月尺度的地磁累积 vs 周尺度的湿度异常:两个暴露窗口放进同一个模型,会互相吃掉吗?

Written by

in

Language

Evidence level: B

一个要量六个月的累积,一个要量三周的异常。两个时钟刻度差了一个数量级,却在同一条死亡曲线上争同一块方差——这不是谁对谁错,是一场必须先设计好的窗口竞争。

一、两个时钟不在同一个刻度上

2026 年一项全美研究给了地磁侧明确的时间常数:2000–2019 年 47 个州的月度呼吸道感染(RTI)死亡数据,把 Kp 指数移动平均从 1 个月扫到 12 个月,6 个月窗口关联最强,85 岁以上女性每升高一个四分位距 RR 1.16(95% CI 1.14–1.18),男性 1.14(1.12–1.16)(PMID 42350784)。这不是”磁暴当天死人多”,而是”连续半年持续扰动”。

气象侧的常数完全不同。绝对湿度(AH)解释豚鼠模型 50% 的传播变异与 90% 的病毒存活变异(PMID 19204283),人群层面表现为流行启动前数周的异常低 AH(PMID 20186267)。广州 42,440 例呼吸系统死亡的 DLNM 分析把窗口定在滞后 0–35 天,非最优 AH 贡献了 21.57% 的死亡(PMID 34789160)。

一个要半年,一个要三周。把这两个暴露放进同一个回归,等于让米尺和千分尺去量同一块布。

二、把尺子张成一张面

标准工具是分布滞后非线性模型(DLNM):把”暴露水平”与”滞后天数”张成交叉基,输出一张二维的暴露-滞后-反应面(PMID 20812303)。双暴露联合建模,就是把 Kp 的月尺度累积基与 AH 的周尺度异常基同时放进模型。技术上可行,三个陷阱立刻出现。

冬季共线。 冬季既是低 AH 期,也常被指为地磁相关窗口,两者在同一时段共同变动。直接并列时,回归会把效应记到”解释力更强”的那一方名下,而记错的一方不会报警。

窗口不是越多越好。 一份职业队列研究做了直接对照:把暴露按滞后年逐段展开的 DLNM,率比(肺癌 1.47、非恶性呼吸病 1.80)显著高于”线性累积暴露”这种简单汇总(1.15、1.23)(PMID 29447338)。但更高的估计值不等于更接近真相——它也可能是把同一份暴露切碎后反复计分。

尺度错配的报应。 用周尺度模型拟合月尺度效应,表现为”所有滞后系数都小、都勉强显著”;用月尺度窗口套周尺度触发,则表现为”效应被平均稀释到看不见”。

三、窗口竞争的三种结局

两个窗口同时进模型,逻辑上只有三种可能,各自对应完全不同的生物学解释:

① 替代:AH 进来后 Kp 系数塌掉。说明地磁侧”效应”很大程度只是冬季气象的代理,免疫通路站不住。

② 加性:两者各自保持显著,系数互不侵蚀。这是双通路独立起作用的证据——也是这个领域真正想看到的。

③ 交互:乘积项显著。意味着”高湿度异常期遇上高地磁累积期”的风险超过两者之和。

裁决工具是模型比较,不是看谁的 p 值小。一份模拟研究给了关键校准:最小 AIC 准则能在绝大多数样本里认出正确模型,但必须把”正确模型或与它相差 4 个 AIC 点以内的等价模型”一起看——AIC 相差小于 4 分的两个窗口设定,统计上根本不可区分(PMID 22095719)。任何”6 个月就是最优窗口”的结论,都必须公布它在 5 个月和 7 个月窗口下的 AIC 差了多少。

四、负的滞后系数不是错误

联合滞后结构里一定会出现某段系数为负。这不是 bug,是死亡置换(harvesting):脆弱个体被提前几天带走,之后几天死亡数反低于基线。

希腊 2013–2017 年流感归因死亡研究把这件事量化了:A(H1N1)pdm09 流行期,65 岁以上人群周归因死亡先快速上升,随后降到零甚至负数(PMID 30968823)。广州 199,777 份死亡证明的 DLNM 分析给出置换时长——A(H1N1)、A(H3N2)、B/Yamagata 为 2–5 天,B/Victoria 长达 5–13 天(PMID 36630176)。

结论很硬:最大滞后必须拉得足够长(至少覆盖一个完整置换周期),否则负系数会被误读成”保护效应”。窗口截断本身就是偏倚来源。

五、伏邪与新感

这套双窗口结构,中医早就分过。

月尺度累积 = 伏邪。“伏气温病”讲邪气潜伏、日久蓄积,条件成熟才发;对应 Kp 六个月移动平均那种”没有一天特别糟,但半年都不太平”的形态。

周尺度异常 = 新感。“新感温病”讲当令之邪即时伤人;对应 AH 在流行启动前数周的异常下探,窗口短、触发明确。

临床上”伏邪”与”新感”从不是二选一——先有伏邪垫底,再遇新感触发,才有大疫。这正是交互项(结局③)要检验的东西,也是”正气存内,邪不可干”的统计翻译:同一个 AH 异常,对不同 Kp 累积背景的人,风险不是同一个数。

六、一条必须先说死的规则

太阳活动与流感大流行的历史关联,是这个领域最容易被”钓”出来的结论。有分析声称几乎所有记录在案的流感大流行都落在太阳黑子极值 ±1 年内(OR 3.87,95% CI 1.08–13.85)(PMID 27136236);而一篇重分析逐条复核后发现,那些显著性来自大流行年份抄录错误、统计方法误用、以及对任意数据选择假设不稳健,修正后全部消失(PMID 28847318)。

这条前车之鉴直接决定设计纪律:窗口必须在看到数据之前指定。1–12 个月全扫一遍再挑最强的那一个,本质是多重比较,不是发现。可行做法是预注册主窗口(Kp 累积取 6 个月、AH 异常取 0–35 天),其余窗口作为敏感性分析并列报告,并把 AIC 差一并公布。

两个窗口争的从来不是”谁是真的”,而是”它们各自负责哪一段时间”。 如果设计允许研究者事后挑窗口,这场竞争就没有裁判。

证据与局限

已验证:
– 地磁累积暴露与 RTI 死亡关联在 6 个月窗口最强,高龄组 RR 1.14–1.16(PMID 42350784)
– 绝对湿度经机制路径调制流感病毒存活与传播,人群表现为流行前数周异常低 AH(PMID 19204283, PMID 20186267)
– AH 与呼吸系统死亡呈非线性关系,非最优 AH 贡献 21.57% 死亡负担(PMID 34789160)
– DLNM 交叉基是暴露-滞后-反应面建模的标准框架(PMID 20812303)
– 流感死亡存在可量化的死亡置换,时长 2–13 天不等(PMID 36630176, PMID 30968823)
– AIC 模型比较的 4 分等价规则已由模拟标定(PMID 22095719)

部分验证:
– 月尺度与周尺度双暴露的联合滞后结构尚无实际研究完成,三种结局(替代/加性/交互)均未被检验
– 冬季共线对两个窗口系数分配的污染比例未量化

待验证:
– 同队列双暴露联合 DLNM:控制 AH 周尺度异常后,Kp 月尺度累积系数是否保留
– 两个窗口的乘积项是否显著(叠加效应是否存在)
– 最大滞后长度选择对死亡置换方向判定的敏感性

相关阅读

证据等级: B
参考文献: PMID 42350784, PMID 19204283, PMID 20186267, PMID 34789160, PMID 20812303, PMID 29447338, PMID 22095719, PMID 30968823, PMID 36630176, PMID 27136236, PMID 28847318


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。

Monthly geomagnetic accumulation vs. weekly humidity anomalies: Will two exposure windows cancel each other out in the same model?

Evidence level: B

One requires measuring six-month accumulation; the other requires measuring three-week anomalies. Two clock scales differ by an order of magnitude, yet they fight for the same variance on a single mortality curve. This is not a matter of right or wrong, but a window competition that must be designed in advance.

I. Two Clocks Not on the Same Tick

A 2026 US-wide study provided geomagnetic side a clear time constant: monthly respiratory tract infection (RTI) mortality data from 47 states between 2000–2019 showed that moving averages of the Kp index from 1 to 12 months yielded the strongest association at a 6-month window, with an RR of 1.16 (95% CI 1.14–1.18) per interquartile range increase for women aged ≥85, and 1.14 (1.12–1.16) for men (PMID 42350784). This is not “more deaths on geomagnetic storm days,” but rather “continuous disturbance over six consecutive months.”

The meteorological constant is entirely different. Absolute humidity (AH) explains 50% of the transmission variability and 90% of the virus survival variability in guinea pig models (PMID 19204283), manifesting at the population level as abnormally low AH in the weeks preceding epidemic onset (PMID 20186267). A DLNM analysis of 42,440 respiratory deaths in Guangzhou set the window at a lag of 0–35 days, where non-optimal AH contributed 21.57% of deaths (PMID 34789160).

One requires six months; the other requires three weeks. Putting both exposures into the same regression is like using a meter stick and a micrometer to measure the same piece of cloth.

II. Stretching the Ruler into a Surface

The standard tool is the Distributed Lag Non-Linear Model (DLNM): stretching “exposure levels” against “lag days” into cross-basis functions to output a two-dimensional exposure-lag-response surface (PMID 20812303). Joint modeling of dual exposures involves placing both the Kp monthly accumulation basis and the AH weekly anomaly basis simultaneously into the model. Technically feasible, three traps immediately arise.

Winter Collinearity. Winter is both a low-AH period and frequently cited as a geomagnetic-related window; both vary simultaneously during this season. When entered directly, regression attributes the effect to whichever variable shows stronger explanatory power, while the misattributed one fails to signal an error.

More Windows Are Not Better. One occupational cohort study performed a direct comparison: DLNMs expanding exposure by lag years showed significantly higher rate ratios (lung cancer 1.47, non-malignant respiratory disease 1.80) compared to “linear cumulative exposure” simple summation (1.15, 1.23) (PMID 29447338). However, higher estimated values do not equal proximity to truth—they may simply reflect scoring the same exposure repeatedly after chopping it into smaller pieces.

Consequences of Scale Mismatch. Fitting a weekly-scale model to a monthly-scale effect manifests as “all lag coefficients small and barely significant”; forcing a monthly-scale window onto weekly-scale triggers manifests as “effects diluted to invisibility by averaging.”

III. Three Outcomes of Window Competition

Logically, there are only three possibilities when two windows enter the model simultaneously, each corresponding to a completely different biological interpretation:

① Substitution: The Kp coefficient collapses after AH enters. This suggests the geomagnetic side’s “effect” is largely a proxy for winter meteorology, and immune pathways are not holding up.

② Additive: Both remain significant with coefficients not eroding each other. This is evidence of dual pathways acting independently—what this field truly hopes to see.

③ Interaction: The product term is significant. Meaning the risk during “high humidity anomaly periods coinciding with high geomagnetic accumulation” exceeds the sum of both.

The adjudication tool is model comparison, not who has the smaller p-value. One simulation study provided key calibration: The minimum AIC criterion can identify the correct model in most samples, but must view “the correct model or equivalent models within 4 AIC points together”—two window settings with an AIC difference less than 4 are statistically indistinguishable (PMID 22095719). Any conclusion claiming “6 months is the optimal window” must disclose the AIC difference under 5-month and 7-month windows.

IV. Negative Lag Coefficients Are Not Errors

Within a joint lag structure, some coefficients will inevitably be negative. This is not a bug but death harvesting (harvesting effect): vulnerable individuals are removed days in advance, causing death counts to subsequently drop below baseline for several days.

A Greek 2013–2017 influenza-attributable mortality study quantified this: during the A(H1N1)pdm09 epidemic, weekly attributable deaths in populations aged ≥65 rose rapidly first, then dropped to zero or even negative values (PMID 30968823). A DLNM analysis of 199,777 death certificates in Guangzhou provided the duration of harvesting—2–5 days for A(H1N1), A(H3N2), and B/Yamagata, and up to 5–13 days for B/Victoria (PMID 36630176).

