Evidence level: B
古人经验的价值不在于”帮我们验证”,而在于”帮我们少找几个地方”。但这套导航到底更灵,还是只是把真信号挡在门外——可以被同一个数据集直接裁决。
正文
上一篇把”运气长滞后是否携带独立增量”拆成了可执行实验。但它留下了一个更底层的工程问题:你凭什么决定去看哪几个滞后窗口? 有两种截然不同的策略。第一种叫”先验导航”——按五运六气的岁运太过不及、客气加临,先推出”这几年该在哪些滞后窗里找”。第二种叫”全滞后盲扫”——不问理论,把 0 到 60 个月所有窗口全跑一遍。哪一种的发现效率更高?这不是信仰问题,是可以拿同一份数据直接比的问题。
先看盲扫要付的代价。把滞后窗口一路拉到月年级,等于同时做了几十次假设检验。若照 α=0.05 的裸阈值判定,61 个窗口里纯靠运气就会冒出约 3 个”显著”结果。环境流行病学的标准解法是把滞后-反应关系放进分布式滞后非线性模型(DLNM),用交叉基一次性建模整个曲线,而不是逐窗口硬扫(PMID 20812303);对时空数据则用错误发现率(FDR)控制”发现里有多少是假的”——Ventrucci 等人为标化死亡比设计的就是一套贝叶斯层次 FDR 估计,用后验概率决定哪些区域”真高”(PMID 20577014)。盲扫不是错,但每多扫一个窗口,就多缴一份假阳性税。
先验导航的诱惑正在于此:只测理论点名的少数窗口,检验次数骤减,每个窗口的功效自然抬高。但这里埋着一个致命前提。Cai 等人在全基因组假设检验里把这件事算清楚了:先验分两种——”神谕先验”知道所有真信号都落在限定子集内,那限制窗口确实省力;但一旦先验漏掉哪怕一小部分真信号,非神谕先验的表现可以比完全不用先验还差(PMID 37101120)。而且他们发现,把检验从全部缩到 100-1000 个候选基因带来的功效增益,往往还比不上队列规模每年 20-40% 的自然增长。换句话说:先验限制是把双刃剑——只有”不漏”才划算,漏了就是自残。 运气先验属于哪一类,至今没人量过。
把知识塞进变量选择,方法学界已有成熟但代价明确的先例。Boulet 等人把临床专家的数值权重经”幂先验”并入贝叶斯变量选择,好处是能挑出纯数据驱动会漏掉的稀有变量,坏处是也会把”数据看相关、专家觉得无关”的变量剔掉(PMID 30963815)。Hill 等人则用经验贝叶斯自动加权路径/网络先验,核心动机就是防止先验设定错误反噬(PMID 22578440)。这两篇共同指向一个教训:先验能用,但必须给”先验错了”留退路。
退路其实已经有现成工具,而且比”硬先验”更聪明。Chen 等人提出”数据自适应收缩”:把无约束 DLM(无偏但低效)往有约束 DLM(高效但可能有偏)方向收缩,让数据自己决定收缩多少、在哪几段收缩,从而在偏差与方差之间取平衡(PMID 29040386)。Antonelli 等人的 spike-and-slab 模型更进一步——把关键窗口的选择交给数据:哪些滞后时段真的关联到结局,由后验包含概率挑出来,而不是由理论预先钉死(PMID 36073640)。这相当于把”先验导航”从一把硬尺子,变成一根可弯的软绳。
于是实验设计就清楚了:同一份多年传染病+气象长序列,跑两条通道。 通道 A——先验导航:严格按运气理论给出的窗口建模。通道 B——全滞后盲扫:0-60 个月全窗口 + BH-FDR 校正。真值来自”真信号标定模拟”:在已知的滞后位置注入已知强度的、与干支分组同相位的效应,这样才知道每个通道到底抓到了几个真的、放跑了几个。核心指标有三个:召回率(真信号抓到几成)、假阳性率、以及最关键的——“神谕分数” f,即被注入的真信号里有多大比例落在先验窗口之内。 这个 f 就是判定运气先验是”神谕”还是”非神谕”的那条线(对应 PMID 37101120 的 f 临界判据)。
三种结局,各自对应运气理论的一种读法。其一,通道 A 召回更高、假阳性更低——先验确实是有信息量的假设生成器,古人的窗口经验被现代数据认账。其二,通道 B 反超——说明运气推导出的窗口与真实效应窗口错位,先验在这里只是多重比较的负担,”古人经验当导航”这一主张当场被否。其三,两者打平——那先验的价值就不在”缩窗口”上,而在 Cai 等人说的另一类事情上:因果结构与通路假设,那本就不是标准假设检验擅长的地方(PMID 37101120)。
无论哪种结局,产出的都不是一句”运气准不准”,而是一个可复用的“神谕分数”标定模板——一个数字,回答”这套古代先验在漏信号这件事上到底有多可靠”。这恰好把”古人经验”重新摆回它该站的位置:它是假设生成器,不是验证机器;它负责提出”该去哪看”,现代数据负责回答”到底有没有”。这,才是科学易学桥接里两者各司其职的标准姿势。
证据与局限
已验证: 逐窗口扫描会累积多重比较假阳性,DLNM 交叉基可整体建模滞后曲线而不必逐窗硬扫(PMID 20812303);时空环境数据的真假发现可用贝叶斯层次 FDR 定量控制(PMID 20577014);先验限制只有在”不漏真信号”时才提升功效,非神谕先验可劣于无先验(PMID 37101120)。
部分验证: 专家/知识先验经幂先验、经验贝叶斯并入变量选择,能捞出数据会漏的稀有变量,但存在先验误设反噬风险(PMID 30963815、PMID 22578440);数据自适应收缩与 spike-and-slab 可把”窗口选择”交给数据,作为硬先验的软替代(PMID 29040386、PMID 36073640)。
待验证: 在真实多年传染病-气象序列上,”运气先验导航窗口 vs 0-60 个月全窗口盲扫”的头对头功效/假阳性对比,以及运气先验的”神谕分数”实测标定——全球尚无直接实验。
相关阅读
证据等级: B
参考文献: PMID 37101120, PMID 20577014, PMID 20812303, PMID 30963815, PMID 22578440, PMID 29040386, PMID 36073640
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。
\”Prior Navigating Luck\” vs \”Full Lag Blind Sweeping\”: Can Ancient Experience Serve as a Hypothesis Generator?
Evidence level: B
