阈值还是坡度:月相-月经耦合衰减的判决逻辑与断点回归的误用

Written by

in

Language

Evidence level: B

破坏”天人相应”的夜间光照,究竟是跨过某条线才生效,还是每多一分就衰减一分?这个问题不是靠堆数据回答的,而是靠模型之间的正面对决——而大多数人选的工具,从一开始就选错了。

正文

“光照破坏天人相应”这句话,藏着一个没被回答的追问:它是”跨过某条线才破坏”,还是”每多一分就少一分”? 两种答案指向完全不同的生活方案——前者说”把卧室灯压到某 lux 以下就保住了”,后者说”不存在安全线,能少一分是一分”。前作已经把耦合的存在(PMID 40991714)和个体剂量-效应曲线该长什么样讲清了,本文处理最后一步、也最容易被糊弄的一步:怎么用数据判决”阈值模型”与”连续坡度模型”这两家。

第一件事:多数人在这一步选错了工具。 常见说法是”用断点回归(RDD)测临界光照强度”。但 RDD 有硬前提——存在基于阈值的分配规则:CD4 计数低于 200 就给药,个体在界线两侧形成准随机(PMID 25061922)。夜间光照没有这条规则,没有人”按 100 lux 划线决定谁被暴露”;强行套用,估出来的不是生理阈值,而是人为挑出的断点,且只在阈值邻域内有效(PMID 37528626)。正确的工具是分段回归/阈值回归:模型本身就是”两段不同斜率在未知转点相接”(PMID 8928781),转点由数据估出而非人为指定(PMID 40896720 给出可复现流程),并可加权纳入研究间异质性(PMID 37251960)。这不是术语之争:RDD 报的是”界线两侧的跳跃”,分段回归报的是”斜率在哪换挡”——月末同步度是连续量,只有后者对得上。

第二件事:模型对决不能查表。 阈值模型有个统计麻烦——零假设下转点不可识别,似然比统计量不服从标准卡方分布(非正则渐近)。所以”p<0.05 就宣布有阈值”是错的,必须用自助法/置换法构造零分布(PMID 29037149 用蒙特卡洛验证了检验的第一类错误率)。可执行的刀法:把线性、折棍(hockey-stick)、S 形、双段线性四种模型放在同一批个体数据上比 AIC/BIC,再用自助法给出转点位置的置信区间——若转点的 95% 区间横跨整个剂量范围,那就是没有阈值,只有坡度。

第三件事:最大的陷阱是”群体 S 形”冒充”个体阈值”。 褪黑素抑制的曲线是 S 形:约 100 lux 达半程(PMID 10922269),1.5 melanopic lux 启动、305 saturate(PMID 30697806)。S 形不是阈值——它没有断点。 更关键的是:若每个人的折点位置各不相同(个体光敏感度可相差一个数量级),把这些个体曲线平均起来,必然得到一条平滑的 S 形。于是群体层面看到”低剂量区几乎无响应、高剂量区饱和”,极易被读成”存在临界线”,而真相可能只是”每人各有各的线、且分布很宽”。判决只有一条路:个体内设计——同一人被施加 ≥5 级光剂量,为每人单独估曲线,再看个体转点是”聚在一处”还是”散开”。

第四件事:剂量不是标量。 “临界光照强度”默认 lux 是唯一自变量。但双因素实验显示时长主效应显著、强度不显著(PMID 21532952),短脉冲每分钟效率更高(PMID 22526883)。若真实杠杆是时长与光谱,任何一维 lux 阈值模型都会被错配,估出的”阈值”只是丢失变量的投影。

第五件事:必须扣掉引力这个共变项。 2025 年 Sci Adv 的分析显示,2010 年后耦合并未消失,而是集中在 1 月,作者归因于日-地-月引力最强的时节(PMID 40991714)。这意味着不控制季节/引力项,1 月残留同步会被误读为”阈值以下仍保留耦合”,把整条阈值往高处推。该研究还给出一个关键分层变量:只有周期长度落在 26–36 天窗口的女性才有条件同步——阈值模型必须把它作为协变量,否则同步度差异会被错算成光照效应。

可执行的判决方案:被试内交叉设计,每人在 ≥6 个周期内经历 ≥5 级夜间黑视素照度(含时长与蓝光比的正交操作),结局用圆形统计的同步指数(Rayleigh 检验的合向量长度 R̄,需做相位折叠处理轴双峰,方法学见 PMID 40991714);先对每个个体拟合四模型族并留出转点置信区间,再检验”个体转点是否同质”与”低于转点区间的斜率是否与零等价”(等价性检验/TOST)。三种结局各有含义:个体转点同质且低剂量区斜率与零等价 → 真阈值,可以划安全线;个体转点离散 → 是混合伪影,只能谈”个体线”;转点区间横跨全量程 → 没有阈值,只有坡度。

对”天人相应”的意义:这一步做完,”光煞”才从”玄”变成”界”。古人说”明暗须有分寸”,现代版的问法更锋利——分寸是一条线还是一个坡度?而这条线,是不是每个人都画在同一个位置?

