换成斜率加区间之后判据翻转率真降了吗:两法头对头其实共用一个旋钮

Written by

in

Language

Evidence level: C

把判据从「显著性检验 + 阈值」改成「斜率 + 置信区间」,看着像换了工具,代数上是同一个式子的两种写法;真正决定结论稳不稳的,只有一个数——判据线离标称线几个标准误。

正文

一、一条从没被验证过的建议

上一轮讨论呼气末容积的年龄异方差,结论是「BP 检验显著、方差比超 1.5」这套两步判据有毛病:功效太低,测不出来时默认滑进「同方差够用」,于是「没测出来」被读成「不存在」。当时给的补救是改成直接报「log(VE) 对年龄的斜率与它的置信区间」。

这建议听上去顺,但它本身没人验证过。换成斜率之后,判据就不会在边界附近翻来翻去了吗?还是只换了地方翻?一个判据形式值不值得换,得看它把「同一真值、两次独立抽样给出相反结论」的概率压下去多少。这个概率叫翻转率,等于 2p(1−p),p 是判 H(同方差够用)的概率。

二、两法是同一个式子的两种写法

先把符号摆平。令 θ = log(VE_老 / VE_青),标称线 θ₀ = log(1.5) = 0.4055,两段各 n 人时 log s² 方差 2/(n−1),相减得 SE(θ̂) = √(8/(n−2))。

  • 两步法:方差比检验 + 点估计过线,两个条件与起来等效边距约 k ≈ 1.0;
  • 斜率区间法:θ̂ + 1.645·SE < θ₀,k = 1.645;
  • 点估计法:θ̂ < θ₀,k = 0。

看清结构了吗——三套判据全落在同一个族里:判 H 当且仅当 θ̂ + k·SE < θ₀。所谓「换成斜率」,无非是把 θ 换个名字(斜率分子分母同除年龄跨度),CI 端点单调对应。换判据形式没改变判决逻辑,只改变 k。所以换法后翻转率若真降了,功劳属于 k 被抬高,不属于「斜率更先进」——因果记错,下次就照着错的原因改。

三、线上翻转率有闭式解,而且不含 n

把真值放在标称线上,θ̂ ~ N(θ₀, SE²),判 H 要求 Z < −k,概率就是 Φ(−k)。翻转率

2·Φ(−k)·(1−Φ(−k))

k = 0 时 0.500;k = 1.0 时 0.267;k = 1.645 时 0.095;k = 1.96 时 0.049;k = 2.326 时 0.020。

注意式子里没有 n。这是最反直觉的一条:加大样本不能让标称线上的翻转率变好。λ 恰好等于 1.5 这事本身就在模棱两可的位置上,样本再多也只是把「模棱两可」测得更精确。加样本压窄模糊带宽度,压不掉线上那一份。

顺带得上界:任何二分判据翻转率最多 0.5,永远降不到零。要低于 10%,边距须超 1.62 个 SE;低于 5%,须超 1.95 个。这条与终点是什么、跟易学有没有关系都无关,是二分决策本身带的(PMID 27209009 整理了这类误读,PMID 16217841、22733454、9598427 量化过二分连续量的代价)。

四、最坏点其实不在标称线上

线上 P(H) = Φ(−k) 只是一个点。把 λ 当变量扫过去,p(λ) = Φ((log λ − θ₀)/SE − k),等于 0.5 的位置是

λ* = 1.5·exp(k·SE)

k = 1.645 时:n = 60,λ = 2.76;n = 120,λ = 2.30;n = 600,λ = 1.81;n = 3000,λ = 1.63。

也就是说最坏点永远落在标称线上方,随样本量增大往线上靠,但路不到 1.50。50% 的翻转率始终存在,只是站在不同位置。所以「λ = 1.5 处翻转率只有 9.5%」有相当的欺骗性——它报的是一个点的读数,而曲线最坏点恒为 50%。要报就报整条曲线,或至少报 λ* 在哪儿。

五、网格实测:三法头对头

我们把三套判据放进 4×4 网格(真 λ ∈ {1.0, 1.2, 1.5, 2.0} × n ∈ {60, 120, 300, 600})各跑两万次,另加厚尾臂。为防「只看翻转率会选错判据」,同时算两端均衡判对率 BA = [P(H│λ=1.0) + P(须分段│λ=2.0)] / 2,0.5 等于抛硬币。

两条值得记住:

第一,线上翻转率与 n 完全无关,k = 1.645 那一列在 n = 60/120/300/600 上分别是 0.097、0.098、0.100、0.092,波动只是抽样误差。

第二,翻转率最好的判据不是分辨力最好的判据。k = 0 的点估计法线上翻转率 0.500(最差),但 BA 在 n = 120 时 0.904;k = 1.645 的区间法线上翻转率 0.098(最好),BA 只有 0.731。两者换的是同一样东西:边距越宽,越不容易翻,也越不容易说对。

六、只报翻转率会被常数列刷满分

这条是本轮最该警惕的方法学硬伤。假设有人提议一张假判据:不管数据,一律答「同方差够用」。它在所有 8 个格子里 p = 0,翻转率 2p(1−p) = 0.000——满分。反过来一律答「须分段」,p = 1,翻转率也是 0.000,同样满分。

任何只用翻转率一个标量评估二分判据的方案,都系统性奖励「不表态」。这不是理论空子:区间法在 n = 60 时判 H 的概率只有 0.048,已接近常数列,它的漂亮翻转率是「几乎恒答须分段」换来的,BA 只有 0.640。要判它合不合格,必须同时报固定检出率下的假阳性或整条 ROC。单看翻转率,等于让考官自己出题。

七、真实代价:模糊带与样本量

按「两端各 90% 判对」划线,k = 1.645 时 n = 120 只能保证 λ ≤ 0.70 干净判 H、λ ≥ 1.36 干净判分段,中间 [0.70, 1.36] 怎么判都有一头不稳。n = 600 收窄到 [1.07, 1.44],n = 2400 才到 [1.27, 1.47]。

换算成样本量:要让「λ ≤ 1.2 时 90% 判 H」,k = 1.0 需每段 n ≥ 838,k = 1.645 需 ≥ 1378(总 2756),k = 1.96 需 ≥ 1690。对照计划 n = 120,缺口 11.5 倍。

不二分更省:直接报 θ̂ 的 95% CI,宽度用倍数表示是 exp(3.92·SE)。n = 120 时上下界比 2.78 倍——「最多两倍差」这句在 n = 120 时根本排除不掉;n = 300 降到 1.90 倍,n = 600 是 1.57 倍。

厚尾还要再打一折。误差若接近 t(5),log s² 的方差是 (2+κ)/df,κ = 6 时是正态四倍,真 SE 等于名义两倍。线上 P(H) 仍 0.050(对称性保住),但整条曲线被横向拉伸一倍:n = 120 的模糊带只相当于正态下 n = 30。这也解释了实测里区间法常比公式预测的差(PMID 14673940 报过 CI 近边界时覆盖率不足,PMID 29707825 报 ICC 类指标随分布改变性质)。

八、结论:换的不是判据,是边距

回到开头那条追问。答案是:换成斜率加区间,翻转率不会自动下降。若新做法把 k 从约 1.0 抬到 1.645,翻转率确实从 0.267 降到 0.095;但功劳属于边距,不属于斜率。若新做法仍用点估计或仍按 p < 0.05 落判决,翻转率一分不降。

所以正确写法是四件事一起报:边距 k、样本量 n、线上 P(H)、最坏点 λ* 的位置。缺任一个,「这判据稳不稳」就没内容。若目标真是可重复性而非凑结论,最省的是根本不二分——直接报 θ̂ 与 CI,让读者看区间跨不跨 1.5(PMID 35841249 提醒:区间重叠不等于无差异,不重叠也不等于有差异,二分场景里加倍成立)。

证据与局限

已验证:判据族 θ̂ + k·SE < θ₀ 的线上翻转率闭式解 2Φ(−k)(1−Φ(−k)),与 n 无关;最坏点位置 λ* = 1.5·exp(k·SE) 随 n 单调向标称线靠拢而不抵达;只报翻转率会被常数列刷满分(恒答者翻转率 0.000)。这三条都是纯代数,可逐位验算,不依赖任何数据。

