Blog

  • 刮风那天和静风那天,是两种「煞」:室内次声能不能给建筑体型分型

    Evidence level: B

    “签名”这个词预设了写字的人固定、读字的人能读。可室内次声谱是三项相乘的乘积,而这三项里最大的一项,恰好是每天在变的那一项。

    正文

    第一层,”签名”是个乘积,不是指纹。 室内测到的谱,等于室外激励、建筑传递、房间响应三者相乘。要从这个乘积里反推中间那一项,前提是另两项已知。现有设计只测乘积本身——测了谱、分了型、做了判别,缺的正是那个识别条件。这不是分析做得不够,是这场观测在数学上还不足以定解。

    第二层,风不是背景噪声,是最大的一项。 次声测量第一大污染源就是风:湍流压力直接压在传感器上,它不是声。管道阵列靠空间平均压它,最多约 20 dB,而且风速越大抑制越差;30–270 米的光纤传感器能到约 30 dB,但平行风向比垂直风向多约 4 dB(PMID 23742329)。4 dB 是什么量级?同一栋楼,风向转九十度,谱就动 4 dB。而”宽楼体与尖塔之间差几个 dB”,至今没人测过。更麻烦的是,那些用来压风噪的管道本身会改写传感器的频响,必须就地标定才能读(PMID 21895058)——你以为是楼在写签名,其实是管路先改了一笔。

    第三层,楼在次声面前不是散射体。 20 Hz 的波长 17 米,4 Hz 的波长 86 米。一栋 20 米宽的楼,比 4 Hz 的波小四倍多,它挡不住声,只是压力场里的一个障碍物,在下游拖出湍流尾流。尾流主频由涡脱落定:取斯特劳哈尔数 0.2,主频等于 0.2 乘风速除以特征尺寸。5 米每秒的风、20 米楼宽,给 0.05 Hz;要落到 4 Hz,特征尺寸得 0.25 米;落 16 Hz,得 0.06 米。进这个带的是檐口、栏杆、线缆、女儿墙这些厘米级细部,不是楼体体量。“宽楼体对尖塔”这个分层,分的恰好是不进带的那一维。

    第四层,房间比楼更能改谱,而且方向反直觉。 矩形房间第一阶轴向模态是声速除以两倍进深。5 米给 34 Hz,10 米给 17 Hz,20 米给 8.6 Hz,40 米给 4.3 Hz。普通住宅在 4–16 Hz 连第一阶模态都没到,室内压力同相、没有驻波——它是个气球,不是共鸣箱。所以”房间模态放大次声”这句话只对大厅堂成立,对卧室不成立。同一栋楼,卧室和大堂测出来的是两样东西;而从单个房间响应反推房间尺度,本身也被证明只有扩散场模型勉强靠得住(PMID 18681590)。

    第五层,人耳那端也不支持”频谱指纹”。 4 Hz 的听阈随刺激时长一直往下掉,掉到 4 秒还没稳;而 16 Hz 和 32 Hz 大约 1 秒就稳了(PMID 32872982)。原因是 4 Hz 被听觉系统解析成一串独立的压力脉冲,不是一个音。所以谱上那个 4 Hz 的峰,在感知上不是”一个很低的音”,是”每秒推你四下”。用三分之一倍频程的谱型去描述它,描述的是一个跟感受不同构的量。

    第六层,判别分析在这套数据上会给出漂亮的假象。 三类体型、每类十来栋、几十个频点当特征,就是典型的高维小样本。小样本下 K 折交叉验证的准确率估计严重偏倚,样本量到 1000 仍可见;嵌套交叉验证与独立测试集才接近无偏,而在合并的训练测试数据上做特征筛选,贡献的偏倚比调参更大(PMID 31697686)。三类平衡准确率的机会水平是 33%。报一个 78% 却不报置换基线,等于什么都没说。

    第七层,就算真分出来了,下游是空的。 72 小时、1.6–20 Hz、约 90 dB 峰值、37 人双盲三臂交叉,入睡后觉醒时间差异 −1.36 分钟(95%CI −6.60 至 3.88,p=0.60),而作为阳性对照的交通噪声恶化了 6.07 分钟(PMID 36946580);28 晚、6 Hz、80–90 dB、38 人,全部行为学指标为零(PMID 33542290)。静息态网络倒是有变化,但方向与自报烦恼度相反(PMID 39427080)。可分性不等于危害性,这两件事必须分开报。

    判决四件事。 一,把风当主变量而不是噪声:所有测点同步记风速风向,按风速分层,先看”体型效应”在静风层里还剩多少。二,先报重复性:同一房间、不同日期、同一测点的谱差异,这是”签名”能不能用的下限,这个数不报,后面全是空谈。三,分层按细部尺寸重划:檐口高度、栏杆间距、女儿墙厚度,不是楼宽。四,判别分析必带置换基线:打乱标签跑一千次,报超出机会水平的增量,不报绝对准确率。

    跟古人怎么接。 形煞讲的是”形”,把形当因、气当果。这里对上的是判断单位:风水从不判”这栋楼”,判的是”这个点”——坐向、穴位、门位。声学的结论是同一个:4–16 Hz 的室内是压力场,处处同相,室内测点没有空间结构;真正有空间结构的,是室外那段尾流。所以”煞”若真有物理对应物,它在室外近场,不在室内。反过来说,把仪器搬进客厅去找签名,找的是被房间抹平之后剩下的那一点。古人量的从来不是一个房子,是一个位置。

    证据与局限

    已验证: 次声测量中风噪是首要污染源,空间平均抑制上限约 20 dB 且随风速下降、光纤传感器可达约 30 dB 并存在约 4 dB 风向依赖(PMID 23742329);风噪抑制管路会改写传感器频响、必须就地标定(PMID 21895058);4 Hz 听阈的时长依赖显著长于 16/32 Hz(PMID 32872982);小样本下 K 折交叉验证偏倚严重、嵌套交叉验证接近无偏(PMID 31697686);72 小时 90 dB 峰值次声双盲暴露对睡眠无显著影响(PMID 36946580);28 晚 6 Hz 80–90 dB 暴露行为学指标为零(PMID 33542290)。

    部分验证: 次声与低频噪声同烦恼度、睡眠问题存在关联,但慢性病证据非常有限、高烦恼率约 10%(PMID 26994804);低频能量占比与调幅深度各自独立抬高烦恼度,等效于最多 8 dB 声压级变化(PMID 29783747);实测建筑内 100 Hz 以下总声级 64–73 dB,远低于有害水平,但”频谱不平衡”可能解释异常感受(PMID 37818147);房间尺度的可靠反推仅扩散场模型成立(PMID 18681590)。

    待验证: 板式宽体、窄檐口、尖角塔楼三型的室内次声是否真的存在可分离的频段-强度结构;若存在,其效应量是否大于风速与风向带来的方差;细部尺寸(檐口/栏杆/女儿墙)是否是比楼体体量更好的分层变量;上述可分性是否对应任何剂量-反应的健康结局。

    相关阅读

    证据等级: B(风噪抑制上限与风向依赖、听阈时长依赖、小样本验证偏倚、双盲暴露阴性结果均有 PubMed 直接支撑;”建筑体型对应次声签名”本身无先例,属本文按声学量纲推算的检验提案)
    参考文献: PMID 23742329, PMID 21895058, PMID 32872982, PMID 31697686, PMID 36946580, PMID 33542290, PMID 39427080, PMID 26994804, PMID 29783747, PMID 37818147, PMID 18681590


    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。

    Wind and Calm Days Are Two Types of \”Sha\”: Can Indoor Infrasound Classify Building Shapes?