证据与局限

已验证: RDD 的适用前提(阈值分配规则、仅邻域内有效)与实施规范(PMID 25061922, 37528626);分段回归/阈值回归可估计未知转点且流程可复现(PMID 8928781, 40896720);阈值模型需自助法/置换法处理非正则渐近(PMID 29037149);加权变点回归可从系统评价估 NOAEL(PMID 37251960);褪黑素抑制曲线为 S 形、约 100 lux 半程、1.5–305 melanopic lux 区间(PMID 10922269, 30697806);时长主效应强于强度、短脉冲效率更高(PMID 21532952, 22526883);2010 年后月相-月经耦合集中于 1 月、周期长度 26–36 天为同步窗口(PMID 40991714)。

部分验证: 上述方法学证据全部来自其他领域(流行病学、毒理学、光生物学),尚无一项以月相-月经同步度为结局做过阈值 vs 坡度的模型对决;”群体 S 形可由个体转点异质混合产生”在光生物学中为公认数理事实,但未在月经同步场景中被量化。

待验证: 月相-月经耦合衰减是否存在统计上可检测的阈值(个体转点是否同质);群体层面的”S 形”中有多少来自个体异质混合;剂量应由强度、时长还是光谱主导(若为多变量,一维阈值模型的形式是否正确);季节/引力项与光照项在观测数据中的可识别边界。

相关阅读

证据等级: B
参考文献: PMID 40991714, PMID 10922269, PMID 30697806, PMID 25061922, PMID 37528626, PMID 8928781, PMID 40896720, PMID 29037149, PMID 37251960, PMID 21532952, PMID 22526883


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。

Threshold or Slope: Judgment Logic and Misuse of Breakpoint Regression in Moon-Phase-Menstrual Coupling Attenuation

Evidence level: B

Does disrupting “Heaven-Human Correspondence” via nocturnal illumination require crossing a specific line to take effect, or does attenuation occur incrementally with every additional unit? This question cannot be answered by piling up data; it requires a head-to-head confrontation between models—yet most practitioners select tools that are fundamentally misaligned from the outset.

Main Text

“Disruption of Heaven-Human Correspondence via Light” hides an unanswered question: Is the effect binary (“cross a threshold to disrupt”) or linear (“attenuate incrementally with every unit increase”)? These answers dictate entirely different life protocols—the former suggests “keep bedroom light below X lux to preserve synchrony,” while the latter implies “no safe threshold exists; minimize exposure as much as possible.” Previous work has clarified the existence of coupling (PMID 40991714) and the expected shape of individual dose-response curves; this article addresses the final, most easily misleading step: how to adjudicate between the “Threshold Model” and the “Continuous Slope Model” using data.

First Issue: Most practitioners select the wrong tool at this step. The common claim is “use Regression Discontinuity Design (RDD) to measure critical light intensity.” However, RDD relies on a hard prerequisite—the existence of a threshold-based allocation rule: e.g., administering medication when CD4 counts drop below 200, creating quasi-randomization on either side of the boundary (PMID 25061922). No such rule exists for nocturnal light; no one “draws a line at 100 lux to decide who is exposed.” Forcing RDD here estimates an artificial discontinuity rather than a physiological threshold, valid only within the immediate neighborhood of that arbitrary line (PMID 37528626). The correct tool is Piecewise Regression / Threshold Regression: the model itself posits “two different slopes meeting at an unknown inflection point” (PMID 8928781), where the turning point is estimated from data rather than arbitrarily specified (PMID 40896720 provides a reproducible workflow), and can incorporate study-level heterogeneity via weighting (PMID 37251960). This is not a terminology dispute: RDD reports “jumps on either side of a line,” while piecewise regression reports “where the slope shifts”—menstrual synchrony is a continuous variable, and only the latter aligns.

Second Issue: Model confrontation cannot rely on lookup tables. The threshold model has a statistical complication—the inflection point is unidentified under the null hypothesis, and the likelihood ratio statistic does not follow a standard chi-square distribution (non-regular asymptotic). Therefore, “declare a threshold exists if p<0.05” is incorrect; one must construct the null distribution via bootstrapping or permutation (PMID 29037149 verified Type I error rates via Monte Carlo). Executable methodology: Fit four model families (linear, hockey-stick, sigmoidal, two-segment linear) on the same individual dataset and compare via AIC/BIC, then use bootstrapping to generate confidence intervals for the turning point—if the 95% CI for the turning point spans the entire dose range, there is no threshold, only a slope.