部分验证:4×4 网格的蒙特卡洛结果与闭式解吻合(正态臂误差 < 0.005),厚尾臂的「真 SE = 2×名义」由峰度公式给出、抽样核验方向一致。但这套模拟用的是「两段独立、方差比的对数近似正态」这个简化;真实数据里两段常相关、分布常偏,量化结论会偏移。文献侧,二分连续量的功效损失有稳定共识(PMID 16217841、22733454、9598427),但「斜率法优于检验法」这一具体主张并未在文献中被直接检验过——本轮的意义恰恰是指出它不必被检验,因为两者等价。

待验证:① 本项目考虑的终点(呼气末容积、DR2 余量)在两段相关不为零时,θ̂ 的 SE 该用哪个表达式,会让上述所有数字整体缩放;② 若把「二分决策」本身取消、改报连续量,临床与工程侧是否接受这种输出形式,属于接受度问题而非统计问题,本项目无法单方面裁定;③ 厚尾臂只测了 t(5) 一个点,κ 从 0 到 6 之间翻转率如何连续变化没有扫。

相关阅读

L1 网络状态:E-004|金|主|肺|《素问·阴阳应象大论》|现代证据「待补」|L1|open —— 该边仅为 L1 同构、状态 open,现代证据栏原文为「待补」;本文讨论的是判据形式的等价性与翻转率,在对应网络中既无节点也无边,只能借 E-004 作话题锚点,故证据等级最高只能标 C,文中所有结论均为方法学推演而非机制验证。

证据等级: C
参考文献: PMID 27209009, PMID 16217841, PMID 22733454, PMID 9598427, PMID 17066378, PMID 23130635, PMID 14673940, PMID 29707825, PMID 35841249


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。

Does the flip rate of the criterion really decrease after adding slope and interval: two methods head-to-head actually share one knob

Evidence level: C

Switching the criterion from “significance test + threshold” to “slope + confidence interval” appears like changing tools, yet algebraically it is merely two ways of writing the same equation. What truly determines the stability of the conclusion is a single number: how many standard errors the criterion line lies from the nominal line.

Body

I. One Suggestion Never Verified

In the previous discussion of age heteroscedasticity in end-tidal volume, the conclusion that “the BP test is significant and the variance ratio exceeds 1.5” (a two-step criterion) was flawed: its power is too low, so when it fails to detect an effect, the default slips into “homoscedasticity suffices,” causing “failure to detect” to be read as “non-existence.” The remedy proposed was to switch directly to reporting “the slope of log(VE) against age and its confidence interval.”

This suggestion sounds logical, yet it itself has never been verified. Does switching to the slope prevent the criterion from flipping back and forth near the boundary, or does it merely flip somewhere else? Whether a criterion form is worth changing depends on how much it reduces the probability of “the same true value, two independent samples yielding opposite conclusions.” This probability is called the reversal rate, equal to 2p(1−p), where p is the probability of accepting H (homoscedasticity suffices).

II. The Two Methods Are Two Ways of Writing the Same Equation

Let us first align the symbols. Let θ = log(VE_old / VE_young), the nominal line θ₀ = log(1.5) = 0.4055, and when each segment has n people, the log variance s² has 2/(n−1), so subtracting gives SE(θ̂) = √(8/(n−2)).

  • Two-step method: variance ratio test + point estimate crossing the line; the two conditions combined yield an approximate margin k ≈ 1.0;
  • Slope interval method: θ̂ + 1.645·SE < θ₀, k = 1.645;
  • Point estimate method: θ̂ < θ₀, k = 0.

Do you see the structure now—all three criteria fall into the same family: accept H if and only if θ̂ + k·SE < θ₀. What “switching to the slope” does is merely rename θ (dividing numerator and denominator by age span), with CI endpoints mapping monotonically. Changing the criterion form does not alter the decision logic, only k. Thus, if the reversal rate truly drops after switching, the credit belongs to k being raised, not “the slope is more advanced”—the causality is misattributed; next time, correct based on the wrong cause.

III. The Online Reversal Rate Has a Closed-Form Solution, and It Does Not Depend on n

Place the true value on the nominal line; then θ̂ ~ N(θ₀, SE²). Accepting H requires Z < −k, so the probability is Φ(−k). The reversal rate is

2·Φ(−k)·(1−Φ(−k))

When k = 0, the rate is 0.500; when k = 1.0, it is 0.267; when k = 1.645, it is 0.095; when k = 1.96, it is 0.049; when k = 2.326, it is 0.020.

