Evidence level: B
配对设计把气象混杂减掉了,但”超额死亡”本身是一个算法产物。Serfling、泊松回归、率差法在同一份数据上给出的数不一样——这不是稳健性检验的附属品,它是结论的一部分。
正文
一 同一个设计,三把尺子量出三个数
配对年份设计的核心动作是”减”:把气象条件相同的两个年份的死亡数相减,剩下的算地磁。但被减掉的那个”基线”不是一个数,是一个算法。
美国四项方法对比(PMID 19453440)显示,泊松回归给出年均 25,470 例流感相关死亡,peri-season 率差法给出 22,454 例,量级接近;但用夏季基线做率差法时系统性偏高。西班牙的方向相反——广义线性模型的估计系统性低于 Serfling,后者在 65 岁以上给出每十万人 57 与 164 例(PMID 22284214)。加拿大的二十年数据结论更直接:Serfling 估计高度依赖”每年排除多少周”,且大多数排除规则下低于泊松模型;拟合出的周期性基线会高估冬季的非流感死亡,从而低估流感超额,甚至产生负的年度估计(PMID 24312225)。
三个国家、三种方向。这说明不存在”更保守的方法”,只有”错法不同”。
这个设计要问的问题本身有据可依:全美 2000–2019 年数据显示,地磁 Kp 指数的六个月移动平均与呼吸道感染死亡显著相关(PMID 42350784),配对设计要检验的正是”气象扣掉之后,这份关联还剩多少”。而绝对湿度与流感活动的关系在中国随气候带反转(PMID 35218485)——这提示”气象”不是一个可以随手扣掉的常数。
二 最硬的一条:方法本身是估计值的预测因子
2018 年一项系统综述纳入 103 项流感死亡负担研究做 meta-regression,结论很直白——建模技术本身是估计值的重要预测因子:以乘数法为参照,Serfling 型模型给出的年度年龄标准化超额死亡率平均高出约 40/10 万(95%CI 3–77,P=0.033);同一批研究里,死因范围从”肺炎与流感”扩到”全因”,平均再增加 43/10 万(95%CI 32–54)(PMID 28679157)。
这对地磁研究的意义很直接:如果三法在同一份数据上的系统性差异(每十万人几十例的量级)大于地磁边际增量本身,那么”地磁显著”这个结论在很大程度上是方法选的,不是数据给的。
三 配对设计不能自动消除方法偏倚——它分成两半
配对设计的威力在减法。但方法偏倚在减法里的命运不一样:一部分是共同偏倚——只要两个年份的死亡基线形状相同,相减时抵消;另一部分是差分偏倚——当两个年份的死亡时间分布本身不同时,它留了下来,直接混进边际增量。
最典型的差分偏倚是死亡位移(harvesting):一次扰动提前带走了本来会在数周后死亡的人。JAMA Netw Open 对 2020–2024 年全球数据的评估显示,位移是普遍现象而非个案(PMID 41609824);BMC 的方法学研究给出了个体层面的反事实校正方案(PMID 37853353)。放回配对设计:若地磁活跃年的死亡曲线更”提前”,Serfling 基线会低估它、泊松模型会部分吸收它、率差法取决于基线期怎么划——三个方法对同一个位移的反应不同,这个差额不是地磁效应。
四 不确定性区间同样是方法的一部分
估计值之外还有区间。在过离散的计数序列上直接套泊松,会得到过窄的置信区间——2026 年 Epidemiol Infect 的模拟研究量化了这一点:泊松假设导致对有效再生数的估计”过度自信”,负二项与准泊松的区间显著更宽(PMID 42178993)。超额死亡领域同样如此:Popul Dev Rev 的敏感性分析发现,仅改变基线方法,立陶宛的超额死亡率估计可以差近 4 倍;仅改变参考期(2010–2019 vs 2017–2019),比利时的估计差 20%(PMID 35600716)。WHO 的方法学报告(PMID 36517599)、更灵活的替代框架(PMID 35383642)与贝叶斯时空变系数模型(PMID 40661232)处理的是同一件事:区间本身也是模型产物。
五 该怎么做
第一,方法事前预注册,并把三法并行跑完,报告”方法 × 纬度”交互,而不是事后挑一个最漂亮的。
第二,把方法差异本身当成一个可报告的效应量:报告”三法估计的极差 ÷ 地磁边际增量”这个比值。比值远大于 1,结论就不可信。
第三,别用 AIC 裁决方法——AIC 差 4 分以内应视为等价,不能据此选模型(PMID 22095719);暴露—滞后—反应场景下另有既定做法(PMID 20812303、PMID 29447338、PMID 23297754、PMID 35436963)。
