Evidence level: B
把全国多疾病长序列合并做”归年后1-3年”检验,看起来数据量更大、结论更稳;但如果不同气候区的滞后结构本来就不同,合并恰恰会把区域信号互相抵消——这不是统计细节,而是”归年到底成不成立”的成败关键。
正文
上一篇《”归年”后第1/2/3年》提出了逐病种检验的方法,但留下一个悬而未决的问题:检验到底该用全国合并数据,还是该按气候区分层? 直觉上”数据越多越可信”,所以很多人默认用全国合并序列。统计学会告诉你,这个直觉在异质性存在时是反的。
一、合并数据会”撒谎”:Simpson 悖论
Simpson 悖论至今仍是最容易被忽视的陷阱。一项肺癌 EGFR 突变研究提供了真实案例:把数据按种族分层(partition)与不分层直接合并(aggregation)做单变量分析,会得出互相矛盾的结论(PMID 22246415, Fu et al. 2012)。分层时每个子群里关联成立,合并后方向却翻转——原因就是种族这一”第三变量”在两次比较里的权重不同。映射到归年检验:如果温带区归年后第 2 年发病上升、亚热带区同期下降,全国合并后两者相加可能恰好抵消,报告出”归年无效应”的假阴性。合并得到的”无显著”结论,可能只是区域信号互相抵消的假象。
二、聚合不只掩盖,还会放大偏倚
更麻烦的是,聚合(aggregation)本身会放大原有偏倚,而不只是让效应消失。Webster 2007 的方法学研究表明:当存在”按组混杂”和”按组效应修饰”时,聚合导致的暴露方差下降会把个体层面的偏倚放大到组水平(PMID 17615079)。也就是说,全国合并不仅可能把真信号冲淡,还可能把一个原本微弱的方向性偏差放大成显著假象——两个方向的错误同时存在。重症医学的 meta 分析综述也明确把 Simpson 悖论与”生态谬误”(ecological fallacy)并列为合并研究最常见的两类错误(PMID 18671838)。
三、气候区:一个有理论依据的分层变量
幸运的是,我们不需要盲选分层变量——气候区是有现代实证支撑的。一项覆盖 94 国、128 万孕妇的弓形虫血清流行率 meta 分析,用随机效应 meta 回归发现:纬度降低、气温升高时感染率系统性上升,且不同气候带的最佳干预策略本就不同(PMID 34329056, Rostami et al. 2021)。另一项隐孢子虫季节性 meta 分析更直接:按 Köppen 气候分区(热带/干旱/温带/寒冷)分层后,驱动因子完全不同——湿热带里降水是主驱动,中纬温带里温度才是主驱动,且滞后关系跨气候带持续存在(PMID 19328462)。这说明”气候区×疾病×滞后”的三重交互是真实现象(与源文引用的 PMID 36141780 中国四大气候区腹泻研究一致),用全国合并会把这个三重交互抹平成”平均效应”。
四、分层不是万能药:亚组功效陷阱
但”分层就一定更好”也是错的。meta 分析亚组研究的方法学警告指出:检测亚组间差异所需的研究数是主分析的 3-4 倍,效应越小、亚组越不均匀,所需数量呈指数上升(PMID 34852862, Cuijpers et al. 2021)。中国四大气候区里高原区站点本就稀少,若强行分层,每层内样本可能不足以支撑检验功效——分层过度会从”消解异质性”变成”切碎数据”。而且研究层面的 meta 回归并不稳健:一项 NSCLC 的对比研究发现,用汇总数据的 meta 回归与个体数据(IPD)回归可能得到方向相反的结论(PMID 15640501)。所以分层必须配套”气候区×疾病”的交互检验,而非简单分组报告。
五、可测方案与三种结局
具体设计:同一批多种气候敏感传染病长序列(流感、RSV、痢疾、伤寒、疟疾、登革热),做双通道——
1. 合并通道:全国合并序列 → 短滞后 DLNM vs 短+1/2/3年滞后 DLNM → ΔAUC/互信息;
2. 分层通道:按温带/亚热带/热带/高原分层各自重跑,再做气候区 meta 回归,报告 I² 下降幅度与气候区调节效应的显著性。
三种结局对应三种解读:若分层后各气候区滞后增量同向且显著、I² 低 → 归年是全国普适信号,合并即够;若分层后各区增量异号、I² 高、气候区是显著调节变量 → 全国合并确实把异质性平均掉了,归年必须按气候区分区表述;若分层后各层均不显著、I² 也低 → 归年在全国范围内无独立增量,是文化叙事。
这里还有一个时间维度的提醒:Ballester 2024 覆盖欧洲 16 国、4 亿人的研究发现,时间聚合本身就会系统性低估效应——周数据把冷相关死亡低估 8.56%、热相关低估 21.56%(PMID 38188278)。无论空间合并还是时间聚合,聚合都会改变估计值,必须显式检验,而不是默认接受。
证据与局限
已验证: Simpson 悖论——分层与合并可得相反结论(PMID 22246415);聚合会放大按组混杂/效应修饰导致的偏倚(PMID 17615079);Simpson 悖论与生态谬误是合并研究的典型错误(PMID 18671838);气候带分层后疾病驱动因子与滞后结构显著不同(PMID 34329056 纬度/气温梯度,PMID 19328462 Köppen 分区);时间聚合系统性低估效应(PMID 38188278)。
部分验证: 气候区是”有依据的分层变量”——弓形虫、隐孢子虫研究支持,但针对中国”归年后1-3年滞后”本身的跨气候区实证尚未完成;meta 回归与 IPD 回归方向可能冲突(PMID 15640501)提示分层结论需谨慎。
待验证: 全国合并 vs 分气候区归年检验的 ΔAUC/互信息差异幅度;气候区作为调节变量的 I² 解释比例;亚组分层在中国四大气候区下的实际功效(高原区站点是否够用,PMID 34852862 的功效约束);空间合并与时间聚合双重作用下的联合偏倚方向。
相关阅读
证据等级: B
参考文献: PMID 22246415, PMID 17615079, PMID 18671838, PMID 34329056, PMID 19328462, PMID 34852862, PMID 15640501, PMID 38188278
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。
National Merging vs. Climate Zones: Will the \”Heterogeneity\” Detected by Anomaly Years Be Averaged Out?
Evidence level: B
Merging multi-disease long-term series across the nation for a “return-year +1 to +3 years” validation appears to yield larger sample sizes and more stable conclusions. However, if the lag structures inherently differ across climate zones, aggregation actively cancels out regional signals rather than reinforcing them. This is not merely a statistical detail; it is the decisive factor determining whether the “return-year” hypothesis holds validity.
Main Text