Note that the formula contains no n. This is the most counterintuitive point: increasing the sample size does not improve the reversal rate on the nominal line. The fact that λ is exactly 1.5 itself lies in an ambiguous position; adding more samples merely measures “ambiguity” more precisely. Narrowing the fuzzy band by adding samples cannot eliminate the portion on the line.

Incidentally, the upper bound: for any binary criterion, the reversal rate is at most 0.5 and never drops to zero. To go below 10%, the margin must exceed 1.62 SE; to go below 5%, it must exceed 1.95 SE. This holds regardless of the endpoint or its relationship to Yijing Studies, as it is inherent to binary decision-making (PMID 27209009 organized such misreadings, while PMID 16217841, 22733454, 9598427 quantified the cost of binary discretization of continuous quantities).

IV. The Worst Point Is Not on the Nominal Line

The probability P(H) = Φ(−k) on the line is just a point. Sweeping λ as a variable, p(λ) = Φ((log λ − θ₀)/SE − k). The position where p equals 0.5 is

λ* = 1.5·exp(k·SE)

When k = 1.645: for n = 60, λ = 2.76; for n = 120, λ = 2.30; for n = 600, λ = 1.81; for n = 3000, λ = 1.63.

Thus, the worst point always lies above the nominal line, approaching it as sample size increases, but never reaching 1.50. A 50% reversal rate always exists, just at different positions. Therefore, the claim “the reversal rate is only 9.5% at λ = 1.5” is quite deceptive—it reports a point reading, while the curve’s worst point remains constant at 50%. If reporting, one must report the entire curve, or at least where λ* lies.

V. Grid Simulation: Three Methods Head-to-Head

We placed the three criteria into a 4×4 grid (true λ ∈ {1.0, 1.2, 1.5, 2.0} × n ∈ {60, 120, 300, 600}), running each two million times, plus thick-tailed arms. To prevent “looking only at reversal rate leads to choosing the wrong criterion,” we also computed the two-end balanced accuracy BA = [P(H│λ=1.0) + P(need segmentation│λ=2.0)] / 2, where 0.5 equals flipping a coin.

Two points are worth remembering:

First, the reversal rate on the line is completely independent of n; for the column k = 1.645, the rates at n = 60/120/300/600 are 0.097, 0.098, 0.100, 0.092 respectively, with fluctuations due only to sampling error.

Second, the criterion with the best reversal rate is not the one with the best resolution. The point estimate method (k = 0) has a reversal rate of 0.500 on the line (worst), yet BA is 0.904 when n = 120; the interval method (k = 1.645) has a reversal rate of 0.098 on the line (best), yet BA is only 0.731. Both switch the same thing: wider margins make it less likely to flip, and less likely to be right.

VI. Reporting Only the Reversal Rate Gets Full Marks from Constant Columns

This is the methodological flaw most worth warning against in this round. Suppose someone proposes a fake criterion: regardless of data, always answer “homoscedasticity suffices.” In all 8 grids, p = 0, so the reversal rate 2p(1−p) = 0.000—full marks. Conversely, always answer “need segmentation,” p = 1, reversal rate is also 0.000, similarly full marks.

Any scheme evaluating binary criteria using only the reversal rate scalar systematically rewards “not stating an opinion.” This is not a theoretical loophole: the interval method, when n = 60, has only a 0.048 probability of accepting H, already close to the constant column; its beautiful reversal rate is earned by “almost always answering ‘need segmentation’,” so BA is only 0.640. To judge whether it qualifies, one must simultaneously report false positives under fixed detection rates or the entire ROC curve. Looking only at reversal rate is like letting the examiner write the questions themselves.

VII. Real Costs: Fuzzy Bands and Sample Size

Drawing lines where “both ends have 90% correct judgment,” when k = 1.645, n = 120 can only may support λ ≤ 0.70 cleanly accepts H, λ ≥ 1.36 cleanly accepts segmentation, while the middle [0.70, 1.36] has uncertainty on both sides regardless of judgment. For n = 600, this narrows to [1.07, 1.44]; n = 2400 reaches [1.27, 1.47].

Converted to sample size: to ensure “λ ≤ 1.2 with 90% acceptance of H,” k = 1.0 requires n ≥ 838 per segment, k = 1.645 requires ≥ 1378 (total 2756), k = 1.96 requires ≥ 1690. Compared to the planned n = 120, the gap is 11.5 times larger.