    Evidence level: B

    The term “signature” presupposes a fixed writer and a capable reader. However, the indoor infrasound spectrum is a product of three factors, among which the largest—and most variable—is the daily fluctuation.

    Main Text

    Layer 1: The “Signature” is a Product, Not a Fingerprint. The spectrum measured indoors equals the product of outdoor excitation, building transmission, and room response. To reverse-engineer the middle term requires knowledge of the other two. Current designs measure only the product itself—they record the spectrum, classify morphology, and make judgments—yet lack the necessary identification conditions. This is not a failure of analysis; rather, the observation is mathematically underdetermined.

    Layer 2: Wind is Not Background Noise; It Is the Dominant Term. The primary pollution source for infrasound measurement is wind: turbulent pressure acts directly on the sensor. It is not sound. While pipe arrays suppress it via spatial averaging to approximately 20 dB, suppression worsens as wind speed increases; 30–270 meter fiber optic sensors can reach ~30 dB but exhibit an additional ~4 dB dependence on wind direction (PMID 23742329). What is a 4 dB shift? For the same building, rotating the wind direction by ninety degrees shifts the spectrum by 4 dB. Yet, the difference between “wide buildings and spires” in decibels remains unmeasured. Complicating matters further, the pipes used to suppress wind noise themselves alter the sensor’s frequency response and must be calibrated in situ to read accurately (PMID 21895058)—you are not reading the building’s signature; the piping has already altered the record.

    Layer 3: Buildings Are Not Scatterers Before Infrasound. At 20 Hz, the wavelength is 17 meters; at 4 Hz, it is 86 meters. A building 20 meters wide is more than four times smaller than the 4 Hz wave; it cannot block sound but acts as an obstacle in the pressure field, generating a turbulent wake downstream. The wake’s dominant frequency is determined by vortex shedding: using a Strouhal number of 0.2, the main frequency equals 0.2 times wind speed divided by characteristic dimension. A 5 m/s wind and a 20 m wide building yield 0.05 Hz; to reach 4 Hz, the characteristic dimension must be 0.25 meters; to reach 16 Hz, it must be 0.06 meters. What enters this band are eaves, railings, cables, and parapets—centimeter-scale details—not the building’s bulk. The distinction between “wide buildings and spires” divides precisely that dimension which does not enter the band.

    Layer 4: Rooms Alter the Spectrum More Than Buildings, and Counter-Intuitively. The first axial mode of a rectangular room is the speed of sound divided by twice the depth. Five meters yields 34 Hz; ten meters, 17 Hz; twenty meters, 8.6 Hz; forty meters, 4.3 Hz. Ordinary residential spaces do not even reach the first mode between 4–16 Hz; indoor pressure is in-phase with no standing waves—it acts as a balloon, not a resonator. Therefore, the statement “room modes amplify infrasound” holds only for halls, not bedrooms. In the same building, bedrooms and halls yield different spectra; furthermore, inferring room scale from single-room response has been shown to rely only on the diffuse field model (PMID 18681590).

    Layer 5: The Human Ear Does Not Support a “Spectral Fingerprint.” The hearing threshold at 4 Hz decreases continuously with stimulus duration, remaining unstable even after 4 seconds; whereas thresholds at 16 Hz and 32 Hz stabilize within approximately 1 second (PMID 32872982). The reason is that the auditory system parses 4 Hz into a series of independent pressure pulses rather than a single tone. Thus, the peak at 4 Hz on the spectrum is not perceived as “a very low tone,” but as “pushing you four times per second.” Describing it with a one-third octave band profile characterizes a quantity structurally distinct from perception.

    Layer 6: Discriminant Analysis Produces Beautiful Illusions on This Data. Three morphology classes, roughly ten buildings each, and dozens of frequency points as features constitute a classic high-dimensional small-sample scenario. Under such conditions, K-fold cross-validation accuracy estimates suffer severe bias even with sample sizes reaching 1000. Nested cross-validation and independent test sets approach unbiasedness, whereas feature selection on merged training-test data introduces bias greater than hyperparameter tuning (PMID 31697686). The chance-level balanced accuracy for three classes is 33%. Reporting a 78% accuracy without reporting the permutation baseline amounts to saying nothing.

    Layer 7: Even If Classified, the Downstream Consequences Are Empty. In a 72-hour, 1.6–20 Hz, ~90 dB peak, three-arm crossover double-blind study with 37 subjects, the difference in wake-up time after sleep onset was −1.36 minutes (95%CI −6.60 to 3.88, p=0.60), while as a positive control, traffic noise worsened sleep by 6.07 minutes (PMID 36946580); in a 28-night, 6 Hz, 80–90 dB study with 38 subjects, all behavioral indices were zero (PMID 33542290). Resting-state networks did show changes, but the direction opposed self-reported distress (PMID 39427080). However, separability does not equate to harm; these two matters must be reported separately.

    Verdict on Four Matters. First, treat wind as the primary variable, not noise: record wind speed and direction at all measurement points simultaneously, stratify by wind speed, and first examine how much “morphology effect” remains in the calm layer. Second, report repeatability first: the spectral differences within the same room across different dates at the same point represent the floor for whether a “signature” is usable; without this number, all subsequent discussion is empty. Third, re-stratify by fine-scale dimensions: eave height, railing spacing, parapet thickness—not building width. Fourth, discriminant analysis must include a permutation baseline: shuffle labels and run 1000 times; report the increment beyond chance, not absolute accuracy.

    How to Connect with Ancients. “Form Sha” discusses “form,” treating form as cause and qi as effect. Here, the correspondence lies in the judgment unit: Feng Shui never judges “this building”; it judges “this point”—orientation,穴位 (acupoint), and gate position. Acoustics concludes the same: indoor space at 4–16 Hz is a pressure field, in-phase everywhere, with no spatial structure among indoor measurement points. The true spatial structure exists in the outdoor wake. Therefore, if “Sha” has a physical counterpart, it lies in the near-field outdoors, not indoors. Conversely, moving instruments into a living room to find a signature seeks only the remnant flattened by the room. Ancients measured never an entire house, but a position.

