Evidence level: B
五运六气”三年化疫”的本质是一个长滞后假说:1-3 年前的气候信息是否携带 0-3 周短滞后气象模型捕捉不到的独立预测力?现代数据给出了值得正视的证据,但”独立性”本身还没有被严格检验过。
正文
五运六气最迷人的地方,也是它最容易被质疑的地方,是”三年化疫”式的长滞后主张:今年岁运太过、某气失常,瘟疫可能在一到三年后才爆发。现代流行病学早就承认气候对疾病有滞后效应,但主流建模默认滞后窗口很短——分布式滞后非线性模型(DLNM)通常只取 0-3 周。于是问题来了:古代理论盯住的那 1-3 年长滞后,到底有没有独立于短滞后气象的信息量? 这不是一个哲学问题,而是可以被现代数据直接检验的问题。
先看最有力的正面证据。香港一项用 22 年结核监测数据、DLNM 最大滞后拉到 20 个月的研究发现:月平均气温 16.8°C 时,滞后 14 个月的结核报告风险达到峰值 RR 1.18(95%CI 1.02-1.35);相对湿度的影响从滞后 9-11 个月一路延伸到 12-17 个月,峰值 RR 1.06(PMID 33436107)。结核的潜伏感染期和诊断延迟天然决定了它需要以月、而不是以天为单位的滞后窗口——这正是”长滞后关联真实存在”最干净的现代实例。
其次是方法学上的突破。南非一项疟疾预测研究在同一个 DLNM 框架里,把降水拆成了两个独立项:短滞后项(当前周至前 15 周)和长滞后项(至前 51 周、约一年),分别用自然样条建模,明确假设”短滞后和长滞后效应可能有独立作用”(PMID 31784563)。这套”长短滞后分离建模”正是检验运气长滞后信息量所需要的操作框架——它把”长滞后有没有用”从一个模糊主张,变成了”长滞后项能否显著改善模型”的可检验问题。
但证据并不是一边倒的。经典的登革热-ENSO 研究用 20 多年数据做小波分析,结论相当泼冷水:在波多黎各,ENSO 与登革的多年尺度关联只是”短暂出现”;在泰国,看似显著的多尺度关联”很可能是伪关联”;在墨西哥则完全观察不到(PMID 19918363)。而另一项东南亚研究在 2-4 年频带看到了登革与 ENSO 的显著交叉相干(PMID 25253208),美国的非霍乱弧菌病也被发现与大尺度海洋现象存在跨尺度滞后关联(PMID 31364581)。合起来看:多年尺度气候-疾病关联在某些地区、某些时期显著,另一些地方则完全消失——这种”时有时无”恰恰提示它可能脆弱、非平稳,需要格外严格的统计检验。
那么”独立性”该怎么检验?关键区分两种可能:其一,长滞后只是短滞后信息的”换皮”——宏观气候周期和短期天气高度相关,长滞后项捡到的是短期天气早已编码的信息,这叫”伪独立”;其二,长滞后确实携带额外信息,比如经年累月的免疫、营养或媒介生态变化(PMID 35523336 就展示了结核住院风险里气象、空气污染物与流感各自的独立效应)。区分两者的标准做法是:在同一数据集上分别拟合”短滞后 DLNM”和”短+长滞后 DLNM”,比较两者的时间外预测力增量(ΔAUC、互信息或 Brier 分解),并做置换检验确认增量不是偶然。如果长滞后项不能带来显著且可复现的增量,那”三年化疫”就只能算文化叙事;如果能够,五运六气的长滞后主张就第一次在概率预测框架内拿到了硬证据。
这里还藏着一个对五运六气特别友好的细节:干支是固定历法标签,用它分组不会像很多机器学习特征那样泄露未来信息,所以”长滞后信息量检验”在方法学上是干净的——不需要担心时间穿越。而运气理论本身提供的先验(岁运太过不及、客气加临),恰好可以当作”应该去哪几个滞后窗口找信号”的导航,而不是把 0-60 个月全部盲目扫一遍——那样做必然撞上多重比较的假阳性墙。换句话说,古人经验在这里的角色是”假设生成器”,现代数据负责”假设检验”,两者各司其职,这正是科学易学桥接的标准姿势。
对中国研究者,这个检验的门槛很低:一省几十年的法定传染病日报 + 对应气象日报 + 一套干支历表即可启动,难点全在协议纪律——统一长短滞后窗口、预先登记模型清单、做时间外验证而非训练集自夸。无论结果如何,产出的都是一个可复用的”长滞后信息量检验”模板,而这正是当前运气医学研究中最稀缺的资产。
证据与局限
已验证: 长滞后气象-疾病关联真实存在,香港结核研究最大滞后 20 个月、峰值滞后 14 个月 RR 1.18(PMID 33436107);南非疟疾 DLNM 已实现”短滞后 vs 长滞后”分离建模(PMID 31784563);气象、污染物、流感对结核住院有独立效应(PMID 35523336)。
部分验证: 多年尺度(2-4 年)气候-疾病关联在东南亚登革-ENSO 上显著(PMID 25253208)、在美非霍乱弧菌病上可测(PMID 31364581),但波多黎各/泰国/墨西哥数据显示其非平稳、可能伪关联(PMID 19918363)。
待验证: “长短滞后分离模型”在同一数据集上的时间外增量检验(ΔAUC/互信息/Brier 分解+置换检验)——长滞后气象信息是否携带独立于短滞后的可复现预测力,全球尚无直接实验。
相关阅读
证据等级: B
参考文献: PMID 33436107, PMID 31784563, PMID 35523336, PMID 25253208, PMID 31364581, PMID 19918363
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。
Long Lag (1-3 Years) Weather-Disease Associations: Do They Carry Information Independent of Short Lags?
Evidence level: B