The conclusion is firm: Maximum lag must be stretched sufficiently long (covering at least one complete harvesting cycle), otherwise negative coefficients will be misread as “protective effects.” Window truncation itself is a source of bias.

V. Stored Pathogen vs. New Attack

This dual-window structure was already distinguished by Traditional Chinese Medicine (TCM).

Monthly-scale accumulation = Stored Pathogen (Fu Xie). “Fu-temperature diseases” describe pathogens lurking and accumulating over time, erupting only when conditions mature; corresponding to the Kp six-month moving average form where “no single day is particularly bad, but six months are never peaceful.”

Weekly-scale anomaly = New Attack (Xin Gan). “New-attack warm diseases” describe immediate injury by the prevailing pathogen; corresponding to AH dipping abnormally in the weeks preceding epidemic onset, with a short window and clear trigger.

Clinically, “Stored Pathogen” and “New Attack” are never mutually exclusive—only with a Stored Pathogen base layer followed by a New Attack trigger does a major epidemic occur. This is precisely what the interaction term (Outcome ③) must test, and also the statistical translation of “when the body’s vital energy is preserved within, the pathogen cannot invade”: the same AH anomaly carries different risk numbers for people with different Kp accumulation backgrounds.

VI. A Rule That Must Be Stated First

The historical association between solar activity and influenza pandemics is the conclusion in this field most easily “hooked” out. One analysis claimed almost all recorded influenza pandemics fell within ±1 year of solar sunspot extrema (OR 3.87, 95% CI 1.08–13.85) (PMID 27136236); however, a re-analysis reviewing each case line-by-line found that those significances stemmed from transcription errors in pandemic years, misuse of statistical methods, and lack of robustness to arbitrary data selection assumptions, all vanishing upon correction (PMID 28847318).

This cautionary tale directly dictates design discipline: Windows must be specified before seeing the data. Sweeping through 1–12 months to pick the strongest one is essentially multiple comparisons, not discovery. A feasible approach is pre-registering primary windows (Kp accumulation at 6 months, AH anomaly at 0–35 days), reporting remaining windows as sensitivity analyses alongside them, and disclosing AIC differences simultaneously.

The two windows are never fighting over “who is true,” but rather “which time period they are each responsible for.” If design allows researchers to pick windows post-hoc, this competition has no referee.

Evidence and Limitations

Verified:
– Strongest geomagnetic accumulation exposure association with RTI deaths at 6-month window, RR 1.14–1.16 in elderly groups (PMID 42350784)
– Absolute humidity modulates influenza virus survival and transmission via mechanistic pathways; population manifestation is abnormally low AH in weeks preceding epidemics (PMID 19204283, PMID 20186267)
– Non-linear relationship between AH and respiratory deaths; non-optimal AH contributes 21.57% of death burden (PMID 34789160)
– DLNM cross-basis is the standard framework for modeling exposure-lag-response surfaces (PMID 20812303)
– Influenza deaths exhibit quantifiable death harvesting with durations ranging from 2–13 days (PMID 36630176, PMID 30968823)
– The 4-point equivalence rule for AIC model comparison calibrated by simulation (PMID 22095719)

Partially Verified:
– No actual studies completed yet on joint lag structures for dual exposures at monthly and weekly scales; none of the three outcomes (substitution/additive/interaction) have been tested
– The proportion of contamination to coefficient allocation by winter collinearity remains unquantified

To Be Verified:
– Dual-exposure joint DLNM in the same cohort: whether Kp monthly accumulation coefficients persist after controlling for AH weekly-scale anomalies
– Whether the product term between two windows is significant (existence of additive/synergistic effects)
– Sensitivity of maximum lag length selection on determining the direction of death harvesting

Related Reading

Evidence Level: B
References: PMID 42350784, PMID 19204283, PMID 20186267, PMID 34789160, PMID 20812303, PMID 29447338, PMID 22095719, PMID 30968823, PMID 36630176, PMID 27136236, PMID 28847318


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 *