    Evidence and Limitations

    Verified: Wind noise is the primary pollution source in infrasound measurement; spatial averaging suppresses it to ~20 dB with degradation as wind speed increases, while fiber optic sensors reach ~30 dB with ~4 dB wind direction dependence (PMID 23742329); wind noise suppression piping alters sensor frequency response requiring in situ calibration (PMID 21895058); the duration dependence of the 4 Hz hearing threshold is significantly longer than at 16/32 Hz (PMID 32872982); K-fold cross-validation shows severe bias in small samples while nested cross-validation approaches unbiasedness (PMID 31697686); 72-hour peak infrasound exposure at 90 dB in double-blind studies shows no significant effect on sleep (PMID 36946580); behavioral indices were zero in 28-night exposure at 6 Hz, 80–90 dB (PMID 33542290).

    Partially Verified: Infrasound and low-frequency noise correlate with distress and sleep issues, but evidence for chronic disease is very limited; high distress rates are ~10% (PMID 26994804); low-frequency energy proportion and modulation depth independently elevate distress, equivalent to a maximum 8 dB sound pressure level change (PMID 29783747); measured total sound levels below 100 Hz indoors range from 64–73 dB, well below harmful levels, but “spectral imbalance” may explain anomalous sensations (PMID 37818147); reliable inference of room scale holds only for the diffuse field model (PMID 18681590).

    To Be Verified: Whether indoor infrasound truly exhibits separable frequency-intensity structures for plate-wide bodies, narrow eaves, and spire towers; if so, whether the effect size exceeds variance caused by wind speed and direction; whether fine-scale dimensions (eaves/railings/parapets) serve as better stratification variables than building bulk; and whether such separability corresponds to any dose-response health outcome.

    Related Reading

    Evidence Level: B (Wind noise suppression limits and wind direction dependence, hearing threshold duration dependence, small-sample validation bias, and negative double-blind exposure results all have direct PubMed support; “building morphology corresponding to infrasound signatures” itself has no precedent, belonging to a testing proposal derived in this article via acoustic dimensional reasoning)
    References: PMID 23742329, PMID 21895058, PMID 32872982, PMID 31697686, PMID 36946580, PMID 33542290, PMID 39427080, PMID 26994804, PMID 29783747, PMID 37818147, PMID 18681590


    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.

  • 残差里没有「那个东西」:想把评级剩下的那部分归给微气候,得先过四道关

    Evidence level: B

    控制物探之后评级还剩一点东西,所有人都想立刻给它起个名字。但残差是个差额,不是个包裹——它里面没有”那个东西”,只有一张还没试完的清单。

    正文

    前一篇说的是:那个增量在别的地方还算不算同一个量。这一篇往后退一步,问一件更根本的事——那个增量到底是什么。最流行的答案有三个字:微气候。听着顺,但它现在还不是结论,只是一个候选,而且是最难验的那类候选。

    第一层,残差是个差额,不是个包裹。 残差在定义上就是”1 减去已解释的部分”,它是一个数,不是一个容器。问”残差里装的是什么”,逻辑上已经有毛病——数据只能告诉你”还剩多少没解释”,不能告诉你”剩下的是谁”。方向只能是”拿候选去撞残差”,不是”从残差里挖出答案”。前者是归纳神话,后者才是可做的事。

    第二层,候选得先跟已测的物探”脱钩”。 微气候不是自由变量,它是地形驱动的;地形同时驱动岩性、断层走向、土壤厚度。所以评级、断层距离、湿度很可能共用一个上游母体:地形。此时”评级残差与湿度显著相关”最省事的解释不是”评级编码了湿度”,而是”它们又碰上了同一个共同原因”。麻烦在于,中介和共同原因在观测数据里长得一模一样,靠相关系数分不开,只能靠时间先后、外生干预,或一个能把地形效应剥净的强设计。

    第三层,拿地形当微气候的替身,替身本身就不合格。 很多这类研究没实测微气候,用的是 DEM 派生地形加林冠结构,默认它们能代替微气候。这个默认已被明确驳回:一项用苔藓分布做判据的研究直接问”高分辨率地形加林冠结构能不能替代微气候实测”,答案是不能(PMID 35077798)。没实测的话,”微气候中介”这一层从头到尾都是替身的替身,可靠性不可能高过那个替身。

    第四层,中介-结局的未测混杂。 间接效应的估计有一个硬前提:中介与结局之间没有未测混杂。而微气候与氡、与健康结局之间恰恰最容易被污染——建筑密闭性、通风习惯、居住时长、社会经济地位,全都同时影响微气候和结局,且几乎测不全。偏倚有多大可以量化:中介 E-value 告诉你”要多强的中介-结局混杂才能把间接效应抹平”(PMID 31348008)。这是必须报的数,实际研究里报的人极少。

    第五层,占比这个数本身就不可解释。 “评级信息里有四成由湿度中介”——这句最像结论,其实最不能写。解释占比(PTE)的综述把三种解释全过了一遍,结论很硬:PTE 只在研究者愿意接受极强假设时才有意义,三种解释都依赖不可验证的前提(PMID 39053571)。能写的是”假设成立时间接效应的点估计、区间与 E-value 各是多少”,不能写”四成归它”。

    第六层,功效过不去,前面全白搭。 中介效应的检验功效远低于总效应——总效应卡在显著边缘,中介基本必然检不出。功效和样本量有现成的仿真工具(PMID 37231326),先算再决定做不做。几十个点位、十几个事件的规模,连总效应都勉强,做中介属于用显微镜看月亮。

    第七层,这类设计还有个专门的坑。 用”到某处的距离”做暴露的生态学研究,即使真实关联为零,二维空间分布本身也能造出显著正关联,这在仿真里被反复验证过(PMID 23557010)。所以”残差与湿度显著相关”在这个设计下的基线假阳性本来就偏高,p 值不能按常规读。能压住这个坑的只有负对照:找”不可能有因果通路但与地质共变”的暴露做负对照暴露,找”不受评级影响但与居住行为相关”的结局做负对照结局;双负对照在部分负对照无效时仍能一致估计(PMID 40983525)。而按生态回归的敏感性框架,这类研究里未测混杂与区内暴露变异,通常比空间残差依赖更该先处理(PMID 12762436)。

    判决。 残差不是一个待拆的包裹,而是一张待补的清单。你不能”拆开它”,只能”一条条试”,每试一条都要过四道关:脱钩、实测、混杂、可解释。过不了的,不是候选,是故事。真正的价值不在最终是否归给微气候——在于它逼着研究把”未测”从托词变成必须报的数。

    易学这一层,古人给的答案其实更老实。 形峦派讲”内局外局”:外局是山川大势,看得见;内局是宅院小环境,藏风聚气、难量。而”气”这个字,古人从没说它等于某一个可测的物理量——它是统称,是”多要素配比之后才成立的那一部分”。所以老风水师看地,很少说”这里凶,因为湿”,他说的是”这一片的配比不对”。用统计的话翻译:他拒绝的正是那个不成立的归因要求。残差在模型里不是缺失数据,它是一个未测的共同因子——一个模型项,不是一个待命名的实体。这一点上,”气”的模糊不是偷懒,是诚实地承认自己只能给整体判断,给不了单项分解。