The essence of the “Three-Year Epidemic” hypothesis in Five Movements and Six Qi (Wuyun Liqi) is a long-lag proposition: Do climatic conditions from 1–3 years prior carry predictive power that short-lag meteorological models (typically 0–3 weeks) fail to capture? Modern data provides evidence worthy of serious consideration, yet the “independence” of this long-lag signal remains unverified through rigorous statistical testing.
Main Text
The most captivating aspect of Yijing Studies, and simultaneously its most vulnerable point to skepticism, is the claim of long-lag effects akin to “Three-Year Epidemics”: if the annual cycle is excessive or a Qi is abnormal, an epidemic might erupt one to three years later. Modern epidemiology long ago acknowledged the lagged effects of climate on disease, but mainstream modeling defaults to short windows—Distributed Lag Non-Linear Models (DLNM) typically capture only 0–3 weeks. Thus, the question arises: Does the long-lag window (1–3 years) targeted by ancient theory contain information independent of short-lag meteorological signals? This is not merely a philosophical inquiry; it is a question directly testable via modern data.
First, consider the strongest positive evidence. A Hong Kong study utilizing 22 years of tuberculosis monitoring data and extending the DLNM maximum lag to 20 months found that at a mean monthly temperature of 16.8°C, tuberculosis reporting risk peaked at a lag of 14 months (RR 1.18, 95%CI 1.02–1.35); similarly, the influence of relative humidity extended from a lag of 9–11 months to 12–17 months, peaking at RR 1.06 (PMID 33436107). The natural incubation period and diagnostic delay of tuberculosis necessitate lag windows measured in months rather than days—this represents one of the cleanest modern instances confirming the reality of “long-lag associations.”
Second, there is a methodological breakthrough. A South African malaria prediction study within the same DLNM framework decomposed precipitation into two independent terms: a short-lag term (current week to 15 weeks prior) and a long-lag term (up to 51 weeks prior, approximately one year), modeled separately using natural splines. They explicitly hypothesized that “short-lag and long-lag effects may act independently” (PMID 31784563). This “separate modeling of short and long lags” constitutes the operational framework required to test the information content of long-lag effects in Yijing theory—it transforms the vague assertion “is long-lag useful?” into the testable question “does the inclusion of a long-lag term significantly improve model performance?”