The previous article, “Return-Year +1/+2/+3 Years,” proposed a disease-specific validation approach but left an unresolved question: Should validation utilize aggregated national data or stratify by climate zone? Intuitively, “more data equals more credibility,” leading many to default to national aggregation. Statistical methodology warns that this intuition is inverted when heterogeneity exists.
I. Aggregated Data Can “Lie”: The Simpson Paradox
The Simpson paradox remains the most frequently overlooked trap. A real-world case study on lung cancer EGFR mutations illustrates this: performing univariate analysis by partitioning data by race versus directly aggregating without stratification yields contradictory conclusions (PMID 22246415, Fu et al. 2012). Stratification confirms associations within subgroups, yet aggregation reverses the direction. This occurs because race acts as a “third variable” with differing weights in the two comparisons. Mapping this to return-year validation: if incidence rises in the second year post-return in temperate zones but declines simultaneously in subtropical zones, national aggregation may result in perfect cancellation, reporting a “null return-year effect” false negative. The “non-significant” conclusion derived from aggregation may merely be an illusion of mutually cancelling regional signals.
II. Aggregation Amplifies Bias, Not Just Masks It
More problematic is that aggregation itself can amplify existing bias rather than simply erasing effects. Webster’s 2007 methodological study demonstrates that when “confounding by subgroup” and “effect modification by subgroup” are present, the reduction in exposure variance caused by aggregation can inflate individual-level bias to the group level (PMID 17615079). In other words, national aggregation not only dilutes true signals but can also amplify a previously weak directional bias into a significant false positive—potentially involving errors in both directions. A meta-analysis review in critical care explicitly lists the Simpson paradox and “ecological fallacy” as the two most common errors in aggregated studies (PMID 18671838).
III. Climate Zones: A Stratification Variable with Theoretical Basis
Fortunately, we do not need to arbitrarily select stratification variables; climate zones are supported by modern empirical evidence. A meta-analysis covering 94 countries and 1.28 million pregnant women found using random-effects meta-regression that infection rates systematically rise with decreasing latitude and increasing temperature, and optimal intervention strategies differ fundamentally across distinct climate belts (PMID 34329056, Rostami et al. 2021). Another seasonal cryptosporidium meta-analysis directly showed that stratifying by Köppen climate classification (tropical/arid/temperate/cold) reveals completely different driving factors: precipitation is the primary driver in wet tropics, while temperature dominates in mid-latitude temperate zones, with lag relationships persisting across climate bands (PMID 19328462). This confirms that the triple interaction of “Climate Zone × Disease × Lag” is a genuine phenomenon (consistent with the Chinese four-climate-zone diarrhea study cited as PMID 36141780). National aggregation flattens this triple interaction into a “mean effect.”