The value of ancient experience lies not in “verifying for us,” but in “helping us search fewer places.” Yet, is this navigation genuinely superior, or does it merely block true signals? This can be adjudicated directly using the same dataset.
Main Text
The previous article decomposed the question of whether “long-term luck carries independent incremental information” into an executable experiment. However, it left a more fundamental engineering question: On what basis do you decide which lag windows to examine? Two distinct strategies exist. The first is “Prior Navigational Guidance”—deducing which lag windows to search based on the excess or deficiency of the annual cycle, and the interaction of the guiding qi (Ke Qi). The second is “Full-Lag Blind Sweeping”—running all windows from 0 to 60 months without theoretical constraints. Which yields higher discovery efficiency? This is not a matter of faith, but one that can be directly compared using the same dataset.
First, consider the cost of blind sweeping. Extending lag windows to the annual level simultaneously performs dozens of hypothesis tests. If a raw threshold of $\alpha=0.05$ is applied, approximately 3 “significant” results would emerge purely by chance among the 61 windows. The standard solution in environmental epidemiology is to incorporate lag-response relationships into a Distributed Lag Nonlinear Model (DLNM) using cross-basis functions to model the entire curve at once, rather than hard-sweeping window by window (PMID 20812303); for spatiotemporal data, the False Discovery Rate (FDR) controls the proportion of false discoveries—Ventrucci et al. designed a Bayesian hierarchical FDR estimator specifically for standardized mortality ratios, using posterior probabilities to determine which regions are “truly elevated” (PMID 20577014). Blind sweeping is not incorrect, but each additional window incurs a tax of false positives.
The allure of prior navigation lies precisely here: by testing only the few windows named by theory, the number of tests drops drastically, naturally increasing power per window. However, this relies on a fatal premise. Cai et al. clarified this in the context of whole-genome hypothesis testing: Priors come in two types—”Oracle priors” know all true signals fall within a restricted subset, making window restriction efficient; but once the prior misses even a small portion of true signals, non-oracle priors can perform worse than using no prior at all (PMID 37101120). Furthermore, they found that the power gain from restricting tests from all candidates to 100–1000 candidate genes often falls short of the natural annual growth in cohort size of 20–40%. In other words: Prior restriction is a double-edged sword—it only pays off if it does not miss signals; missing them is self-harm. It remains unknown whether “luck priors” fall into the oracle or non-oracle category.