However, the evidence is not unanimous. Classic Dengue-ENSO studies using over 20 years of data via wavelet analysis yielded sobering conclusions: in Puerto Rico, the multi-year association between ENSO and Dengue appeared only “transiently”; in Thailand, seemingly significant multi-scale associations were likely “spurious”; and in Mexico, no association was observed at all (PMID 19918363). Conversely, another Southeast Asian study identified significant cross-coherence between Dengue and ENSO within the 2–4 year band (PMID 25253208), and non-cholera Vibrio diseases in the US were found to exhibit cross-scale lagged associations with large-scale oceanic phenomena (PMID 31364581). Taken together: Multi-year scale climate-disease associations are significant in some regions and periods but vanish in others—this “intermittent” nature suggests fragility and non-stationarity, demanding exceptionally rigorous statistical validation.
So, how should “independence” be tested? The key distinction lies between two possibilities: First, long-lag is merely a “re-skinned” version of short-lag information—macro-climatic cycles are highly correlated with short-term weather, meaning the long-lag term captures information already encoded in short-term weather; this is “pseudo-independence.” Second, long-lag genuinely carries additional information, such as cumulative changes in immunity, nutrition, or vector ecology over years (PMID 35523336 demonstrates independent effects of meteorology, air pollutants, and influenza on tuberculosis hospitalization). The standard procedure to distinguish these is to fit both a “short-lag DLNM” and a “short+long-lag DLNM” on the same dataset, compare their out-of-sample predictive increments (ΔAUC, mutual information, or Brier decomposition), and perform permutation tests to confirm the increment is not accidental. If the long-lag term fails to yield significant and reproducible increments, the “Three-Year Epidemic” remains a cultural narrative; if it succeeds, Yijing’s long-lag claims receive hard evidence within a probabilistic prediction framework.
Here lies a detail particularly friendly to Yijing theory: the Heavenly Stems and Earthly Branches are fixed calendrical labels. Grouping by them does not leak future information like many machine learning features, making the “long-lag information test” methodologically clean—no need to worry about time travel. Furthermore, the priors provided by Yijing theory itself (excessive/deficient annual cycles, visiting Qi) can serve as a navigation map indicating “which lag windows to search for signals,” rather than blindly scanning all 0–60 months, which would inevitably collide with the wall of multiple comparison false positives. In other words, ancient experience here functions as a “hypothesis generator,” while modern data performs “hypothesis testing.” This division of labor is the standard posture for bridging science and Yijing Studies.
For Chinese researchers, the threshold for this test is low: one province’s decades of statutory infectious disease daily reports plus corresponding meteorological daily reports and a set of Heavenly Stems and Earthly Branchs tables suffice to initiate the study. The difficulty lies entirely in protocol discipline—standardizing short and long lag windows, pre-registering model lists, and conducting out-of-sample validation rather than boasting on training set performance. Regardless of the outcome, the output is a reusable “long-lag information test” template, which is precisely the most scarce asset in current Yijing medical research.
Evidence and Limitations
Verified: Long-lag meteorological-disease associations are real; Hong Kong tuberculosis study shows maximum lag of 20 months with peak at 14 months (RR 1.18) (PMID 33436107); South African malaria DLNM has implemented “short-lag vs. long-lag” separation modeling (PMID 31784563); meteorology, pollutants, and influenza have independent effects on tuberculosis hospitalization (PMID 35523336).
Partially Verified: Multi-year scale (2–4 years) climate-disease associations are significant in Southeast Asian Dengue-ENSO cases (PMID 25253208) and measurable in US non-cholera Vibrio cases (PMID 31364581), but Puerto Rico/Thailand/Mexico data indicates non-stationarity and potential spuriousness (PMID 19918363).
Pending Verification: Out-of-sample incremental testing of “short and long lag separation models” on the same dataset (ΔAUC/mutual information/Brier decomposition + permutation test)—whether long-lag meteorological information carries reproducible predictive power independent of short-lags. No direct global experiment exists yet.
Related Reading
- Five Movements Six Qi + Modern Statistical Models: Synergy or Interference?
- Pure DLNM vs. Yijing Screening + DLNM: Head-to-Head Comparison of Three Models
- Yijing Long-Lag Multi-Disease Independent Information Validation
- Research Pool: Spatiotemporal Frequency Topology
Evidence Level: B
References: PMID 33436107, PMID 31784563, PMID 35523336, PMID 25253208, PMID 31364581, PMID 19918363
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