第四,分层与及时评估:按年龄与亚型分层能暴露位移效应(PMID 30968823),AIC 驱动的模型选择步骤(PMID 26781954)与早期建立的及时监测方法(PMID 9209852)可直接复用。
六 易学落点
“数”从来不是中立的。同一片天时,用不同历法、不同起算点去量,得到不同的”气数”。《素问》讲”必先岁气,无伐天和”——先要定准这一年的”岁气”,而”定准”这件事本身就包含”用哪套算法”。
本篇的分辨点:若三法的差异主要是共同偏倚,配对设计仍可信,结论稳健;若差异主要来自差分偏倚(死亡位移),那”地磁边际增量”必须带着方法名一起说——它是”Serfling 下的增量”还是”泊松下的增量”,是两个命题。
证据与局限
已验证: 同一份数据上四种超额死亡方法量级接近但存在系统性差异(PMID 19453440);Serfling 估计高度依赖排除周数规则并可能产生负估计(PMID 24312225);建模技术本身是流感死亡负担估计值的显著预测因子(PMID 28679157);泊松假设在过离散序列上导致区间过窄(PMID 42178993);改变基线方法与参考期可使超额死亡估计相差数倍(PMID 35600716)。
部分验证: 死亡位移的全球普遍性(PMID 41609824,生态学汇总层面);个体层面反事实校正的可行性(PMID 37853353);灵活超额死亡框架与贝叶斯时空变系数模型的表现(PMID 35383642、PMID 40661232,模拟研究为主)。
待验证: ①”方法偏倚在配对设计中分成共同偏倚与差分偏倚”这一分解至今没有任何研究直接量化过,因而无法判断地磁边际增量中有多少来自差分偏倚;②三法在”地磁活跃年 vs 平静年”这一特定对比下的估计差异与覆盖概率从未被比较过;③若差分偏倚确实占主导,是否存在一个可事前指定的方法组合能把偏倚压到边际增量以下,仍是开放问题。
相关阅读
- 地磁活跃期对呼吸道超额死亡的边际增量:气象条件配对年份设计
- 把北美的配对年份设计搬到东亚和欧洲:地磁边际增量能否随纬度增强?
- 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?
- 环境-生命频率接口研究池
证据等级: B
参考文献: PMID 19453440, PMID 24312225, PMID 28679157, PMID 22284214, PMID 30968823, PMID 9209852, PMID 26781954, PMID 35600716, PMID 42178993, PMID 41609824, PMID 37853353, PMID 35383642, PMID 40661232, PMID 36517599, PMID 22095719, PMID 20812303, PMID 29447338, PMID 23297754, PMID 35436963, PMID 42350784, PMID 35218485
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 magnetic marginal increment still hold with an alternative excess mortality algorithm?
Evidence level: B
A matched-pair design mitigates meteorological confounding, yet “excess mortality” itself is an algorithmic construct. Serfling, Poisson regression, and rate difference methods yield different figures on the same dataset—this is not merely an artifact of robustness testing; it constitutes part of the conclusion.
Main Text
I. One Design, Three Rulers Yielding Three Numbers
The core operation of the matched-pair year design is “subtraction”: subtracting death counts from two years with identical meteorological conditions to isolate the geomagnetic signal. However, the subtracted “baseline” is not a single number; it is an algorithm.