Third Issue: The largest trap is “Population Sigmoid” masquerading as “Individual Threshold.” Melatonin suppression curves are sigmoidal: reaching half-maximum at ~100 lux (PMID 10922269), with onset at 1.5 melanopic lux and saturation at 305 (PMID 30697806). A sigmoid is not a threshold—it has no discontinuity. More critically: if each individual’s inflection point varies significantly (individual light sensitivity can differ by an order of magnitude), averaging these individual curves necessarily yields a smooth sigmoid. Thus, at the population level, observing “no response at low doses and saturation at high doses” is easily misread as “a critical line exists,” while the truth may simply be “everyone has their own line, and the distribution is wide.” The adjudication path is singular: Within-subject design—expose each person to ≥5 levels of nocturnal melatonin-suppressing illuminance (orthogonally manipulating duration and blue-light ratio), fit curves individually for each person, then assess whether individual turning points “cluster together” or “scatter.”

Fourth Issue: Dose is not a scalar. “Critical light intensity” defaults to lux as the sole independent variable. However, two-factor experiments show significant main effects for duration but not intensity (PMID 21532952), with short pulses being more efficient per minute (PMID 22526883). If the true lever is duration or spectrum, any univariate lux threshold model will be misspecified; the estimated “threshold” is merely a projection of lost variables.

Fifth Issue: Must control for gravitational confounding. A 2025 Sci Adv analysis indicates that post-2010 coupling did not disappear but concentrated in January, attributed by authors to the period of strongest sun-earth-moon gravitational forces (PMID 40991714). This means failing to control for season/gravity terms will misread January residual synchrony as “coupling persists below the threshold,” artificially pushing the entire threshold upward. The study also provides a key stratification variable: only women with cycle lengths falling within the 26–36 day window exhibit conditional synchrony—the threshold model must include this as a covariate; otherwise, synchrony differences will be erroneously calculated as light effects.

Executable Adjudication Protocol: Within-subject crossover design, where each participant experiences ≥5 levels of nocturnal melatonin-suppressing illuminance (including orthogonal manipulation of duration and blue-light ratio) across ≥6 cycles. Outcome: synchrony index via circular statistics (mean vector length $\bar{R}$ from Rayleigh test, requiring phase-folding handling for bimodal axes; methodology see PMID 40991714). First, fit the four model families for each individual and derive confidence intervals for the turning point; then test “homogeneity of individual turning points” and “whether slopes below the turning point interval are equivalent to zero” (equivalence testing/TOST). Three outcomes carry distinct meanings:
1. Homogeneous turning points + slope at low doses equivalent to zero → True threshold; a safe line can be drawn.
2. Dispersed turning points → Mixed artifact; only “individual lines” can be discussed.
3. Turning point interval spans full range → No threshold, only a slope.

Significance for “Heaven-Human Correspondence”: Once this step is completed, “Light Sha” transforms from “symbolic” to “bounded.” Ancient texts say “light and darkness must have measure”; the modern formulation is sharper—is the measure a line or a slope? And is this line drawn at the same position for everyone?

Evidence and Limitations

Verified: RDD applicability prerequisites (threshold allocation rule, validity only within neighborhood) and implementation standards (PMID 25061922, 37528626); piecewise/threshold regression capability to estimate unknown turning points with reproducible workflows (PMID 8928781, 40896720); necessity of bootstrapping/permutation for non-regular asymptotics in threshold models (PMID 29037149); weighted change-point regression for estimating NOAEL from systematic reviews (PMID 37251960); melatonin suppression curves are sigmoidal (~100 lux half-max, 1.5–305 melanopic lux range) (PMID 10922269, 30697806); duration main effect stronger than intensity, short pulses more efficient (PMID 21532952, 22526883); post-2010 moon-phase-menstrual coupling concentrated in January, 26–36 day cycle window for synchrony (PMID 40991714).

Partially Verified: All methodological evidence above originates from other fields (epidemiology, toxicology, photobiology); no study yet has conducted a threshold vs. slope model confrontation with moon-phase-menstrual synchrony as the outcome; “population sigmoid arising from heterogeneous individual turning points” is an accepted mathematical fact in photobiology but not yet quantified in the menstrual synchrony context.

To Be Verified: Whether statistically detectable thresholds exist for moon-phase-menstrual coupling attenuation (homogeneity of individual turning points); the proportion of the population-level “sigmoid” arising from individual heterogeneity; whether dose should be driven by intensity, duration, or spectrum (if multivariate, is the univariate threshold model form correct); identifiable boundaries between season/gravity terms and light terms in observational data.

Related Reading

Evidence Level: B
References: PMID 40991714, PMID 10922269, PMID 30697806, PMID 25061922, PMID 37528626, PMID 8928781, PMID 40896720, PMID 29037149, PMID 37251960, PMID 21532952, PMID 22526883


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 *