    证据与局限

    已验证: DEM 派生地形加林冠结构不能替代微气候实测(PMID 35077798);PTE 仅在极强假设下可解释(PMID 39053571);中介-结局未测混杂可用中介 E-value 定量(PMID 31348008);中介分析功效远低于总效应、已有仿真算法(PMID 37231326);邻近型生态设计在零效应下仍产生显著正关联(PMID 23557010);双负对照在部分负对照无效时仍可一致估计(PMID 40983525);生态回归中未测混杂与区域内暴露变异的偏倚通常重于空间残差依赖(PMID 12762436);E-value 可作为未测混杂强度的统一标尺(PMID 28693043)。

    部分验证: 地形与林冠结构确实联合约束微气候,但解释份额随景观类型变化(PMID 30246358);深度学习氡潜力制图中地形类因子参与但非主导(PMID 34673157、33254926)。

    待验证: 凶地评级残差与实测微气候/水文/景观结构之间是否存在非共同原因的解释路径;在可行样本量下这条路径是否具备可检出的功效;负对照设计能否把该类研究的基线假阳性压到可接受水平。

    相关阅读

    证据等级: B(地形不可替代微气候、PTE 不可解释、中介 E-value、中介功效、生态设计假阳性、双负对照均有 PubMed 直接支撑;”凶地评级残差的未测维度归因”本身无先例,属研究设计提案)
    参考文献: PMID 35077798, PMID 39053571, PMID 31348008, PMID 37231326, PMID 23557010, PMID 40983525, PMID 12762436, PMID 28693043, PMID 30246358, PMID 34673157, PMID 33254926


    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。

    What’s Missing in the Residuals: Four Hurdles Before Attributing Credit Risk to Microclimates

    Evidence level: B

    After controlling for geophysical surveying, a residual remains. Everyone wants to immediately name it. But a residual is a difference, not a package—it contains no “something,” only an unfinished list of candidates.

    Main Text

    The previous article addressed whether that increment represents the same quantity elsewhere. This one steps back to ask a more fundamental question: what is that increment exactly. The most popular answer consists of three characters: microclimate. It sounds plausible, but it is not yet a conclusion; rather, it is a candidate, specifically the most difficult to verify.

    First layer: The residual is a difference, not a container. By definition, the residual equals “1 minus the explained portion.” It is a number, not a vessel. Asking “what is inside the residual” is logically flawed—data can only tell you “how much remains unexplained,” not “who remains.” The direction must be “to collide candidates against the residual,” not “to excavate answers from the residual.” The former is an inductive myth; the latter is actionable.

    Second layer: Candidates must first “decouple” from measured geophysical data. Microclimate is not a free variable; it is driven by topography. Topography simultaneously drives lithology, fault orientation, and soil thickness. Thus, rating residuals, fault distance, and humidity likely share an upstream common cause: topography. Consequently, the most parsimonious explanation for a “significant correlation between rating residuals and humidity” is not that “rating encodes humidity,” but that “they encounter the same common cause.” The complication lies in the fact that mediators and common causes look identical in observational data; correlation coefficients cannot separate them. Only temporal precedence, exogenous intervention, or a strong design capable of stripping away topographic effects can resolve this.

    Third layer: Using topography as a proxy for microclimate is flawed. Many such studies do not measure microclimate directly; instead, they use DEM-derived topography and canopy structure, defaulting to the assumption that these can substitute for microclimate. This default has been explicitly rejected: a study using moss distribution as a criterion directly asked whether “high-resolution topography plus canopy structure can substitute for microclimate measurement,” and the answer is no (PMID 35077798). Without direct measurement, the “microclimate mediator” layer is entirely composed of proxies for proxies; its reliability cannot exceed that of the proxy itself.

    Fourth layer: Unmeasured confounding in the mediator-outcome pathway. Estimating indirect effects rests on a hard prerequisite: no unmeasured confounding between the mediator and the outcome. Yet microclimate and radon, or health outcomes, are precisely the most susceptible to contamination—building airtightness, ventilation habits, duration of residence, and socioeconomic status all simultaneously affect both microclimate and outcomes, and are rarely fully measured. The magnitude of bias can be quantified: the mediator E-value tells you “how strong the mediator-outcome confounding must be to nullify the indirect effect” (PMID 31348008). This is a number that must be reported, yet actual studies report it rarely.

    Fifth layer: The proportion itself is uninterpretable. “Forty percent of rating information is mediated by humidity”—this sentence sounds most like a conclusion, yet it is the least justifiable. Reviews explaining proportions (PTE) have examined three interpretations, concluding very firmly: PTE only holds meaning when researchers accept extremely strong assumptions, and all three interpretations rely on unverifiable premises (PMID 39053571). What can be written is “what are the point estimate, interval, and E-value for the assumed effect size,” not “forty percent is attributed to it.”

    Sixth layer: Power is insufficient, rendering previous efforts moot. The power to detect mediator effects is far lower than for total effects—total effects often sit on the edge of significance, while mediator effects are almost certainly undetectable. Power and sample size have ready-made simulation tools (PMID 37231326); calculate first, then decide. With dozens of sites and tens of events, detecting total effects is already a stretch; attempting mediator analysis is like using a microscope to view the moon.

    Seventh layer: This design has a specific pitfall. Ecological studies using “distance to a location” as exposure can generate significant positive associations even if the true association is zero, due to the two-dimensional spatial distribution itself—a phenomenon repeatedly verified in simulations (PMID 23557010). Therefore, the baseline false-positive rate for “significant correlation between residuals and humidity” under this design is inherently elevated; p-values cannot be read conventionally. Only negative controls can mitigate this pitfall: find an exposure that “has no causal pathway but covaries with geology” as a negative control exposure, and find an outcome that “is unaffected by rating but correlates with living behavior” as a negative control outcome; dual negative controls provide consistent estimates even when partial negative controls fail (PMID 40983525). Furthermore, within the sensitivity framework for ecological regression, unmeasured confounding and within-area exposure variation typically pose greater bias than spatial residual dependence (PMID 12762436).

    Verdict. The residual is not a package to be unpacked, but a list to be filled. You cannot “unpack it”; you can only “try items one by one,” and each must pass four gates: decoupling, direct measurement, confounding control, and interpretability. If it fails, it is not a candidate, but a story. True value does not lie in whether we ultimately attribute the residual to microclimate, but in forcing research to transform “unmeasured” from an excuse into a reported number.

    On the Yijing layer, the ancients offered a more straightforward answer. The Xing Luan school distinguishes between “internal and external formations”: the external formation refers to the grand mountain ranges, visible to the eye; the internal formation refers to the courtyard’s small environment, where wind gathers and qi accumulates, yet is difficult to quantify. Regarding the character “Qi” (energy), ancients never equated it with a single measurable physical quantity—it is a general term, representing “the portion that holds only after multiple factors are proportioned.” Thus, old Feng Shui masters rarely say “this is凶 (inauspicious) because of dampness”; they say “the proportioning here is incorrect.” Translated into statistical terms: he rejects the very attribution requirement that is invalid. In the model, the residual is not missing data; it is an unmeasured common factor—a model term, not an entity waiting to be named. Here, the ambiguity of “Qi” is not laziness, but a honest admission that one can only provide holistic judgment, not single-item decomposition.