Comparisons across four US methods (PMID 19453440) show that Poisson regression yields an annual average of 25,470 influenza-related deaths, while the peri-season rate difference method yields 22,454; their magnitudes are close. However, using a summer baseline for the rate difference method results in a systematic overestimation. The direction is reversed in Spain, where generalized linear models estimate systematically lower values than Serfling; the latter reports 57 and 164 cases per 100,000 for those aged 65+ (PMID 22284214). Canada’s twenty-year dataset offers a more direct conclusion: Serfling estimates are highly dependent on “how many weeks are excluded annually,” and under most exclusion rules, they fall below Poisson models; the fitted periodic baseline overestimates non-influenza winter deaths, thereby underestimating influenza excess and even generating negative annual estimates (PMID 24312225).
Three countries, three divergent directions. This indicates there is no “more conservative method,” only “different methods applied.”
The question this design seeks to answer is grounded in evidence: US data from 2000–2019 show a significant correlation between the six-month moving average of geomagnetic Kp indices and respiratory infection deaths (PMID 42350784); the matched-pair design aims to verify “how much of this association remains after meteorological factors are subtracted.” Meanwhile, the relationship between absolute humidity and influenza activity reverses across climate zones in China (PMID 35218485)—suggesting “meteorology” is not a constant that can be arbitrarily subtracted.
II. The Hardest Constraint: Methodology Itself Is a Predictor of Estimates
A 2018 systematic review including 103 influenza mortality burden studies conducted a meta-regression with a straightforward conclusion: modeling technique itself is a significant predictor of estimates. Using the multiplier method as a reference, Serfling-type models yield an annual age-standardized excess mortality rate approximately 40/100,000 higher on average (95%CI 3–77, P=0.033); within the same body of research, expanding the cause-of-death scope from “pneumonia and influenza” to “all-cause” adds another average 43/100,000 (95%CI 32–54) (PMID 28679157).
The implication for geomagnetic research is direct: if the systematic differences between three methods on the same dataset (on the order of tens per 100,000) exceed the geomagnetic marginal increment itself, then the conclusion of “geomagnetic significance” is largely determined by method selection rather than data.
III. Matched-Pair Design Does Not Automatically Eliminate Methodological Bias—It Splits It
The power of the matched-pair design lies in subtraction. However, the fate of methodological bias within this operation differs: one portion is common bias—offset when subtracted if the baseline shapes of both years are identical; the other is differential bias—which persists when the temporal distributions of deaths differ between years and directly contaminates the marginal increment.
The most typical differential bias is death displacement (harvesting), where a single disturbance prematurely removes individuals who would have died weeks later. An evaluation of global data from 2020–2024 by JAMA Netw Open shows displacement is a universal phenomenon, not an outlier (PMID 41609824); BMC methodological research provides individual-level counterfactual correction schemes (PMID 37853353). Reintroducing the matched-pair design: if death curves in geomagnetic active years are more “advanced,” Serfling baselines will underestimate them, Poisson models will partially absorb them, and the rate difference method depends on how the baseline period is defined—three methods react differently to the same displacement; this discrepancy is not a geomagnetic effect.
IV. Confidence Intervals Are Also Part of the Methodology
Beyond point estimates lie intervals. Directly applying Poisson assumptions to overly discrete count sequences yields confidence intervals that are too narrow—a 2026 simulation study in Epidemiol Infect quantified this: Poisson assumptions lead to “overconfidence” in effective reproduction number estimates, while negative binomial and quasi-Poisson intervals are significantly wider (PMID 42178993). The same holds true in the excess mortality domain: a sensitivity analysis in Popul Dev Rev found that altering only the baseline method can cause Lithuanian excess mortality estimates to differ by nearly 4-fold; changing only the reference period (2010–2019 vs. 2017–2019) causes Belgian estimates to differ by 20% (PMID 35600716). WHO methodological reports (PMID 36517599), more flexible alternative frameworks (PMID 35383642), and Bayesian spatiotemporal varying coefficient models (PMID 40661232) address the same issue: intervals are also model products.
