暖化趋势强度对「去趋势通道检验功效」的敏感性标定:趋势多强时,去趋势通道就「看不见」干支了

Written by

in

Language

Evidence level: B

上一篇说”去趋势后还能站住脚的干支信号才算真信号”。但去趋势这个动作本身会吃掉检验的力气——趋势越强,先把趋势削掉,再去找剩下那点干支信号,就越可能”什么都找不到”。所以”去趋势后不显著”不能直接读成”没有干支”,得先补一层功效标定:趋势到底多强时,这个通道就瞎了。

正文

上一篇(#426)把检验五运六气的第一道关卡立了起来:面对 1961 年以来那条单向的暖化趋势,只看”含趋势对比”会得到假的符合,必须走”去趋势通道”——把趋势削掉之后,再看干支分组还剩多少效应。但这里立刻冒出一个更尖锐、也更容易被忽略的问题:去趋势通道自己会不会把真信号也一起削掉? 如果答案是”会”,那么”去趋势后不显著”这句话就是有歧义的——它可能是”真的没有干支信号”,也可能是”趋势太强,检验被压垮了”.

检验功效只有在”真情况已知”时才能算。 功效(power = 1 − β)的定义是”当效应真实存在时,检验能把它揪出来的概率”。问题是:真实的地球上,我们并不知道干支到底有没有真效应、有多大——所以从真实数据里根本算不出功效,只能从数据”反推”出一个显不显著。想量化”去趋势通道还剩多少把握”,唯一的路是合成序列标定:造一批”真情况已知”的人工序列,人为塞进一个设定好的干支效应,再跑完全一样的去趋势流程,看能检出多少。这正是流行病学里做中断时间序列(ITS)研究时的标准做法——研究者用模拟数据估计”给多少年数据、多大效应,检验才有把握”(PMID 21640554,PMID 31886433)。空气污染时间序列研究里也有同样的工具:用模拟生成带时间趋势和气象混杂的日度序列,逐场景估计功效(PMID 22995599)。这套方法完全可以平移到干支检验上。

趋势是通过两条路把功效吃掉的。 第一条路最直观:去趋势的动作本身会去掉一部分真信号。如果干支效应和暖化趋势在时间上部分同向(比如”火太过年偏热”恰好与暖化抬升段叠加),那么线性或多项式去趋势在削掉暖化的同时,也会连带削掉一块干支效应——剩下的残差里,真信号的量级被系统性缩小了。第二条路更隐蔽:趋势会抬高残差的方差和自相关。非平稳序列里,单调趋势的能量会向相邻频段泄漏,把背景噪声抬起来。噪声一抬高,信噪比就下降,同样大小的真效应更难被检出。更麻烦的是,自相关本身就是功效的杀手——在中断时间序列的模拟里,效应量固定时,成功检出率会随自相关上升而下降(PMID 21640554);而 Mann-Kendall 类检验在自相关数据上,其统计量的真实分布会明显偏离高斯假设、方差被估计得偏小,导致显著性判断失真(PMID 41712666)。两条路叠加,效果就是:趋势越强,去趋势通道”看不见干支”的概率越高——不是因为干支不存在,而是因为检验的眼睛被趋势蒙住了。

标定怎么做:画一条”趋势强度 vs 检出功效”的曲线。 具体设计很直接:生成一批合成序列,每一批里塞入一个固定的、已知量级的干支效应(可以直接设为零,用来测误报率;也可以设成理论预期的小效应,用来测检出率),然后把叠加的暖化趋势强度从 0 一路拉到观测值的 2 倍——0 倍是”没暖化”的对照,1 倍是”真实观测到的暖化”,2 倍是”暖化再翻一倍”的压力测试。对每个趋势强度,跑完整的双通道流程(含趋势 vs 去趋势),重复上千次,数出”成功检出真效应”的比例就是功效,”无中生有报出假效应”的比例就是误报率。把这两条比例随趋势强度画出来,会得到一条随趋势上升而下滑的功效曲线,以及一条随趋势上升而抬头的误报曲线。两条曲线交叉、或者功效跌破常用门槛(比如 0.8)的那个点,就是功效崩溃阈值。

有了阈值,真实数据才有参照系。 这套标定真正的价值不在合成序列本身,而在于给真实数据的结论装了一把尺子:假如真实观测的暖化强度恰好落在功效曲线还很平缓的区间,”去趋势后不显著”就可以放心读成”没有可检出的干支信号”;但如果观测暖化强度已经逼近甚至越过了崩溃阈值,那么”去趋势后不显著”就只是“检不出来”,而不是”不存在”——二者在逻辑上天差地别,前者是阴性结论,后者是”证据不足、无法下判”。工程上还应当把两条机制分开标定:一条序列只加趋势、不加自相关,另一条只加自相关、不加趋势,看是”去趋势削掉真效应”占主导,还是”方差膨胀抬高噪声”占主导——这决定了补救方向是改去趋势方法,还是干脆换成对方差更稳健的方差校正型检验。

