两个读数之间的误差相关到底有多大:扰动臂实验把责任指向定标还是算法

Written by

in

Language

Evidence level: C

同一根呼吸曲线、同一个容积轴,两次读数之间的误差相关 ρ 从未在本领域被测过——它的大小、符号和方向,决定了整套「余量」终点的最小可检测变化是放宽还是收严。

正文

在呼吸功能检测里,「余量」这个终点是呼气末容积(EELV)与肺活量变化(DR2)两个读数的差。上一篇文章(《余量的最小可检测变化,一直按两分量独立在算》)已经证明:一旦两个读数的测量误差相关(ρ≠0),独立假设就会把最小可检测变化高估 1.58 倍到 4.47 倍(ρ=0.6 到 0.95);反过来,若 ρ 为负,名义 5% 的假阳性率会涨到 8.6%–11%。换句话说,整道判据门槛都挂在 ρ 这一个数上。

可问题是:这个 ρ 在本领域从来没有人测过。「同一条 SVC-FOT 曲线、同一个容积轴」这条设计理由,只能说明 ρ 应当为正,给不出量级;而 ρ 的符号和大小恰好决定结论方向——ρ>0 意味着独立假设高估门槛(自罚),ρ<0 意味着低估门槛(自宽)。更麻烦的是,ρ 越大越估不准:ρ=0.9 时估偏 0.05 就能让 MDC95 差 22%,ρ=0.95 时差 41%;样本量 n=30 时 ρ=0.8 的 95% 置信区间半宽约 0.14,映射到 MDC95 上上下相差 2.29 倍。所以「估出 ρ」这件事本身需要一份精度预算,不能只报一个点估计。

本主题给出的方案是:用扰动臂实验把「ρ 归因」拆开。第一步,取 ≥30 名健康成人,隔天各测 5 次 SVC-FOT,同一算法逐次同时检出 DR2 与呼气末容积(均折算为 %TLC),把每次的两个读数配成对,用两次测量之差的 2×2 协方差阵直接估计 s₁、s₂ 与 ρ,并给出 ρ 的 Fisher z 变换 95% 置信区间。配对测量误差的完整表达式必须含 2ρs₁s₂ 这一项——这是经典测量学早已给出的框架(Bruynesteyn 在类风湿关节进展判定中明确提出:分数统计上的相依不等于误差上的相关,两者要分开处理,PMID 15286006);分量层面的 s₁、s₂ 与 ICC 则由标准重测信度框架给出(Weir 的 ICC+SEM 框架,PMID 15705040);每个个体多次观测的一致性评估方法(Bland–Altman 扩展)则提供了「把重复测量之差当新变量」的统计出处(PMID 17613642)。

第二步是关键:设两个扰动臂。(a) 换容积轴定标方法——例如以已知容积注射器标定 vs 以潮气量积分标定;(b) 换 DR2 检出算法——换切线判据或换平滑窗。每个臂内重估 ρ,比较臂间 ρ 的差。这个设计的逻辑是:如果 ρ 主要由「共用采集通道」这一物理事实决定,那么换定标、换算法都不该改变它太多,责任在测量原理;如果某一臂让 ρ 明显下降,那么 ρ 的实际来源是被该环节注入的,责任就在那一环。

判据分四档。第一档:主臂 ρ>0.6 且两个扰动臂的 ρ 都同向为正,(a) 臂使 ρ 下降的幅度大于 (b) 臂的 2 倍 → 责任在定标,须把定标方法登记为误差结构的一部分并优先改造。第二档:(b) 臂使 ρ 下降更多 → 责任在检出算法,优先换算法。第三档:两臂 ρ 变动都 <0.1 且 ρ 的 95% 置信区间宽度 >0.3 → ρ 在当前样本量下不可估,须先扩样(或改用不依赖 ρ 的组水平终点),在此之前任何含 ρ 的 MDC95 都不得写入结论。第四档:主臂 ρ<0.2 或为负 → 与「同曲线同轴」的设计理由相冲,须逐条排查两个读数是否实际来自不同处理链(如不同平滑窗、不同采样率),排查不出则视为未测。

这个设计的价值在于:它把「误差相关」从一个隐假设变成可测量的量,再把「ρ 偏大」的责任定位到具体环节。肺功能检测文献已经给出 ρ 可能量级的旁证:FOT 的会话内重复性在分量层面有明确数据(PMID 33914403),日间变异 5%–15%(PMID 23598954);但分量层面的重复性好,恰恰不能推出「两分量误差互不相关」——两个高度可重复的读数,可以因为共用同一采集通道而高度相关。这也是为什么必须有扰动臂,而不是只看 ICC。