IV. Stratification Is Not a Panacea: The Subgroup Power Trap
However, “stratification is always better” is also incorrect. Methodological warnings regarding subgroup analysis in meta-regression indicate that the number of studies required to detect differences between subgroups is 3-4 times that of the main analysis. Smaller effects and more heterogeneous subgroups lead to exponential increases in required sample size (PMID 34852862, Cuijpers et al. 2021). Within China’s four major climate zones, the plateau region has inherently sparse stations; forcing stratification may result in insufficient sample sizes per layer to support statistical power—excessive stratification transforms “dissolving heterogeneity” into “chopping up data.” Furthermore, meta-regression at the study level is not robust: a comparative study on NSCLC found that meta-regression using aggregated data versus individual patient data (IPD) regression could yield opposite conclusions (PMID 15640501). Therefore, stratification must be paired with interaction testing (“Climate Zone × Disease”) rather than simple subgroup reporting.
V. Measurable Protocols and Three Possible Outcomes
Specific design: Using the same long-term series of multiple climate-sensitive infectious diseases (influenza, RSV, dysentery, typhoid, malaria, dengue), implement a dual-channel approach:
1. Aggregation Channel: National aggregated series → Short-lag DLNM vs. Short +1/2/3-year lag DLNM → ΔAUC/Mutual Information;
2. Stratification Channel: Re-run separately for Temperate/Subtropical/Tropical/Plateau strata, then perform climate-zone meta-regression, reporting I² reduction magnitude and significance of climate-zone moderation effects.
Three outcomes correspond to three interpretations:
* If stratified increments are unidirectional and significant across zones with low I² → Return-year is a universal national signal; aggregation suffices.
* If stratified increments are bidirectional, I² is high, and climate zone is a significant moderator → National aggregation indeed averaged out heterogeneity; return-year must be reported stratified by climate zone.
* If none of the strata show significance despite low I² → Return-year has no independent incremental effect nationally; it is a cultural narrative.
A temporal reminder is also warranted: Ballester’s 2024 study covering 16 European countries and 400 million people found that temporal aggregation itself systematically underestimates effects—cooling-related mortality was underestimated by 8.56% and heat-related by 21.56% when using weekly data (PMID 38188278). Whether through spatial aggregation or temporal aggregation, aggregation alters estimates; it must be explicitly tested rather than assumed.
Evidence and Limitations
Verified: Simpson paradox—stratification and aggregation yield opposite conclusions (PMID 22246415); aggregation amplifies bias from confounding by subgroup/effect modification (PMID 17615079); Simpson paradox and ecological fallacy are typical errors in aggregated studies (PMID 18671838); disease driving factors and lag structures differ significantly after climate-zone stratification (PMID 34329056 latitude/temperature gradient, PMID 19328462 Köppen classification); temporal aggregation systematically underestimates effects (PMID 38188278).
Partially Verified: Climate zones are an “evidence-based stratification variable”—supported by toxoplasmosis and cryptosporidium studies, but cross-climate-zone empirical validation for China’s specific “return-year +1 to +3 year lag” itself remains incomplete; potential conflict between meta-regression and IPD regression directions (PMID 15640501) suggests caution in stratified conclusions.
To Be Verified: Magnitude of ΔAUC/mutual information differences between national aggregation and stratified climate-zone return-year validation; proportion of I² explained by climate zones as a moderator; actual statistical power of subgroup stratification within China’s four major climate zones (whether plateau stations suffice, constrained by PMID 34852862); joint bias direction under combined spatial aggregation and temporal aggregation.
Related Reading
- “Return-Year” +1/+2/+3 Years: Can Multi-Disease Incidence Increments Validate “Rigid and Soft Loss of Control”?
- Climate Zone Attribution Verdict on Heterogeneity of Phase Drift in Luck Cycle Periodicity](/research/science-yixue/yunqi-phase-drift-heterogeneity-climate-zone-attribution)
- Lag Scale of “Three-Year Epidemic Control”: Systematic Validation Across Multiple Diseases](/research/science-yixue/yunqi-three-year-epidemic-lag-multidisease-validation)
- Research Pool ②: Spatiotemporal Frequency Topology](/research/pools/pool-2-yunqi-spatiotemporal-frequency-topology)
Evidence Level: B
References: PMID 22246415, PMID 17615079, PMID 18671838, PMID 34329056, PMID 19328462, PMID 34852862, PMID 15640501, PMID 38188278
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