    Evidence and Limitations

    Verified: DEM-derived topography plus canopy structure cannot substitute for microclimate measurement (PMID 35077798); PTE is only interpretable under extremely strong assumptions (PMID 39053571); unmeasured confounding in mediator-outcome pathways can be quantified using mediator E-value (PMID 31348008); power for mediator analysis is far lower than total effects, with simulation algorithms available (PMID 37231326); proximity-based ecological designs generate significant positive associations even under zero effects (PMID 23557010); dual negative controls provide consistent estimates even when partial negative controls fail (PMID 40983525); in ecological regression, bias from unmeasured confounding and within-area exposure variation typically exceeds dependence on spatial residuals (PMID 12762436); E-value serves as a unified scale for unmeasured confounding strength (PMID 28693043).

    Partially Verified: Topography and canopy structure do jointly constrain microclimate, but explanatory shares vary by landscape type (PMID 30246358); in deep learning radon potential mapping, topographic factors participate but are not dominant (PMID 34673157, 33254926).

    To Be Verified: Whether non-common-cause explanatory paths exist between inauspicious rating residuals and measured microclimate/hydrology/landscape structure; whether this path possesses detectable power given feasible sample sizes; whether negative control designs can suppress baseline false-positive rates in such studies to acceptable levels.

    Related Reading

    Evidence Level: B (Topography cannot substitute microclimate, PTE is uninterpretable, mediator E-value, mediator power, ecological design false positives, and dual negative controls all have direct PubMed support; “attribution of unmeasured dimensions to inauspicious rating residuals” itself has no precedent and belongs to research design proposals)
    References: PMID 35077798, PMID 39053571, PMID 31348008, PMID 37231326, PMID 23557010, PMID 40983525, PMID 12762436, PMID 28693043, PMID 30246358, PMID 34673157, PMID 33254926


    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.

  • 换个县还灵不灵:跨地质背景验证前,得先摆上桌的三个数

    Evidence level: B

    跨区验证最容易犯的错,是把”性能掉多少”当成问题;真问题是——那个被验证的”增量”,在另一个地质背景里还是不是同一个量。

    正文

    一个县里做出来的凶地评级增量,拿到花岗岩区、碳酸盐区、平原区各验一遍,看 ΔAUC 的符号是否一致——这个设计听起来很干净。但它跳过了三个更靠前的数。

    第一层,外部验证的门槛不是点数,是事件数。 34031906 给出二分类结局外部验证的最小样本量算法:目标是让校准(O/E 比与校准斜率)、判别(C 统计量)、临床净收益三项都达到预设精度。示例里事件率 0.018,算出来要 9835 人、177 个事件,而且这个数字是被”校准斜率”这一条顶上去的。33150684 是连续结局版本,四条准则(R²、calibration-in-the-large、校准斜率、结局方差)取最大者。回到凶地评级:一个县四十个点,超标率按 15% 算,事件数是 6 个。六个事件连校准斜率的区间都撑不开,ΔAUC 的区间宽度会盖住整个可能的效应范围。这不是验证失败,是没在验证。

    第二层,外部验证的先验是掉,不是涨。 25441703 复查 127 个新提出的风险模型:只有 32 个(25%)找得到至少一次外部验证,其中 22 个由完全不同的作者完成;AUC 相对推导中位下降 0.05,19 个模型降幅达到或超过 0.03,没有一个上升达到 0.03。所以判据不能写成”跨区 ΔAUC 显著为正”,得写成”跌幅落在 0 到 0.05 之间”。前者在要求数据推翻一个已知规律。

    第三层,判别与校准是两条独立的命。 36378565 把 43 个重症死亡模型拿去两个独立队列验证,只有 11 个(26%)能重建;能验的那几个,AUC 在 0.63 到 0.83 之间、可接受至优秀,但校准普遍差,重校准截距与斜率后才到中等。31842878 说得更直:校准是预测分析的阿喀琉斯之踵。跨区”方向一致”若只盯 AUC,漏掉的正是最要命的一半:同一个分数在两县代表的绝对风险完全不同。

    第四层,可分配的空间本身是地区属性。 19577346 用方差分析给出硬数字:英格兰与威尔士,填图基岩与表土地质解释室内氡总变异的约 25%;在沉积地质单元氡潜势反差强的地区升到 37%;而在铀矿化这类混杂控制横切地质界线的地区只有 14%。同一个国家、同一套变量、同一个方法,解释力从 14% 到 37%,差 2.6 倍。这意味着”评级还能再加多少”的上限在两地差几倍:ΔR²=0.04 在 37% 的区是锦上添花,在 14% 的区可能是仅剩的空间。跨区比 ΔR² 的大小,等于拿两把不同刻度的尺子量同一件事。

    第五层,变量重要性也随区漂移。 34863751 在意大利 Euganean Hills 用回归克里金建地生氡潜势图,结论是潜势主要由基岩类型驱动,但断层密度与地形控制氡在土体中的运移与地表呼出;33254926 在德国用 4448 个实测点加 40 km 空间块交叉验证,也是地质变量主导。两句合起来看:哪个变量”最重要”是区属性。于是同一个凶地评级,在 A 区可能是断层的粗代理,在 B 区是岩性的粗代理——符号一致或翻转,测的其实是”评级在两区替谁说话”。

    第六层,判一致性的方法不是分区各跑一遍。 11701102 用仿真量化过:在真实无差异的数据上,正式交互检验 5% 假阳性(正确),而按亚组分别做显著性判断会在 7% 到 66% 的情形下”发现”差异,且常报错差异的类型;交互检验功效普遍更好,但要稳定检出,交互大小需达到总效应两倍左右。正确结构是:一个模型加区×评级交互项,配 ICEMAN 评可信度(32778601),用 GRADE 36 区分”结果变异”与”研究特征变异”(36898507)。而交互检验的功效要求反过来说明:四个区各四十点,这个设计无论数据长什么样都得不出”方向一致”。

    第七层,若系数本身就是空间非平稳的,”一致”这个前提就错了。 41873991 用多尺度地理加权回归在全国尺度上看到显著的空间非平稳,全局模型把局部的系数差异抹平。所以该问的不是”是否处处一致”,而是”在哪个尺度上可迁移”。出路有一条实证:35614485 在多个医疗数据库间做迁移,融合集成比单库模型判别更好、跨库更一致,但校准在未见数据上一致地差,必须用新数据重校准。能搬的是排序,不能搬的是水平。

    落到判决上,四件事:① 先算事件数,按校准斜率反推每区最小样本量,达不到只做区内不下跨区结论;② 判据写成”跌幅在 0 至 0.05 之间”,不写”显著为正”;③ 报判别必报校准,两区各给校准曲线而不是只给一个 AUC 差;④ 一致性用”区×评级”交互项加 ICEMAN 打分,不用分区各跑显著性。