需要特别说明:本项目上一篇文章已经指出,含 ρ 的 MDC95 公式一旦启用,必须同时检查差值信度 r_DD 是否为正——ρ 升高会让差值信度塌得比门槛降得快(分量信度 0.80 时,ρ=0.6 已让 r_DD=−0.25)。因此本文测出的 ρ 不是终点,而是下一篇文章(差值信度会不会先破产)的输入。

L1 网络状态:本主题对应边 E-004「金 → 肺」,《素问·阴阳应象大论》,现代证据「待补」,层级 L1(同构),状态 open——L1 网络中该边尚未进入实证阶段,现代证据列标注为「待补」。

证据与局限

已验证: 配对测量误差须用含 2ρs₁s₂ 的完整表达式(PMID 15286006);分量信度可用 ICC+SEM 标准框架估计(PMID 15705040);多次观测下的一致性评估方法成熟(PMID 17613642);FOT 会话内重复性有数据(PMID 33914403);日间变异 5%–15%(PMID 23598954)。
部分验证: 含 ρ 的 MDC95 修正公式、ρ 的 Fisher z 置信区间方法学均有文献支持;但 ρ 本身在本领域(呼吸余量终点)从未实测,其量级、符号均未知。
待验证: 主臂 ρ 的实测值与符号;两个扰动臂(换定标 / 换算法)各自对 ρ 的影响幅度;「ρ 主要来自共用通道」与「ρ 由某环节注入」两种假说的区分。下一步检验差值信度 r_DD 在实测 ρ 下是否仍为正(对应关联课题 #1283)。

相关阅读

证据等级: C
参考文献: PMID 15286006, PMID 15705040, PMID 17613642, PMID 33914403, PMID 23598954


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。

How Large Is the Error Correlation Between Two Readings: Does the Perturbation Arm Experiment Point to Calibration or Algorithm?

Evidence level: C

The error correlation $\rho$ between two readings on the same respiratory curve and volume axis has never been measured in this field—its magnitude, sign, and direction determine whether the minimum detectable change (MDC) endpoint of the entire “margin” is relaxed or tightened.

Body

In respiratory function testing, the endpoint “margin” is defined as the difference between end-expiratory lung volume (EELV) and the change in vital capacity (DR2). A previous article (“The Minimum Detectable Change of the Margin Has Always Been Calculated Assuming Two Components Are Independent”) demonstrated that once measurement errors between two readings are correlated ($\rho \neq 0$), the assumption of independence inflates the MDC by 1.58 to 4.47 times (for $\rho = 0.6$ to $0.95$). Conversely, if $\rho$ is negative, the nominal false-positive rate of 5% rises to 8.6%–11%. In short, the entire decision threshold hinges on this single value, $\rho$.

The problem is: this $\rho$ has never been measured in this field. The design rationale—”the same SVC-FOT curve, the same volume axis”—only implies that $\rho$ should be positive; it provides no magnitude. Yet the sign and size of $\rho$ precisely determine the conclusion’s direction: $\rho > 0$ means the independence assumption overestimates the threshold (self-penalty), while $\rho < 0$ means it underestimates it (self-loosening). More critically, larger $\rho$ leads to greater estimation error: at $\rho = 0.9$, an estimation bias of just 0.05 causes the MDC${95}$ to differ by 22%; at $\rho = 0.95$, the difference reaches 41%. With a sample size $n=30$, the 95% confidence interval half-width for $\rho = 0.8$ is approximately 0.14, which maps to a 2.29-fold difference in MDC$$. Therefore, “estimating $\rho$” itself requires an error budget; reporting a single point estimate is insufficient.

This topic proposes a solution: use perturbation arms to decompose the “attribution of $\rho$.”

Step 1: Recruit $\ge 30$ healthy adults, each tested 5 times on alternate days using SVC-FOT. Using the same algorithm, simultaneously detect DR2 and EELV (both converted to %TLC) on each occasion. Pair the two readings from each test, and directly estimate $s_1$, $s_2$, and $\rho$ using the $2 \times 2$ covariance matrix of the difference between two measurements. Provide the Fisher z-transformed 95% confidence interval for $\rho$. The complete expression for paired measurement error must include the term $2\rho s_1 s_2$—a framework long established in classical metrology (Bruynesteyn explicitly stated in rheumatoid joint progression assessment that statistical dependence of scores does not equal error correlation; the two must be handled separately, PMID 15286006). Component-level $s_1$, $s_2$, and ICC are derived from the standard retest reliability framework (Weir’s ICC+SEM framework, PMID 15705040). The method for assessing consistency across multiple observations per individual (Bland–Altman extension) provides the statistical basis for “treating the difference of repeated measurements as a new variable” (PMID 17613642).

