counts effectively cut further into “effective information quantity.”

So how do we set the criterion? Simulation grid: n∈{10,20,30} × repetition counts 2–10 assigned via discrete uniform random distribution × ρ∈{0.1,0.3}. Three metrics compete under identical conditions—case-count metric, repetition-weighted metric, mixed-model metric—comparing coverage and interval width. If case-count metric coverage remains ≥0.94 and relative width cost vs weighted metric <10%, the starting rule stands unchanged: in unbalanced scenarios, directly adopt the case-count metric; engineering-wise this avoids the complexity of “recording repetitions per case then weighting.” If coverage <0.93, unbalanced scenarios must upgrade to weighted or mixed-model metrics, and sample size declaration forms must add a “repetition completeness” field, making each case’s repetition count an explicit declared item rather than implicit loss.

Moreover, the mode of imbalance itself matters, at least decomposable into three tiers. First tier is uniform fluctuation: each case’s repetition count floats randomly from 2 to 10, missing data unrelated to subjects—this is the mildest setting in simulations; case-count metric likely suffers only slight precision loss. Second tier is subject-correlated fluctuation: cases with high state volatility are more likely to receive supplementary readings, stable cases pass in one round; repetition count correlates with true value; here “mean per case” systematically favors subjects with more repetitions, error becomes bias rather than noise, and t-intervals cannot be salvaged regardless of calculation. Third tier is result-correlated fluctuation: cases with suspicious adjudication results undergo repeated supplementary testing; repetition count itself carries result information—this is the worst tier; mixed-model unbiasedness holds only under the assumption of random missingness. Literature’s conclusion on mixed-model unbiasedness is proven under “missing completely at random or random” conditions (PMID 32033617); applied to adjudication scenes, this means: repetition completeness must not only be declared, but the reason for missingness must also be declared. First simulation round addresses only Tier 1, peeling away imbalance effects from bias; Tiers 2 and 3 are for subsequent rounds, not to be tackled all at once.

In Yijing Studies context, this matter is more fundamental than it appears: pulse readings, Qi reception, syndrome differentiation labels—all retest objects—have their adjudication repetition counts dictated by on-site conditions. If the case-count metric fails under unbalanced scenarios, all prior conclusions about “safe zones of case-count metrics” must add a footnote—valid only under balanced design.

Evidence and Limitations

Verified: Unequal repetition counts/cluster sizes degrade coverage of naive t-class metrics (PMID 19672840 simulation); mixed models provide unbiased reliability and interval estimates for unbalanced repeated measures data (PMID 32033617); retest reliability estimation is sensitive to sample size (PMID 40119099).

Partially Verified: Weighted metrics and mixed models as alternatives for unbalanced scenarios have direct methodological literature (PMID 27681820), but the specific comparison of “repetition counts 2–10 random fluctuation + three metrics head-to-head” belongs to this project’s simulation design; results pending.

To Be Verified: Criteria of coverage ≥0.94 and width cost <10%; interaction between repetition imbalance and adjudicator reliability (whether missed readings concentrate on certain case types, causing non-random missingness); implementation cost of repetition completeness field in sample size declaration workflow.

Related Reading

Evidence Level: B
References: PMID 19672840, PMID 27681820, PMID 32033617, PMID 40119099
L1 Network Status: This topic (unequal repetition counts × robustness of case-count metric) has not yet entered the L1 corresponding network; nearest neighbor is E-060 label reliability edge, current status “conflict (unresolved, with two new unit mismatches/precision interval mismatches)” (PMID 30383398 / 31730403: diagnostic consistency among TCM practitioners in open populations is “low”; accidental correction consistency statistics are often misapplied; Gwet AC2 recommended).

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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *