κ 该取哪个标准差:三条口径里有两条会把年龄趋势整个消掉

Written by

in

Language

Evidence level: C

把「老年段的 SD 比年轻段大多少」这个比值换成「用样本自身的 SD 算」,ψ 里的 λ 就没了——不是估得不准,是那个量根本不在式子里了。

正文

前一篇把 z 化之后的残余异方差写成 ψ = λ/κ²,其中 λ 是两段测量误差方差之比、κ 是两段标准差的比值。当时留了个尾巴:κ 到底该用参考方程的 SD,还是样本自己算出来的 SD?直觉上这是同一个东西的三种估法,选个最准的就行。把三条口径的闭式写出来,会发现直觉错了——它们算的根本不是同一个量。

设年轻段信度 ρ_O、老年段信度 ρ_Y,两段观测方差为 VT,则个体间方差 VB = ρ·VT、误差方差 VE = (1−ρ)·VT,所以 λ = (1−ρ_Y)/(1−ρ_O) 而观测 SD 之比 κ_obs = √[λ(1−ρ_Y)/(1−ρ_O)]。第二个式子看着绕,其实只说一件事:λ 被 ρ 绑住了,不是自由参数。于是:

ψ_obs = λ/κ_obs² = (1−ρ_O)/(1−ρ_Y)

λ 消掉了。再看只含个体间分量的口径:κ_VB² = λ·ρ_Y/ρ_O,代进去

ψ_VB = λ/κ_VB² = ρ_O/ρ_Y

λ 也消掉了。三条口径里只有参考方程那一路(ψ_ref = λ/κ_ref²)还留着 λ,因为 κ_ref 是外部方程给的,不随这批数据的 λ 变。

这就不是精度问题了。选 κ_obs,等于宣布「年龄趋势我不管,我只看两段测量误差差多少」;选 κ_VB,等于宣布「我只看两段真个体差差多少」;选 κ_ref,才是在回答「z 化之后误差还剩多少不均」。三句话说的是三件事,而原文设计的判据把它们当成一件事的三个版本在比。

第二个后果更硬:ψ_obs 与 ψ_VB 必然反号。ψ_obs < 1 ⇔ ρ_Y > ρ_O,而 ψ_VB < 1 ⇔ ρ_Y < ρ_O,两者除非 ρ 相等,否则永远一个说「z 化改善精度」、一个说「z 化恶化精度」。取 ρ_O = 0.80、ρ_Y = 0.85、λ = 1.5 一组完全正常的值:ψ_obs = 1.333(恶化 33%)、ψ_VB = 0.941(改善 6%),结论间距 1.42 倍。所以「κ 估偏 0.3 会让结论从该 z 化翻成不该 z 化」这个担心说轻了——不用估偏,只要换口径,结论就翻。

而这个换算极其敏感。κ_VB/κ_obs = √[ρ_Y(1−ρ_O)/(ρ_O(1−ρ_Y))],在 ρ_O = 0.80 处,ρ 差 0.05(0.80→0.85)就给 κ 差 19.0%,ρ 差 0.10 给 50.0%。反过来说,原设计那条「κ_VB 与 κ_obs 差 > 0.20 ⇒ 年龄趋势主要来自测量误差」的阈值,在 ρ 尺度上只对应约 0.052 的差——小到在常规样本量下几乎估不出来。方向也不对:差大于 0.20 说明两段信度确实不同,而信度不同恰恰是两口径已经反号的条件,此时该回答的是「我们要哪一个 ψ」,不是「以哪个 κ 为准」。

还有一层尺度错位。判据挂在 κ 上,判决挂在 ψ 上,而 ψ 是 κ 的平方:κ 差 20% 对应 ψ 差 44%,κ 差 30% 对应 69%。用 κ 做门槛是在用平方根的尺子审一个平方尺度的量,任何阈值都会被系统性低估近一倍。

还有一层容易被忽略:ψ_ref 的方向也不是自由的。当 κ_ref 恰好等于 κ_obs 时,ψ_ref 与 ψ_obs 逐位相等(λ/κ_obs² 化简即得),这说明参考方程那一路并不提供独立信息,它只是把 λ 从观测口径里重新装了回来。所以「参考方程偏陡会让 ψ 被低估」这句话的准确表述是:κ_ref 偏离 κ_obs 的方向,就是 ψ_ref 偏离 ψ_obs 的反方向,且偏离量按平方放大。取 ρ_O = 0.80、ρ_Y = 0.85、λ = 1.5,若参考方程把年龄趋势拟合得偏陡 27%(κ_ref = 1.35 而 κ_obs = 1.06),ψ_ref 从 1.333 掉到 0.823,符号直接翻;反之拟合偏平(κ_ref = 0.90),ψ_ref 升到 1.852。同一批数据,一个参考方程的选择就能给出「改善 18%」和「恶化 85%」两个相反答案——这是本文最该被记住的一条。

那么实践中该怎么做?至少可以固定一条报告格式:ψ 必须成对出现,且必须标明分子分母各是哪种方差。如果论文只写「z 化后残余异方差为 1.33」,读者无法判断这是在说「两段测量误差之比」还是「两段真个体差之比」,而这两者的数值在多数场景下必然分居 1 的两侧。其次,凡是引用外部参考方程的,κ_ref 与 κ_obs 的差要连同区间一起报,不要只报「是否超过 0.15」——因为这条线在 n = 120 下的误报率已在 24% 量级,一个「超标」很可能只是抽样噪声。

最后是样本量。做一遍蒙特卡洛(真值全等、即 ρ 与 λ 都无年龄差,这是判据集的健康检查):判据一(|κ_ref−κ_obs|/κ_obs > 0.15)在 n = 120/段 时就误报 24.4%,n = 60 误报 41.4%;判据三(三值两两差 < 0.10,本意是「κ 稳健可直接用」)在 n = 120 只命中 7.2%,要到 n ≈ 1200 才 96%。于是低样本量下 37%~50% 的情况三条判据全都不触发,高样本量下判据一早就误报完了、判据三才轮得上——判据集的可行域基本是空的。这与前一篇 #1281 的「两条判据只有一个自由度」是同一类病,但病灶不同:那里是判据之间没有交点,这里是判据的样本量需求与触发顺序互相冲突。

务实一点的修法有三条。其一,别再用 κ 做判据的载体,直接把 ψ 的两条口径并列报出来((1−ρ_O)/(1−ρ_Y) 与 ρ_O/ρ_Y),读者一眼能看到它们反号,也知道各自在说什么。其二,若确实要用参考方程,就必须单独报 κ_ref 与 κ_obs 的差及其区间,而不是报「是否超标」——因为 0.15 这条线在 n = 120 下的误报率就在 24% 量级。其三,把「年龄趋势归因」这件事交回给方差分解(前一篇的三层模型),别指望一个 SD 比值同时承担归因和校正两个任务。

证据与局限

已验证: 三条 κ 口径到 ψ 的闭式推导可代数核验,λ 在观测与分量两路口径中被完全消去;ψ_obs 与 ψ_VB 反号是可证的(等价于 ρ_Y 与 ρ_O 的大小关系),与 λ 取值无关。κ 差到 ψ 差的平方放大关系是恒等式。

部分验证: 判据在 n = 60~2400 各档的误报率与命中率来自纯模拟(对数正态近似,k = 3 次重复,Fisher z 近似估 ρ 的抽样误差),未用真实队列数据复核;实际数据的长尾与缺失模式可能使精度更差。

待验证: 真实肺功能队列中 ρ_O 与 ρ_Y 的实际差有多大(若两段信度几乎相等,整场口径之争可判定为不值得处理);以及当参考方程本身来自 LMS 方法的偏态拟合时,κ_ref 的偏差方向是否系统性偏大。

相关阅读

证据等级: C
参考文献: PMID 29707825, PMID 30208858, PMID 24338600, PMID 18006882, PMID 22743675, PMID 38741358

L1 网络状态:E-004 | 金 | 主 | 肺 | 《素问·阴阳应象大论》 | 待补 | L1 | open


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。

Which standard deviation to choose: two of the three diameters will completely eliminate the age trend

Evidence level: C

Replace the ratio “how much larger is the SD for older ages compared to younger ages” with “calculated from the sample’s own SD,” and λ in ψ disappears—not because it is estimated inaccurately, but because that quantity is not present in the formula at all.

Body

The previous article expressed residual heteroscedasticity after z-transformation as ψ = λ/κ², where λ is the ratio of two measurement error variances and κ is the ratio of two standard deviations. A lingering question remained: should κ be derived from the reference equation or calculated directly from the sample? Intuitively, these represent three different estimations of the same underlying quantity; selecting the most accurate one should suffice. However, writing out the closed-form expressions for all three definitions reveals that this intuition is incorrect—they do not estimate the same quantity.

Let ρ_O denote reliability for the younger segment and ρ_Y for the older segment, with total variance VT. Then between-person variance VB = ρ·VT and error variance VE = (1−ρ)·VT, yielding λ = (1−ρ_Y)/(1−ρ_O) and observed SD ratio κ_obs = √[λ(1−ρ_Y)/(1−ρ_O)]. Although the second expression appears complex, it conveys a single point: λ is constrained by ρ and is not a free parameter. Thus:

ψ_obs = λ/κ_obs² = (1−ρ_O)/(1−ρ_Y)

λ cancels out. Now consider the definition based solely on between-person components: κ_VB² = λ·ρ_Y/ρ_O, substituted into

ψ_VB = λ/κ_VB² = ρ_O/ρ_Y

λ again cancels out. Among the three definitions, only the reference equation path (ψ_ref = λ/κ_ref²) retains λ, because κ_ref is supplied externally and does not vary with the λ of this dataset.

This is not merely a matter of precision. Choosing κ_obs declares “I disregard the age trend and focus solely on differences in measurement error between the two segments”; choosing κ_VB declares “I focus exclusively on true differences between individuals”; choosing κ_ref actually answers “how much heteroscedasticity remains after z-transformation.” These three statements address three distinct issues, yet the original design treats them as three versions of the same criterion.

A second consequence is more severe: ψ_obs and ψ_VB must have opposite signs. ψ_obs < 1 ⇔ ρ_Y > ρ_O, while ψ_VB < 1 ⇔ ρ_Y < ρ_O. Unless ρ is identical, one will always claim “z-transformation improves precision” while the other claims “z-transformation worsens precision.” Consider ρ_O = 0.80, ρ_Y = 0.85, λ = 1.5—a completely normal set of values: ψ_obs = 1.333 (worsening by 33%), ψ_VB = 0.941 (improvement of 6%). The conclusions differ by a factor of 1.42. Thus, the concern that “an estimation error of 0.3 in κ would flip the conclusion from ‘should z-transform’ to ‘should not z-transform’” is understated—no estimation error is needed; simply changing the definition flips the conclusion.

This conversion is extremely sensitive. κ_VB/κ_obs = √[ρ_Y(1−ρ_O)/(ρ_O(1−ρ_Y))]; at ρ_O = 0.80, a difference of 0.05 (0.80→0.85) yields a 19.0% difference in κ, while a difference of 0.10 yields 50.0%. Conversely, the original criterion stating “κ_VB differs from κ_obs by > 0.20 ⇒ age trend primarily stems from measurement error” corresponds to only about 0.052 difference on the ρ scale—too small to estimate reliably in conventional sample sizes. Moreover, the direction is wrong: a difference greater than 0.20 indicates that reliabilities indeed differ, and this difference is precisely the condition under which the two definitions yield opposite signs. In such cases, the question to answer is “which ψ do we want,” not “which κ should we use.”

A further issue involves scale mismatch. The criterion depends on κ, the decision rests on ψ, yet ψ is the square of κ: a 20% difference in κ corresponds to a 44% difference in ψ, and a 30% difference in κ corresponds to a 69% difference in ψ. Using κ as a threshold applies a square-root scale to evaluate a squared-scale quantity, causing any threshold to be systematically underestimated by nearly a factor of two.

Another often-overlooked point: the direction of ψ_ref is not free either. When κ_ref exactly equals κ_obs, ψ_ref and ψ_obs are identical at each digit (simplifying λ/κ_obs² yields this result), indicating that the reference equation path provides no independent information—it merely reintroduces λ into the observed definition. Therefore, the accurate statement is: “If κ_ref deviates from κ_obs in a certain direction, then ψ_ref deviates from ψ_obs in the opposite direction, and the deviation is magnified by squaring.” For ρ_O = 0.80, ρ_Y = 0.85, λ = 1.5, if the reference equation overestimates the age trend by 27% (κ_ref = 1.35 while κ_obs = 1.06), ψ_ref drops from 1.333 to 0.823, flipping the sign; conversely, if it underestimates (κ_ref = 0.90), ψ_ref rises to 1.852. With the same dataset, choosing one reference equation can yield two opposite answers—”improvement of 18%” versus “worsening of 85%”—a fact this article most deserves to be remembered for.

So what should be done in practice? At minimum, establish a reporting format: ψ must appear as pairs, clearly indicating which variance constitutes the numerator and which the denominator. If a paper states only “residual heteroscedasticity after z-transformation is 1.33,” readers cannot determine whether this refers to “the ratio of measurement errors between two segments” or “the ratio of true individual differences between two segments,” and these values will inevitably lie on opposite sides of 1 in most scenarios. Furthermore, whenever external reference equations are cited, the difference between κ_ref and κ_obs—including confidence intervals—must be reported, not merely whether it exceeds 0.15—because this threshold has a false-positive rate of approximately 24% when n = 120; a “violation” may simply reflect sampling noise.

Finally, consider sample size. A Monte Carlo simulation (with true values identical, i.e., no age difference in ρ and λ, serving as a health check for the criteria set) reveals: Criterion 1 (|κ_ref−κ_obs|/κ_obs > 0.15) yields a false-positive rate of 24.4% at n = 120 per segment and 41.4% at n = 60; Criterion 3 (pairwise differences among three values < 0.10, originally intended as “κ is robust and can be used directly”) hits only 7.2% at n = 120, reaching 96% only when n ≈ 1200. Thus, at low sample sizes, 37%–50% of cases fail to trigger any criterion; at high sample sizes, Criterion 1 triggers false positives long before Criterion 3 becomes relevant—the feasible region of the criteria set is essentially empty. This mirrors the “two criteria share only one degree of freedom” problem from the previous article #1281, though the pathology differs: there, the criteria had no intersection; here, the sample size requirements and triggering order of the criteria conflict with each other.

Pragmatic solutions include three steps. First, stop using κ as the carrier of the criterion; instead, report both definitions of ψ side-by-side ((1−ρ_O)/(1−ρ_Y) and ρ_O/ρ_Y), allowing readers to immediately see their opposite signs and understand what each represents. Second, if the reference equation must be used, separately report the difference between κ_ref and κ_obs along with its interval, rather than merely stating “whether it exceeds 0.15″—because the false-positive rate of the 0.15 line is already around 24% when n = 120. Third, return the task of “attributing age trends” to variance decomposition (the three-layer model from the previous article), rather than expecting a single SD ratio to simultaneously handle attribution and correction.

Evidence and Limitations

Verified: The closed-form derivations of the three κ definitions into ψ can be algebraically verified; λ is completely eliminated in both the observed and component-based definitions; the opposite signs of ψ_obs and ψ_VB are provable (equivalent to the magnitude relationship between ρ_Y and ρ_O), independent of λ values. The squared-magnification relationship where κ differences translate into ψ differences is an identity.

Partially Verified: False-positive rates and hit rates for the criteria across sample sizes n = 60 to 2400 stem from pure simulation (using log-normal approximation, k = 3 repetitions, Fisher z approximation to estimate sampling error of ρ), not validated against real cohort data; long-tail distributions and missing patterns in actual data may result in even lower precision.

Pending Verification: The actual magnitude of difference between ρ_O and ρ_Y in real pulmonary function cohorts (if reliabilities for both segments are nearly equal, the entire debate over definitions may be deemed unworthy of addressing); and whether κ_ref exhibits systematic overestimation when the reference equation itself stems from skewed fitting via LMS methods.

Related Reading

Evidence Level: C
References: PMID 29707825, PMID 30208858, PMID 24338600, PMID 18006882, PMID 22743675, PMID 38741358

L1 Network Status: E-004 | Metal | Main | Lung | Su Wen · Yin-Yang Ying Xiang Da Lun | Pending | L1 | open


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 *