    易学这一侧,古人的规矩恰好是对着这件事说的。形峦派讲「一方山水一方龙」,又讲看地要「先看祖宗、再看父母」——祖宗是发脉的高山、父母是结穴的近身之山,同一条龙从祖宗到父母气脉连续,换个县的另一条山脉,祖宗父母全换,走法不能照搬。这不是保守,是经验上早发现了”同一套标准跨不过山脉”。更具体的落点是「山川有异,风气不同」——古书论形势从不给全国通用刻度,要求就地取材、随龙定穴。用统计话说:古人不假设系数平稳,只假设判据的结构(龙、穴、砂、水、向这五个位置有先后与主从)平稳。这正好对上第七层:能搬的是顺序,不能搬的是权重。

    证据与局限

    证据与局限

    已验证: 外部验证的最小样本量由校准精度主导,事件数而非点数决定可行性(34031906、33150684);外部验证的判别力普遍下降、几乎从不上升(25441703);判别可保持而校准常崩,重校准可部分修复(36378565、31842878);地质变量对室内氡的解释力在同一国家内从 14% 到 37% 不等(19577346);亚组内部分析的假阳性远高于交互检验(11701102);空间非平稳在区域尺度上普遍存在(41873991)。

    部分验证: 跨库集成可提升迁移后的判别一致性,但校准仍需重校(35614485);不同地区驱动变量排序不同(34863751、33254926)。

    待验证: 凶地评级的增量在多地质背景间是否方向一致;”可分配空间”差异是否足以解释跨区增量幅度的差异;把”局”操作化为可编码层之后,交互项能否检出。

    相关阅读

    证据等级: B(外部验证方法学、校准失效、地质解释力跨区差异均有 PubMed 直接支撑;”凶地评级增量的跨区方向一致性”本身无先例,属研究设计提案)
    参考文献: PMID 34031906, PMID 33150684, PMID 25441703, PMID 36378565, PMID 31842878, PMID 19577346, PMID 34863751, PMID 33254926, PMID 11701102, PMID 32778601, PMID 36898507, PMID 41873991, PMID 35614485


    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 it still work in a different county: Three numbers must be laid on the table before validating across geological backgrounds

    Evidence level: B

    The most common error in cross-regional validation is treating “how much performance drops” as the problem; the real issue is whether the validated “increment” remains the same quantity in a different geological background.

    Main Text

    An incremental risk rating derived from a single county, applied to granite, carbonate, and plain regions to check if the sign of ΔAUC remains consistent—this design sounds clean. However, it overlooks three more fundamental numbers.

    First Layer: The threshold for external validation is not sample size, but event count. 34031906 provides an algorithm for the minimum sample size required for external validation in binary outcomes: the goal is to ensure calibration (O/E ratio vs. calibration slope), discrimination (C-statistic), and clinical net benefit all meet preset precision levels. In the example, with an event rate of 0.018, the calculation yields 9,835 people and 177 events; this number is driven upward by the “calibration slope” criterion. 33150684 addresses continuous outcomes, using the maximum of four criteria (R², calibration-in-the-large, calibration slope, outcome variance). Returning to risk rating: a county with 40 points and an overage rate of 15% yields only 6 events. Six events cannot even span the interval for the calibration slope; the width of the ΔAUC interval would swallow the entire possible effect range. This is not a validation failure; it is a failure to validate.

    Second Layer: The prior for external validation is decline, not increase. 25441703 reviewed 127 newly proposed risk models: only 32 (25%) found at least one external validation, of which 22 were completed by authors entirely different from the original; AUC relative to the derivation median dropped by 0.05, and in 19 models the drop reached or exceeded 0.03, with none rising by 0.03. Therefore, the criterion cannot be written as “cross-regional ΔAUC significantly positive”; it must be written as “drop falls between 0 and 0.05.” The former demands data overturn a known pattern.

    Third Layer: Discrimination and calibration are two separate lives. 36378565 took 43 severe mortality models to two independent cohorts for validation; only 11 (26%) could be reconstructed. Among those validated, AUC ranged from 0.63 to 0.83 (acceptable to excellent), but calibration was generally poor, reaching only moderate levels after recalibrating intercept and slope. 31842878 states it more directly: calibration is the Achilles’ heel of predictive analysis. If cross-regional “directional consistency” focuses only on AUC, it misses the most critical half: the same score represents vastly different absolute risks in two counties.

    Fourth Layer: The allocatable space itself is a regional attribute. 19577346 used ANOVA to provide hard numbers: in England and Wales, mapping bedrock versus surface geology explained approximately 25% of the total variance in indoor radon; in sedimentary geological units with strong radon potential contrast, this rose to 37%; whereas in areas with uranium mineralization where mixed controls cross-cut geological boundaries, it was only 14%. Same country, same variables, same method—explanatory power varied from 14% to 37%, a 2.6-fold difference. This means the ceiling for “how much more rating can be added” differs by several times between regions: a ΔR² of 0.04 is icing on the cake in the 37% region but may represent the remaining space in the 14% region. Comparing cross-regional ΔR² magnitudes is like measuring the same thing with two rulers of different scales.

    Fifth Layer: Variable importance also drifts by region. 34863751 built a geogenic radon potential map using regression kriging in the Euganean Hills, Italy, concluding that potential is primarily driven by bedrock type, while fault density and topography control radon migration in soil and surface exhalation; 33254926 used 4,448 measured points plus 40 km spatial blocks for cross-validation in Germany, also finding geological variables dominant. Taken together: which variable is “most important” is a regional attribute. Thus, the same risk rating in Region A might be a coarse proxy for faults, while in Region B it is a coarse proxy for lithology—whether the sign remains consistent or flips, what is actually measured is “who the rating speaks for in each region.”

    Sixth Layer: The method to judge consistency is not running separate analyses per region. 11701102 quantified this via simulation: on truly null data, formal interaction testing yields a 5% false positive rate (correct), whereas performing significance tests separately by subgroup would “discover” differences in 7% to 66% of cases, often misreporting the type of difference. Interaction testing generally has better power, but to stably detect it, interaction size needs to be roughly double the main effect. The correct structure is: one model adding a region × rating interaction term, assessed for credibility using ICEMAN (32778601), and distinguishing “outcome heterogeneity” from “study characteristic heterogeneity” using GRADE 36 (36898507). The power requirements for interaction testing conversely demonstrate that with four regions of 40 points each, this design will never yield “directional consistency” regardless of data appearance.

    Seventh Layer: If coefficients themselves are spatially non-stationary, the premise of “consistency” is wrong. 41873991 observed significant spatial non-stationarity at a national scale using multi-scale geographically weighted regression, where global models flatten local coefficient differences. Therefore, the question should not be “is it consistent everywhere,” but “at what scale is it transferable.” One empirical path exists: 35614485 performed migration across multiple medical databases, finding that ensemble models showed better discrimination and cross-database consistency than single-database models, but calibration was consistently poor on unseen data and required recalibration with new data. What can be transferred is ranking; what cannot is level.