Step 2 is critical: Set up two perturbation arms. (a) Change the volume axis calibration method—for example, calibrate using a known-volume syringe versus calibrating via tidal volume integration. (b) Change the DR2 detection algorithm—for example, switch from tangent-based criteria to smoothing windows. Re-estimate $\rho$ within each arm and compare the difference in $\rho$ between arms. The logic of this design is: if $\rho$ is primarily determined by the physical fact of “shared acquisition channel,” then changing calibration or algorithm should not alter it significantly; responsibility lies with the measurement principle. If one arm causes $\rho$ to drop noticeably, then the actual source of $\rho$ is injected at that stage; responsibility lies there.

The decision criteria are divided into four tiers:

  1. Tier 1: Main arm $\rho > 0.6$ and both perturbation arms yield positive $\rho$, but arm (a) reduces $\rho$ by more than twice the amount arm (b) does → Responsibility lies with calibration. The calibration method must be registered as part of the error structure and prioritized for modification.
  2. Tier 2: Arm (b) reduces $\rho$ more than arm (a) → Responsibility lies with the detection algorithm. Prioritize changing the algorithm.
  3. Tier 3: Both arms show $\rho$ variation $< 0.1$ and the 95% confidence interval width for $\rho$ exceeds 0.3 → $\rho$ cannot be estimated under the current sample size; sampling must be expanded first (or switch to group-level endpoints not dependent on $\rho$). Until then, any MDC$_{95}$ involving $\rho$ must not be included in conclusions.
  4. Tier 4: Main arm $\rho < 0.2$ or negative → This contradicts the design rationale of “same curve, same axis.” Each reading must be checked individually to determine if they actually originate from different processing chains (e.g., different smoothing windows, different sampling rates). If such discrepancies cannot be identified, $\rho$ is considered unmeasured.

The value of this design lies in transforming “error correlation” from an implicit assumption into a measurable quantity, and precisely locating responsibility for “overestimated $\rho$” to specific stages. Respiratory function literature already provides collateral evidence of possible magnitudes for $\rho$: FOT intra-session repeatability has clear component-level data (PMID 33914403), and daily variation ranges from 5%–15% (PMID 23598954). However, good repeatability at the component level does not imply “errors between two components are uncorrelated.” Two highly repeatable readings can be strongly correlated simply because they share the same acquisition channel. This is why perturbation arms are necessary, rather than relying solely on ICC.

Special note: A previous article in this project indicated that once the MDC${95}$ formula involving $\rho$ is activated, the difference reliability $r = -0.25$). Thus, the $\rho$ measured here is not the endpoint but input for the next article (whether difference reliability fails first).}$ must simultaneously be checked for positivity—rising $\rho$ causes $r_{DD}$ to collapse faster than the threshold drops (when component reliability is 0.80, $\rho = 0.6$ already yields $r_{DD

L1 Network Status: This topic corresponds to edge E-004 “Metal → Lung” in Su Wen · Yin-Yang Ying Xiang Da Lun. Modern evidence is “pending,” level L1 (isomorphic), status open—the edge has not yet entered the empirical phase in the L1 network; modern evidence is marked as “pending.”

Evidence and Limitations

Verified: Paired measurement error requires the complete expression including $2\rho s_1 s_2$ (PMID 15286006); component reliability can be estimated using the ICC+SEM standard framework (PMID 15705040); consistency assessment under multiple observations is mature (PMID 17613642); FOT intra-session repeatability has data (PMID 33914403); daily variation is 5%–15% (PMID 23598954).

Partially Verified: The MDC$_{95}$ correction formula involving $\rho$ and the Fisher z confidence interval methodology for $\rho$ are supported by literature; however, $\rho$ itself has never been directly measured in this field (respiratory margin endpoint), so its magnitude and sign remain unknown.

Pending Verification: The measured value and sign of $\rho$ in the main arm; the magnitude of $\rho$ reduction caused by each perturbation arm (changing calibration vs. changing algorithm); distinguishing between the hypothesis that “$\rho$ mainly stems from the shared channel” and that “$\rho$ is injected at a specific stage.” Next, verify whether difference reliability $r_{DD}$ remains positive under the measured $\rho$ (corresponding to related topic #1283).

Related Reading

Evidence Level: C
References: PMID 15286006, PMID 15705040, PMID 17613642, PMID 33914403, PMID 23598954


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 *