一句话总结:上一篇把趋势从信号里赶了出去,这一篇要问的是——赶走趋势的时候,有没有连信号一起赶走了。 只有在合成序列上标定出”趋势强度 vs 检出功效”的曲线,找到功效崩溃的临界点,”去趋势后不显著”这句结论才配有一个确定的含义:是在说真话,还是只是被趋势堵住了嘴。

证据与局限

已验证: 检验功效必须依赖”真情况已知”才能计算,模拟法是时间序列功效估计的标准路径(PMID 21640554,PMID 31886433,PMID 22995599);在固定效应量下,自相关上升会降低检验的成功检出率、抬高 I 类错误(PMID 21640554);Mann-Kendall 类检验在自相关数据上的统计量分布偏离高斯假设、方差被低估,导致显著性判断失真(PMID 41712666);模拟法样本量与功效估计在普通最小二乘中断时间序列中已成熟应用(PMID 30881136)。

部分验证: 上述功效标定框架均来自气候、水文、流行病学领域的时间序列研究,其”趋势强度-功效”曲线的形状与自相关结构、序列长度、去趋势方法强相关;但尚无任何研究把”干支分组效应 × 暖化趋势梯度 × 合成序列功效曲线”组合成完整的崩溃阈值标定流程,本课题属方法迁移性探索。

待验证: ① 在人为设定的固定干支效应下,去趋势通道的检出功效随叠加暖化趋势强度(0—2 倍观测值)下降的实际曲线形状与崩溃阈值位置;② “去趋势削掉真效应”与”方差膨胀抬高噪声”两条机制各自的贡献占比(可分解标定);③ 在观测暖化强度下,崩溃阈值是否已被越过——即既有”去趋势后不显著”的结论中,有多少属于”真阴性”,有多少属于”功效不足无法判定”。

相关阅读

证据等级: B
参考文献: PMID 41712666, PMID 21640554, PMID 31886433, PMID 22995599, PMID 30881136


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。

Sensitivity Calibration of Trend Strength on the Power of Detrended Channel Tests: When Trends Are Too Strong, the Dry Branches Become Invisible in the Detrended Channel

Evidence level: B

The previous article stated that “only stem-branch signals standing firm after detrending count as true signals.” However, the act of detrending itself consumes testing power—stronger trends mean more must be shaved off before searching for the remaining stem-branch signals, making it increasingly likely to find “nothing.” Therefore, “non-significance after detrending” cannot be directly read as “no stem-branch signal”; it requires an additional layer of validity calibration: at what trend strength does this channel go blind?

Main Text

The previous article (#426) established the first checkpoint for validating Five Movements and Six Qi: facing the unidirectional warming trend since 1961, observing only “trend-inclusive comparisons” yields false agreement; one must proceed through a “detrending channel”—shave off the trend, then assess how much effect remains within stem-branch groupings. Yet, a sharper and often overlooked question immediately arises: Will the detrending channel itself shave off true signals? If the answer is “yes,” then the statement “non-significant after detrending” becomes ambiguous—it could mean “there truly are no stem-branch signals,” or it could mean “the trend is too strong, crushing the test.”

Testing power can only be calculated when the “true state” is known. Power ($power = 1 – \beta$) is defined as “the probability that the test identifies the effect when it truly exists.” The problem is: on the real Earth, we do not know whether stems and branches possess a true effect or its magnitude; thus, power cannot be calculated directly from real data but must be “inferred” regarding significance. To quantify “how much grip remains in the detrending channel,” the only path is synthetic sequence calibration: generate a batch of artificial sequences with “known true states,” artificially inject a predefined stem-branch effect, run the identical detrending process, and observe how many are detected. This is the standard practice in epidemiology for Interrupted Time Series (ITS) studies—researchers use simulated data to estimate “how many years of data and what effect size are needed for the test to have confidence” (PMID 21640554, PMID 31886433). Similar tools exist in air pollution time series research: generating daily sequences with temporal trends and meteorological confounders via simulation, estimating power scenario by scenario (PMID 22995599). This methodology can be fully translated to stem-branch testing.