    In terms of judgment, four things must be done: ① First calculate event counts and back-calculate minimum sample sizes per region based on calibration slope; if not met, do not draw cross-regional conclusions; ② Write the criterion as “drop falls between 0 and 0.05,” not “significantly positive”; ③ Report both discrimination and calibration, providing calibration curves for each region rather than just a single ΔAUC; ④ Use “region × rating” interaction terms with ICEMAN scoring, avoiding separate significance tests per region.

    On the Yijing Studies side, ancient rules align perfectly with this matter. The Form School speaks of “One Landscape, One Dragon” and notes that site selection requires “Looking at Ancestors First, Then Parents“—Ancestors are the high mountains initiating the lineage, Parents are the close-range mountains forming the tomb site; if the dragon’s energy flows continuously from Ancestors to Parents, changing counties means a different mountain range where both Ancestors and Parents change, so the method cannot be copied blindly. This is not conservatism; it is an empirical realization that “a single set of standards cannot cross mountain ranges.” The more specific application is “Different Mountains, Different Winds“—ancient texts on form never provided a universal scale for the nation, demanding local materials and determining the tomb site following the dragon’s path. In statistical terms: ancients did not assume coefficient stationarity; they only assumed the structure of the criterion (the sequence and hierarchy of Dragon, Tomb, Sand, Water, and Direction) was stationary. This aligns perfectly with Layer Seven: what can be transferred is the sequence; what cannot is the weight.

    Evidence and Limitations

    Evidence and Limitations

    Verified: Minimum sample size for external validation is driven by calibration precision, determined by event count rather than point count (34031906, 33150684); discrimination generally declines in external validation with almost no increase (25441703); discrimination can be maintained while calibration often collapses, partially repairable via recalibration (36378565, 31842878); explanatory power of geological variables for indoor radon varies within the same country from 14% to 37% (19577346); subgroup internal analysis yields far higher false positives than interaction testing (11701102); spatial non-stationarity is common at regional scales (41873991).

    Partially Verified: Cross-database ensembles can improve post-migration discrimination consistency, but calibration still requires recalibration (35614485); driving variable rankings differ across regions (34863751, 33254926).

    To Be Verified: Whether the incremental risk rating maintains directional consistency across multiple geological backgrounds; whether differences in “allocatable space” sufficiently explain variations in cross-regional increment magnitude; whether interaction terms can detect effects after operationalizing the “local” as an encoded layer.

    Related Reading

    Evidence Level: B (External validation methodology, calibration failure, and cross-regional differences in geological explanatory power all have direct PubMed support; “cross-regional directional consistency of risk rating increments” itself has no precedent, belonging to a research design proposal)
    References: PMID 34031906, PMID 33150684, PMID 25441703, PMID 36378565, PMID 31842878, PMID 19577346, PMID 34863751, PMID 33254926, PMID 11701102, PMID 32778601, PMID 36898507, PMID 41873991, PMID 35614485


    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.

  • 「控制断层距离后评级还剩多少」:共线的时候,减法是三选一,而数据不告诉你选哪个

    Evidence level: B

    当两个变量相关到 r=0.9 时,「控制掉一个、看另一个还剩多少」不再是一个动作,而是三个互不相同的动作;而选择哪一个,数据本身没有发言权。

    正文

    一个很自然的问法:凶地评级和岩性、电导率这些物探变量,到底是不是同一件事?做法也顺手——把断层距离当协变量控制掉,再看评级和岩性的系数还显不显著。这一步看似只是「去掉一个干扰项」,实际已做了三个假设:断层在因果上位于上游、两个效应可线性相加、公共解释的那块方差可被切干净。三层没有一层白送。

    第一层,减法本身有三条路,给三个数。残差化(评级先对断层距离回归、取残差再对岩性回归)、联合模型(同时放进方程)、分层比较(先只放断层距离,再放岩性看 ΔR²)——三条路在完全共线时给出三个不同答案,而且都不算错。因为 R² 可分解成「独有 A + 独有 B + 共同」三块,25495950 用模拟数据把这套分解(commonality analysis)做了一遍,指出共同方差既不能归给 A 也不能归给 B,只能单列。所以「控制后还剩多少」的答案,取决于你把共同那块判给了谁。

    第二层,VIF 是放大器指标,不是信息量指标。有人用 20 名射手做多重目标追踪,年龄与训练年限 r=0.928、VIF=11.7,于是两个追踪指标变成了「显著的偏相关(r≈0.49,p<0.03)」,而训练年限的系数翻了号(B=-2.12,尽管零阶相关是 +0.65);换成只放年龄的模型(VIF<2.5),这些效应全部退到 p>0.13(42807114)。这就是共线性下的典型产物:一个被放大的标准误,加一个被误读成机制的反号系数。VIF 高只说明「这个系数不好读」,不说明「这个变量没有信息」。

    第三层,「共同方差」不是零也不是对半分。25495950 的结论是共同部分往往占大头,此时「独有贡献」数学上接近 0 是必然的,不是「岩性没用」。39280972 给了个约定式解法:按「平均共享方差」把总 R² 拆成可相加的个体 R²,在 GAM 里对付 concurvity。但那是约定不是发现——它让数字能加总,没让不可归属变得可归属。

    第四层,残差化在空间数据上更糟。取残差后拿它当因变量会产生生成回归量问题;更要紧的是空间结构:40705491 证明自变量与空间相关残差相关时,标准空间回归有偏、区间覆盖率不达标,须用两阶段(去空间趋势后再估参数)才拿得到名义覆盖率。断层距离和岩性都是空间强结构变量,残差也带着空间结构,于是「残差里的独立信号」很容易是没去干净的时空趋势。

    第五层,非线性与交互是默认,不是例外。断层距离对室内氡的效应是阈值型的:小于 150 m 的住宅有 22.8% 超过 200 Bq/m³,150–700 m 降到 15.2%;而只在富铀基岩上算,这两个数是 30.7% 和 17.5%(26630035)。注意后一对数字差距更大——说明断层距离与岩性之间有交互,不是可加的两项。把断层距离当线性协变量控制掉,等于把交互项硬塞进主效应,残差里剩下的「岩性信号」必然是脏的。30649165 给了同构证据:连续协变量的非线性被线性模型忽略时,估计会产生实质偏倚。

    第六层,「不显著」不等于「没有独立贡献」。共线性抬高标准误、压低功效,于是「控制后 p>0.05」在几十个点的样本里几乎必然发生,无论真值如何。42807114 那篇自己算了:要检出 f²=0.15 需 N≈68,实际样本是 20。这是设计的算术,不是证据。要判「没有独立贡献」,须做等效检验(先定最小有意义增量,如 ΔR²=0.02,再看区间是否整个落在界内),而不是看 p 值。另,把断层距离切成「近/远」两档是最常见的自伤——41802432 指出任意切点会造成信息损失与功效下降。

    第七层,特征重要性的排序在共线下不稳。41749845 在乳腺癌多组学数据上发现,不到 0.1% 的特征被移除,监督模型的 Top-20 排序就会翻盘;41616464 在 1055 种化学品的 QSAR 数据上给出更刺眼的对照:XGBoost 准确率最高(0.8569)但排序最不稳定,无监督方法准确率略低(0.819–0.849)却完全稳定。机制在 19460890 与 20187966 里写得很清楚:Gini 重要性在相关预测变量下有偏,置换重要性在零假设下无偏、但在相关存在时会把重要性摊给整组。所以「岩性的 SHAP 值高于电导率」这类话,在共线设计里不能当结论。

