去趋势的三条岔路:同一个中子通量与病毒突变率,三种洗法能洗出三个结论

Written by

in

Language

Evidence level: B

上一篇已经把”天时改疫”翻译成一道可算的统计题。这一篇要说的是:这道题在动手计算之前,还横着一个不起眼的岔路口——你怎么把趋势洗掉,决定你会拿到什么答案。

正文

上一篇(neutron-flux-virus-mutation-lag-cross-correlation)把”岁运影响疫变”从”拿疫情年份去对太阳极小期”改写成了”时序对时序”:中子监测站的通量序列,对病毒系统发育的 root-to-tip(RTT)突变速率序列,做滞后互相关。设计是对的,但它内部埋了一个动词——去趋势。而这一个动词,是整条链路上最容易让结论掉头的地方。

为什么非洗不可。 两条真实的地球/生物序列,几乎必然都带趋势。中子通量侧随太阳活动周期涨落,突变率侧则有”时间依赖速率现象”(TDRP):速率估计随观测时间尺度系统性衰减,机制被建模为”最快进化位点饱和 + 纯化选择”的组合(PMID 34478645)。再叠加 RNA 病毒本就存在的谱系间速率变异(PMID 11821909 对 50 种 RNA 病毒的定量系统发育分析显示严格分子钟常被拒绝),”突变率序列”本身就是一个带强趋势、带低频漂移的量。两条各自带趋势的序列做相关,即使物理上毫无关系,也能因共享低频漂移获得虚假的正相关——这在气候检测与归因里已经被系统论证过(PMID 35345504)。更麻烦的是”趋势”这个词本身不唯一:对非线性非平稳序列,趋势的定义取决于你用哪种自适应分解(PMID 17846430),而温度类序列的长程持续性又会让”去掉趋势”这个动作本身产生方法依赖(PMID 22181223)。所以去趋势不是清洁步骤,而是建模选择。

三条岔路,是三台不同的机器。 一阶差分(Δxₜ = xₜ − xₜ₋₁)是最常见的做法,它的传递函数是 1 − z⁻¹——一台在低频增益趋零、在奈奎斯特频率增益为 2 的高通滤波器。它的副作用是双重的:既把低频真效应一起清零,又会在白噪声上引入约 −0.5 的负自相关。低通滤波减去低通成分(或滑动平均去趋势)表面上收窄了频段,但有限冲激响应带来的边缘效应和相位失真会跟着进残差。谐波回归走的是第三条路:把趋势项和季节谐波一起放进设计矩阵做最小二乘,再从残差里取信号——它相当于在频域定点扣除若干频率,既不做全频段高通,也不引入差分式负自相关,但代价是”扣几个谐波、趋势用几次多项式”变成一串人为的自由度选择。三种做法的共同反讽是:它们都声称”去掉了趋势”,却给出了三份互不相同的残差序列,而残差的自相关结构,正是决定下一步互相关显著性的唯一输入(PMID 39815636 对随机/确定性趋势处理的系统讨论)。

洗完之后,显著性还得重算。 互相关的 p 值绝不能按样本量 n 直接查表。序列的残余自相关会把独立观测的”有效数量”压到远低于 n,此时名义 5% 的检验可能实际有 30% 以上的假阳性——这正是单例研究与过程控制数据里反复被证实的推断陷阱(PMID 26463724、PMID 23705957)。正确做法有三条并行的路:一是有效自由度校正,用序列自身的谱密度换算出等价的独立观测数,再据此构造相关系数的置信区间(PMID 39846001 给出了适用于生理信号的参数化方案);二是预白化(prewhitening),先按 ARIMA 结构把序列白化再算互相关——这个方法已经在真实流行病学问题上落地,比如韩国自杀相关网络搜索与自杀率的时间关系检验就明确采用了预白化互相关(PMID 41662218);三是重采样/替代数据,在彩色噪声的框架下用时间域与波域的重采样构造零分布(PMID 11169871),对功率律互相关过程也有专门的检验(PMID 22304166)。此外,当序列存在长程记忆时,”趋势是否显著”本身就要用考虑记忆的显著性框架来判(PMID 37271384)。

最难的是把话说对。 三种洗法跑完,如果结论一致,说明”通道”的信号强到能扛过方法选择——这是最理想的结局。如果不一致,则眼下的结论其实是方法的函数,而不是自然的函数,此时任何单点显著都不该被当成证据。还有两个方向相反的表述陷阱必须避免:去趋势后不显著,不等于通道不存在(可能是真阴性,也可能是方法把真效应一起切掉了);去趋势后显著,也不等于通道存在(可能只是残余自相关下的假自由度)。上一篇引过的反证在这里尤其重要:2008–2009 与 2019–2020 两次太阳深极小期中子通量同样偏高,却并未对应同等的病毒登场节奏(PMID 35194566)。

对易学的意义。 这一篇没有一个新假说,它只做一件事:把”天时改疫”能不能被检验这件事,与”检验者怎么选方法”彻底分开。做法就是把三种去趋势算子在同一份中子通量序列、同一份突变率序列上并行跑,把三套相关系数、三套置信区间、三套显著性并排列出来,并预先登记方法选择与判定规则。若三套一致指向零,那”岁运直接改写病毒基因组”这条通道基本可以放下;若三套都指向非零且方向一致,它才值得被认真追查。五运六气的科学价值从来不在于”某年对上了”,而在于”它敢被放在一套会失败的检验里”——去趋势方法学正是这套检验里最容易被糊过去、也最必须写清楚的一环。

证据与局限

已验证: 病毒替换速率存在时间依赖衰减(TDRP),机制为位点饱和与纯化选择的组合(PMID 34478645);RNA 病毒进化速率存在显著谱系间变异,严格分子钟常被拒绝(PMID 11821909);带趋势的非平稳序列做相关/回归可产生伪结果(PMID 35345504);趋势与去趋势的定义依赖方法选择(PMID 17846430);自相关会压缩有效自由度、膨胀 I 类错误(PMID 26463724, PMID 23705957);有效自由度校正(PMID 39846001)、预白化互相关(PMID 41662218)、时间/波域重采样(PMID 11169871)、功率律互相关检验(PMID 22304166)、多尺度去趋势互相关(PMID 39605794)均为成熟可用工具。

部分验证: 预白化互相关在流行病学与公共卫生监测中已有实际应用案例(PMID 41662218),但尚未成为病毒进化-天体指数类研究的常规步骤;长程记忆下的趋势显著性框架(PMID 37271384)验证于大气污染物趋势,尚未在病毒速率序列上标定。

待验证: 同一中子通量—突变率序列在一阶差分、低通滤波去趋势、谐波回归三种算子下的相关系数与显著性漂移幅度;去趋势后残差自相关结构(及其对应的有效自由度)对互相关显著性阈值的影响量级;三种算子下结论一致性是否构成”通道存在”的充分判据。

相关阅读

证据等级: B(TDRP 机制、RNA 病毒速率变异、伪相关与去趋势方法学、有效自由度/预白化/替代数据工具均有 PubMed 直接支撑;”三种去趋势算子导致互相关显著性漂移”这一具体命题尚无直接实证,属方法学类比与间接证据链)
参考文献: PMID 34478645, PMID 11821909, PMID 35345504, PMID 17846430, PMID 22181223, PMID 39815636, PMID 39846001, PMID 41662218, PMID 11169871, PMID 22304166, PMID 39605794, PMID 37271384, PMID 26463724, PMID 23705957, PMID 35194566


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。

Three Paths to De-trending: Same Neutron Flux and Virus Mutation Rate Yield Three Different Conclusions via Three Washing Methods

Evidence level: B

The previous article translated “heavenly timing alters epidemics” into a calculable statistical problem. This piece addresses the fork encountered before calculation begins: how you choose to de-trend determines which answer you receive.

Main Text

The previous article (neutron-flux-virus-mutation-lag-cross-correlation) reframed “solar cycle influences epidemic shifts” from matching epidemic years to solar minima into a time-series-to-time-series analysis: cross-correlating neutron monitor flux sequences with root-to-tip (RTT) mutation rate sequences from viral phylogenies. The design is sound, but it contains an embedded verb—de-trending. This single verb is the most critical point where conclusions can diverge along the entire chain.

Why de-trending is non-negotiable. Two real Earth/biological sequences almost certainly carry trends. On the neutron flux side, fluctuations follow solar activity cycles. On the mutation rate side, there is a “Time-Dependent Rate Phenomenon” (TDRP): rate estimates systematically decay with observation time scales, modeled as a combination of “fastest-evolving site saturation + purifying selection” (PMID 34478645). Adding intrinsic rate variation across lineages in RNA viruses (PMID 11821909 shows strict molecular clocks are frequently rejected in quantitative phylogenetic analyses of 50 RNA viruses), the “mutation rate sequence” itself is a quantity with strong trends and low-frequency drift. Correlating two trend-bearing sequences can yield spurious positive correlation due to shared low-frequency drift even if they are physically unrelated—a phenomenon systematically demonstrated in climate detection and attribution (PMID 35345504). Complicating matters further, the definition of “trend” is not unique for nonlinear non-stationary sequences; it depends on the adaptive decomposition method chosen (PMID 17846430), and long-range persistence in temperature-like sequences means the act of “removing trend” itself introduces method dependence (PMID 22181223). Thus, de-trending is not a cleaning step but a modeling choice.

Three paths correspond to three different machines. First-order differencing (Δxₜ = xₜ − xₜ₋₁) is the most common approach; its transfer function is 1 − z⁻¹—a high-pass filter with zero gain at low frequencies and a gain of 2 at the Nyquist frequency. Its side effects are dual: it zeroes out low-frequency true effects while introducing approximately −0.5 negative autocorrelation on white noise. Subtracting low-pass components (or using sliding averages) narrows the band superficially, but finite impulse response introduces edge effects and phase distortion into the residuals. Harmonic regression takes a third path: including trend terms and seasonal harmonics in the design matrix for least squares, then extracting signals from residuals. This effectively subtracts specific frequencies in the frequency domain without full-band high-pass filtering or introducing differential-based negative autocorrelation, but the cost is that “which harmonics to subtract and how many polynomial terms for trend” becomes a series of arbitrary degree-of-freedom choices. The common irony is that all three claim to “remove trend” yet produce three distinct residual sequences, and the autocorrelation structure of these residuals is the sole input determining the significance of the subsequent cross-correlation (PMID 39815636 offers a systematic discussion on stochastic vs. deterministic trend handling).

Recalculate significance after cleaning. Cross-correlation p-values cannot be looked up directly using sample size $n$. Residual autocorrelation compresses the “effective number of independent observations” far below $n$; at this point, a nominal 5% test may yield over 30% false positives—a trap repeatedly confirmed in single-case studies and process control data (PMID 26463724, PMID 23705957). Correct approaches include three parallel paths: First, effective degrees of freedom correction, converting the sequence’s spectral density into an equivalent number of independent observations to construct confidence intervals for the correlation coefficient (PMID 39846001 provides a parameterized scheme suitable for physiological signals). Second, pre-whitening, where sequences are whitened via ARIMA structure before calculating cross-correlation—this method has already been applied to real epidemiological issues, such as the time relationship test between suicide-related network searches and suicide rates in Korea explicitly using pre-whitened cross-correlation (PMID 41662218). Third, resampling/alternative data, constructing null distributions via time-domain and wavelet-domain resampling under a colored noise framework (PMID 11169871), with specific tests available for power-law cross-correlation processes (PMID 22304166). Furthermore, when sequences exhibit long-range memory, the significance of “trend” itself must be judged using a significance framework accounting for memory (PMID 37271384).

The hardest part is speaking correctly. After running all three cleaning methods, if conclusions align, it indicates the “channel” signal is strong enough to withstand methodological choices—the ideal outcome. If they do not align, the current conclusion is actually a function of the method, not nature; thus, no single-point significance should be treated as evidence. Two opposing rhetorical traps must also be avoided: non-significance after de-trending does not mean the channel is absent (could be true negative or method cutting out true effects); significance after de-trending does not mean the channel exists (could be false degrees of freedom under residual autocorrelation). The counter-evidence cited previously is particularly important here: neutron flux was similarly elevated during the 2008–2009 and 2019–2020 solar grand minima, yet did not correspond to equivalent viral emergence rhythms (PMID 35194566).

Significance for Yijing Studies. This article proposes no new hypothesis; it performs one task: completely separates whether “heavenly timing alters epidemics” can be tested from how the “tester selects a method.” The procedure involves running all three de-trending operators in parallel on the same neutron flux sequence and the same mutation rate sequence, listing three sets of correlation coefficients, three sets of confidence intervals, and three sets of significance levels in advance, while registering method selection and decision rules. If all three point to zero, the channel of “solar cycles directly rewriting viral genomes” can be largely set aside. If all three point to non-zero with consistent direction, it warrants serious investigation. The scientific value of the Five Movements and Six Qi (Wu Yun Liu Qi) never lies in “matching a specific year,” but in its willingness to be placed within a suite of tests prone to failure—de-trending methodology is precisely the most easily overlooked yet must-be-clarified link in this testing suite.

Evidence and Limitations

Verified: Time-dependent decay exists in virus substitution rates (TDRP), mechanism being a combination of site saturation and purifying selection (PMID 34478645); RNA virus evolutionary rates exhibit significant lineage-to-lineage variation, strict molecular clocks frequently rejected (PMID 11821909); correlating/regressing non-stationary sequences with trends produces spurious results (PMID 35345504); definitions of trend and de-trending depend on methodological choices (PMID 17846430); autocorrelation compresses effective degrees of freedom and inflates Type I errors (PMID 26463724, PMID 23705957); effective degrees of freedom correction (PMID 39846001), pre-whitened cross-correlation (PMID 41662218), time/wavelet domain resampling (PMID 11169871), power-law cross-correlation tests (PMID 22304166), and multi-scale de-trending cross-correlation (PMID 39605794) are mature, available tools.

Partially Verified: Pre-whitened cross-correlation has practical application cases in epidemiology and public health monitoring (PMID 41662218), but has not yet become a routine step in virus evolution-astronomical index studies; trend significance frameworks under long-range memory (PMID 37271384) are validated for atmospheric pollutant trends but not yet calibrated on virus rate sequences.

To be Verified: Magnitude of correlation coefficient and significance drift across first-order differencing, low-pass filtering de-trending, and harmonic regression operators on the same neutron flux-mutation rate sequence; magnitude of impact of residual autocorrelation structure (and corresponding effective degrees of freedom) on cross-correlation significance thresholds after de-trending; whether consistency across three operators constitutes a sufficient criterion for “channel existence.”

Related Reading

Evidence Level: B (TDRP mechanisms, RNA virus rate variation, spurious correlation and de-trending methodology, effective degrees of freedom/pre-whitening/alternative data tools all have direct PubMed support; the specific proposition “three de-trending operators cause cross-correlation significance drift” lacks direct empirical evidence, belonging to methodological analogy and indirect evidence chains)
References: PMID 34478645, PMID 11821909, PMID 35345504, PMID 17846430, PMID 22181223, PMID 39815636, PMID 39846001, PMID 41662218, PMID 11169871, PMID 22304166, PMID 39605794, PMID 37271384, PMID 26463724, PMID 23705957, PMID 35194566


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 *