V. What Should Be Done
First, pre-register methods and run all three concurrently, reporting the “method × latitude” interaction rather than selecting the most aesthetically pleasing one post-hoc.
Second, treat methodological differences themselves as a reportable effect size: report the ratio of “range of three-method estimates ÷ geomagnetic marginal increment.” If this ratio is far greater than 1, the conclusion is unreliable.
Third, do not use AIC to adjudicate methods—differences within 4 points should be considered equivalent; models should not be selected based solely on this metric (PMID 22095719); established practices exist for exposure-lag-response scenarios (PMID 20812303, PMID 29447338, PMID 23297754, PMID 35436963).
Fourth, stratify and evaluate timely: stratification by age and subtype can reveal displacement effects (PMID 30968823); AIC-driven model selection steps (PMID 26781954) and early-established timely monitoring methods (PMID 9209852) can be directly reused.
VI. Yijing Application Points
Numbers are never neutral. Under the same celestial timing, using different calendars or starting points yields different “qi counts.” The Su Wen states, “First determine the annual qi; do not violate the harmony of heaven and man”—one must first accurately define the year’s “annual qi,” and the act of “accurate definition” inherently includes “which algorithm to use.”
The distinguishing point of this paper: if the differences among three methods stem primarily from common bias, the matched-pair design remains credible and conclusions are robust; if differences arise mainly from differential bias (death displacement), the “geomagnetic marginal increment” must be stated alongside its method name—it is an “increment under Serfling” or an “increment under Poisson,” two distinct propositions.
Evidence and Limitations
Verified: Four excess mortality methods yield similar magnitudes but exhibit systematic differences on the same dataset (PMID 19453440); Serfling estimates are highly dependent on exclusion week rules and may produce negative estimates (PMID 24312225); modeling technique is a significant predictor of influenza mortality burden estimates (PMID 28679157); Poisson assumptions on over-discrete sequences lead to overly narrow intervals (PMID 42178993); altering baseline methods and reference periods can cause excess mortality estimates to differ by orders of magnitude (PMID 35600716).
Partially Verified: Global universality of death displacement (PMID 41609824, ecological summary level); feasibility of individual-level counterfactual corrections (PMID 37853353); performance of flexible excess mortality frameworks and Bayesian spatiotemporal varying coefficient models (PMID 35383642, PMID 40661232, primarily simulation studies).
To Be Verified: ① The decomposition of “methodological bias into common and differential bias within matched-pair designs” has never been directly quantified in any study to date, making it impossible to determine the proportion of differential bias within the geomagnetic marginal increment; ② The difference in estimates and coverage probability between three methods under the specific comparison of “geomagnetic active years vs. calm years” has never been compared; ③ If differential bias indeed dominates, whether a pre-specified method combination exists to suppress bias below the marginal increment remains an open question.
Related Reading
- Marginal Increment of Geomagnetic Activity on Respiratory Excess Mortality: Matched-Pair Year Design Under Meteorological Conditions
- Transplanting the North American Matched-Pair Year Design to East Asia and Europe: Can Geomagnetic Marginal Increments Intensify with Latitude?
- Monthly-Scale Geomagnetic Accumulation vs. Weekly Humidity Anomalies: Do Two Exposure Windows in the Same Model Eat Each Other Up?
- Environment-Life Frequency Interface Research Pool
Evidence Level: B
References: PMID 19453440, PMID 24312225, PMID 28679157, PMID 22284214, PMID 30968823, PMID 9209852, PMID 26781954, PMID 35600716, PMID 42178993, PMID 41609824, PMID 37853353, PMID 35383642, PMID 40661232, PMID 36517599, PMID 22095719, PMID 20812303, PMID 29447338, PMID 23297754, PMID 35436963, PMID 42350784, PMID 35218485
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.
Leave a Reply