Embedding knowledge into variable selection has established precedents in the methods literature with clear costs. Boulet et al. incorporated clinical expert numerical weights into Bayesian variable selection via “power priors.” The benefit is the ability to identify rare variables missed by pure data-driven approaches; the drawback is the exclusion of variables where “data show correlation but experts deem irrelevant” (PMID 30963815). Hill et al. used empirical Bayes to automatically weight path/network priors, with a core motivation to prevent the backlash of erroneous prior specification (PMID 22578440). Together, these point to a lesson: priors can be used, but a retreat path must be provided for “wrong priors.”
Such a retreat path already exists with ready-made tools, and is arguably smarter than “hard priors.” Chen et al. proposed “data-adaptive shrinkage”: shrinking an unconstrained DLM (unbiased but inefficient) toward a constrained DLM (efficient but potentially biased), allowing data to decide the degree and location of shrinkage, thus balancing bias and variance (PMID 29040386). Antonelli et al.’s spike-and-slab model goes further—delegating window selection to the data: which lag periods truly associate with the outcome are selected by posterior inclusion probabilities, not pre-fixed by theory (PMID 36073640). This transforms “prior navigation” from a rigid ruler into a flexible soft rope.
Thus, the experimental design is clear: Run two channels on the same multi-year infectious disease and meteorological time series. Channel A—Prior Navigation: Model strictly according to windows specified by luck theory. Channel B—Full-Lag Blind Sweeping: All windows from 0–60 months + BH-FDR correction. Truth comes from “True Signal Calibration Simulation”: injecting known-strength effects at known lag positions with phases aligned to Heavenly Stems and Earthly Branches, thereby determining how many true signals each channel captures and how many it misses. Three core metrics are required: Recall (proportion of true signals captured), False Positive Rate, and most critically—the “Oracle Score” f, representing the proportion of injected true signals falling within the prior windows. This f serves as the criterion line for judging whether luck priors are “oracle” or “non-oracle” (corresponding to the f critical criterion in PMID 37101120).
Three outcomes correspond to three readings of luck theory. First, if Channel A shows higher recall and lower false positives—the prior is indeed an informative hypothesis generator, and ancient window experience is validated by modern data. Second, if Channel B surpasses it—indicates a misalignment between derived windows and true effect windows; here, the prior acts merely as a burden for multiple comparisons, and the claim that “ancient experience serves as navigation” is refuted. Third, if they are tied—the value of the prior lies not in “shrinking windows” but in another category described by Cai et al.: causal structure and pathway assumptions, which is not the standard domain of hypothesis testing (PMID 37101120).
Regardless of the outcome, the product is not a statement on “how accurate luck is,” but a reusable “Oracle Score” calibration template—a number answering “how reliable this ancient prior is at missing signals.” This precisely repositions “ancient experience” to its proper role: it is a hypothesis generator, not a verification machine; it is responsible for saying “where to look,” while modern data answers “whether there is anything there.” This is the standard posture for both parties to fulfill their roles in bridging science and Yijing Studies.
Evidence and Limitations
Verified: Window-by-window scanning accumulates multiple comparison false positives; DLNM cross-basis functions can model lag curves globally without hard-sweeping windows (PMID 20812303); Bayesian hierarchical FDR can quantitatively control true/false discovery in spatiotemporal environmental data (PMID 20577014); Prior restriction only boosts power if “no true signals are missed,” and non-oracle priors can underperform no-prior approaches (PMID 37101120).
Partially Verified: Expert/knowledge priors incorporated via power priors or empirical Bayes into variable selection can recover rare variables missed by data, but carry risks of prior misspecification backlash (PMID 30963815, PMID 22578440); Data-adaptive shrinkage and spike-and-slab models can delegate “window selection” to data as a soft alternative to hard priors (PMID 29040386, PMID 36073640).
To Be Verified: Head-to-head comparisons of efficacy and false positive rates between “luck prior navigation windows vs. 0–60 month full-blind sweeping” on real multi-year infectious disease-meteorological series, and empirical calibration of the “Oracle Score” for luck priors—no direct experiments exist globally yet.
Related Reading
- Validation of “Long Lag Independent Increment” in Separation Models
- Luck Long Lag (1–3 Years) Meteorological-Disease Association: Does It Carry Information Independent of Short Lags?
- Pure DLNM vs. Luck Rough Screening + DLNM: Head-to-Head Comparison of Three Models
- Research Pool: Spatiotemporal Frequency Topology
Evidence Level: B
References: PMID 37101120, PMID 20577014, PMID 20812303, PMID 30963815, PMID 22578440, PMID 29040386, PMID 36073640
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