Non-binary reporting saves more: directly reporting θ̂’s 95% CI, with width expressed as a multiple, is exp(3.92·SE). When n = 120, the upper/lower bounds ratio is 2.78 times—”at most twice the difference” cannot be excluded when n = 120; for n = 300 it drops to 1.90 times, and n = 600 reaches 1.57 times.

Thick tails require another discount. If error approaches t(5), the variance of log s² is (2+κ)/df, where κ = 6 makes it four times the normal; the true SE equals twice the nominal. P(H) on the line remains 0.050 (symmetry preserved), but the entire curve is stretched horizontally by a factor of two: the fuzzy band for n = 120 corresponds only to the normal case with n = 30. This also explains why in simulations the interval method often performs worse than formula predictions (PMID 14673940 reported insufficient coverage of CIs near boundaries, PMID 29707825 reported ICC-type metrics changing properties with distribution).

VIII. Conclusion: What Changed Is Not the Criterion, but the Margin

Returning to the initial question. The answer is: switching to slope plus interval does not automatically reduce the reversal rate. If the new approach raises k from about 1.0 to 1.645, the reversal rate indeed drops from 0.267 to 0.095; but the credit belongs to the margin, not the slope. If the new approach still uses point estimates or still decides based on p < 0.05, the reversal rate drops not at all.

Therefore, the correct reporting involves four things together: the margin k, sample size n, P(H) on the line, and the position of the worst point λ*. Missing any one leaves “is this criterion stable” without content. If the goal is reproducibility rather than forcing a conclusion, the most economical approach is to avoid binary decision-making entirely—directly report θ̂ and CI, letting readers see whether the interval spans 1.5 (PMID 35841249 reminds: overlapping intervals do not mean no difference, non-overlapping does not mean there is a difference; doubling holds in binary scenarios).

Evidence and Limitations

Verified: The closed-form solution for the reversal rate on the line, 2Φ(−k)(1−Φ(−k)), for the criterion family θ̂ + k·SE < θ₀, is independent of n; the worst-point position λ* = 1.5·exp(k·SE) monotonically approaches the nominal line as n increases but never reaches it; reporting only the reversal rate gets full marks from constant columns (always-answering yields reversal rate 0.000). These three are pure algebra, verifiable digit-by-digit, without relying on any data.

Partially verified: Monte Carlo results on the 4×4 grid match the closed-form solution (normal arm error < 0.005); the thick-tailed arm’s “true SE = 2× nominal” is given by the kurtosis formula, with sampling verification consistent in direction. However, this simulation uses the simplification that “two independent segments have log variance ratio approximately normal”; real data often has correlated segments and skewed distributions, so quantitative conclusions may shift. On the literature side, there is stable consensus on power loss from binary discretization of continuous quantities (PMID 16217841, 22733454, 9598427), but the specific claim “slope method outperforms test method” has not been directly tested in literature—the significance of this round is precisely to point out that it does not need testing, because the two are equivalent.

To be verified: ① For endpoints considered in this project (end-tidal volume, DR2 residual), when correlation between segments is non-zero, which expression for θ̂’s SE should be used, causing all numbers above to scale; ② If the “binary decision” itself is cancelled and continuous quantities are reported, whether clinical and engineering sides accept this output format is an acceptance issue rather than a statistical one, which this project cannot unilaterally decide; ③ The thick-tailed arm tested only t(5) one point; how the reversal rate changes continuously as κ varies from 0 to 6 has not been swept.

Related Reading

L1 Network Status: E-004 | Metal | Principal | Lung | Su Wen · Yin-Yang Ying Xiang Da Lun | Modern evidence “to be supplemented” | L1 | open — This edge is only an L1 isomorphism, status open; the original text in the modern evidence column reads “to be supplemented”; this article discusses equivalence of criterion forms and reversal rate, so in the corresponding network there are neither nodes nor edges, and we can only borrow E-004 as a topic anchor, hence the highest evidence level can be marked only C. All conclusions in this text are methodological derivations, not mechanism verifications.

Evidence Level: C
References: PMID 27209009, PMID 16217841, PMID 22733454, PMID 9598427, PMID 17066378, PMID 23130635, PMID 14673940, PMID 29707825, PMID 35841249


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 *