Trends consume power via two pathways. The first is most intuitive: The act of detrending removes part of the true signal. If the stem-branch effect and warming trend align partially in time (e.g., “excessive fire year heat” coinciding with a warming uplift), then linear or polynomial detrending, while shaving off warming, will also concurrently shave off a portion of the stem-branch effect—the residual signal’s magnitude is systematically reduced. The second pathway is more subtle: Trends inflate the variance and autocorrelation of residuals. In non-stationary sequences, monotonic trend energy leaks into adjacent frequency bands, raising the background noise. Once noise rises, the signal-to-noise ratio drops, making it harder to detect effects of the same magnitude. More麻烦的是 (troublesome), autocorrelation itself is a killer of power—in ITS simulations, with fixed effect sizes, detection success rates decline as autocorrelation rises (PMID 21640554); while Mann-Kendall type tests on autocorrelated data, their test statistics’ true distributions deviate significantly from Gaussian assumptions and variances are underestimated, leading to distorted significance judgments (PMID 41712666). With both pathways叠加 (superimposed), the result is: The stronger the trend, the higher the probability that the detrending channel becomes “blind to stems and branches”—not because stems and branches do not exist, but because the test’s eyes are blinded by the trend.

How to calibrate: Plot a “Trend Strength vs. Detection Power” curve. The design is straightforward: generate a batch of synthetic sequences, each containing a fixed, known-magnitude stem-branch effect (can be set to zero to measure false positive rates; or set to a theoretical small effect to measure detection rates), then ramp up the superimposed warming trend intensity from 0 all the way to twice the observed value—0x is the “no warming” control, 1x is the “observed warming,” and 2x is a “double-warming” stress test. For each trend intensity, run the full dual-channel process (including trend vs. detrending), repeat thousands of times, and count the proportion of “successful detection of true effects” as power, and the proportion of “finding false effects from nothing” as false positive rate. Plotting these two proportions against trend intensity yields a power curve declining with rising trends and a false positive curve rising with rising trends. The point where the two curves cross, or where power drops below a common threshold (e.g., 0.8), is the power collapse threshold.

With a threshold, real data gains a reference frame. The true value of this calibration lies not in the synthetic sequences themselves, but in equipping conclusions from real data with a ruler: if the observed warming intensity falls within a region where the power curve is still flat, “non-significant after detrending” can be confidently read as “no detectable stem-branch signal”; but if the observed warming intensity approaches or exceeds the collapse threshold, then “non-significant after detrending” is merely “undetectable,” not “non-existent”—logically distinct, the former is a negative conclusion, the latter is “insufficient evidence to judge.” Engineering-wise, both mechanisms should be calibrated separately: one sequence adds only trend without autocorrelation, the other adds only autocorrelation without trend—to determine whether “detrending shaving off true effects” dominates or “variance inflation raising noise” dominates—deciding whether remediation should target the detrending method or switch to a variance-corrected test with more robust variance estimation.

In summary: The previous article drove the trend out of the signal; this one asks—when driving the trend out, did we also drive out the signal? Only by calibrating the “trend intensity vs. detection power” curve on synthetic sequences and finding the critical point of power collapse does the conclusion “non-significant after detrending” deserve a definite meaning: is it speaking truth, or merely being silenced by the trend?

Evidence and Limitations

Verified: Testing power relies on “known true state” for calculation; simulation is the standard path for time series power estimation (PMID 21640554, PMID 31886433, PMID 22995599); under fixed effect sizes, rising autocorrelation reduces detection success rates and inflates Type I error (PMID 21640554); Mann-Kendall type test statistics on autocorrelated data deviate from Gaussian assumptions and suffer variance underestimation, leading to distorted significance judgments (PMID 41712666); simulation-based sample size and power estimation are mature in Ordinary Least Squares Interrupted Time Series (PMID 30881136).

Partially Verified: The above power calibration frameworks originate from time series research in climate, hydrology, and epidemiology; their “trend intensity-power” curve shapes are strongly correlated with autocorrelation structures, sequence lengths, and detrending methods; however, no study yet combines “stem-branch grouping effects × warming trend gradient × synthetic sequence power curves” into a complete collapse threshold calibration process; this topic belongs to methodological transfer exploration.

To Be Verified: ① Under artificially fixed stem-branch effects, the actual curve shape and collapse threshold location of detection power in the detrending channel as superimposed warming trend intensity increases (0–2x observed values); ② The respective contribution proportions of the two mechanisms: “detrending shaving off true effects” vs. “variance inflation raising noise” (decomposable calibration); ③ Under observed warming intensity, whether the collapse threshold has been crossed—i.e., among conclusions of “non-significant after detrending,” how many belong to “true negatives” and how many to “insufficient power for judgment.”

Related Reading

Evidence Level: B
References: PMID 41712666, PMID 21640554, PMID 31886433, PMID 22995599, PMID 30881136


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 *