    这一层正好照出传统那套做法的结构。形峦派讲「龙、穴、砂、水、向」,从不把这五样当五个可互相控制的自变量——口诀「寻龙先看祖与宗,次看过峡与束气」讲的是先后与层级,不是并列与权重。龙定骨、砂定势、水定界,谁在上游谁在下游,古人心里有序次;现代统计把这种次序叫「因果层级」,把缺了次序的并列变量叫「共线」。所以这个问法古人不会提:他先问「这一处的龙是谁、砂是谁」,定了主从,再谈分量。分寸在于:那是可辩论的假设(各派对「龙」的定义都不同),不是已验证的因果图;把它当先验写进 DAG 再用数据检验,才是能往前走的用法。

    怎么办,四件事。不做残差化,做联合模型加 commonality 分解,报「独有评级 / 独有岩性 / 共同」三块数。断层距离按非线性进模型,并显式放入「断层距离 × 岩性」交互项。用空间块交叉验证、必要时用两阶段去空间趋势(40705491 的 DSR 思路),否则残差里的空间结构会把独立贡献喂大。特征重要性只报重抽样下的排序一致性,不报单次 SHAP 排名。

    证据与局限

    已验证:共线性会使系数标准误膨胀、系数可反号,这是统计学的确定结论,42807114 提供了一个 r=0.928、VIF=11.7 的完整现场记录。共同方差在方差分解中不可归属,25495950 与 39280972 从两个方向确认。相关预测变量下特征重要性排序不稳定,41749845、41616464、19460890、20187966 四篇独立复现。

    部分验证:断层距离对氡的阈值型效应有实测支撑(26630035,640 个地下室测点),但 150 m 这个切点是数据驱动的、不是物理常数;「断层距离与岩性存在交互」是从两组分层数字(22.8%/15.2% 与 30.7%/17.5%)反推的,原研究没有报告交互项检验。空间两阶段估计(40705491)是在方法学模拟中验证的,尚未在氡—地质场景里落地。

    待验证:commonality 分解在「评级—物探」这个具体变量组上的实际拆分比例,目前没有任何实测;「等效检验界值」取 0.02 还是 0.05,需要按决策后果定,本文给的是示例值而非标准。

    相关阅读

    证据等级: B
    参考文献: PMID 26630035, PMID 36497982, PMID 25495950, PMID 39280972, PMID 41749845, PMID 41616464, PMID 19460890, PMID 20187966, PMID 42807114, PMID 40705491, PMID 30649165, PMID 41802432, PMID 36323796


    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。

    How Much Rating Remains After Controlling for Fault Distance: When Collinearity Occurs, Subtraction Offers Three Choices But Data Reveals None

    Evidence level: B

    When two variables correlate at $r=0.9$, the question of “controlling for one to see how much remains with the other” ceases to be a single action and becomes three mutually distinct actions; furthermore, the data itself holds no authority on which choice to make.

    Main Text

    A natural inquiry arises: Is the rating of a “bad site” identical to geological properties and geophysical variables such as electrical conductivity? The practical approach is straightforward: treat fault distance as a covariate to control for, then examine whether the coefficient linking rating to lithology remains significant. This step appears merely as “removing one confounder,” yet it implicitly assumes three conditions: that the fault lies upstream in causality, that the two effects are linearly additive, and that the shared variance can be cleanly partitioned. None of these layers is free of assumption.

    First, subtraction itself offers three pathways yielding three numbers: residualization (regressing rating against fault distance first, taking the residuals, then regressing against lithology), joint modeling (simultaneously including both in the equation), or stratified comparison (first modeling only fault distance, then adding lithology to observe $\Delta R^2$). Under perfect collinearity, these three paths yield different answers, none of which are inherently “wrong.” Because $R^2$ decomposes into “Unique A,” “Unique B,” and “Shared” components, 25495950 utilized simulation data to perform a commonality analysis demonstrating that shared variance cannot be attributed solely to A or B but must be listed separately. Therefore, the answer to “how much remains after controlling” depends entirely on how one assigns the shared component.

    Second, VIF is an amplification indicator, not an information quantity indicator. Consider a study with 20 archers tracking multiple targets: age and training duration correlate at $r=0.928$ with a VIF of 11.7. Consequently, two tracking indicators become “significant partial correlations ($r\approx0.49$, $p<0.03$),” while the coefficient for training duration flips sign ($B=-2.12$, despite a zero-order correlation of +0.65); conversely, in a model containing only age (VIF<2.5), these effects retreat to $p>0.13$ (42807114). This represents a typical artifact under collinearity: an inflated standard error paired with a misinterpreted reversed-sign coefficient. A high VIF merely indicates that “this coefficient is difficult to interpret,” not that “this variable lacks information.”

    Third, “shared variance” is neither zero nor exactly half. 25495950 concludes that the shared component often dominates; thus, when “unique contribution” mathematically approaches zero, it is a necessity, not evidence that “lithology is useless.” 39280972 offers a conventional solution: decomposing total $R^2$ into additive individual $R^2$s based on “average shared variance” to handle concurvity in GAMs. However, this is convention rather than discovery—it allows numbers to sum but does not render the unattributable attributable.

    Fourth, residualization performs worse with spatial data. Using residuals as the dependent variable introduces endogeneity issues; more critically, regarding spatial structure: 40705491 demonstrates that when independent variables correlate with spatial residuals, standard spatial regression yields biased estimates and fails to meet nominal coverage rates, requiring a two-stage approach (removing spatial trends before estimating parameters) to achieve nominal coverage. Since both fault distance and lithology are variables with strong spatial structure, their residuals also carry spatial structure; thus, the “independent signal” remaining in the residuals is likely contaminated by unremoved spatiotemporal trends.

    Fifth, nonlinearity and interaction are defaults, not exceptions. The effect of fault distance on indoor radon is threshold-based: residences within 150 m show 22.8% exceeding 200 Bq/m³, dropping to 15.2% between 150–700 m; however, calculating only on uranium-rich bedrock yields 30.7% and 17.5%, respectively (26630035). Note the larger gap in the latter pair—indicating interaction between fault distance and lithology, not merely additive terms. Treating fault distance as a linear covariate to control for is equivalent to forcing an interaction term into the main effect; the remaining “lithology signal” in the residuals will inevitably be contaminated. 30649165 provides isomorphic evidence: ignoring nonlinearity in continuous covariates via linear models produces substantial bias in estimates.

    Sixth, “non-significance” does not equate to “no independent contribution.” Collinearity inflates standard errors and reduces power, making “post-control $p>0.05$” nearly inevitable in samples of dozens of points regardless of the true value. 42807114 calculated this internally: detecting $f^2=0.15$ requires $N\approx68$, yet the actual sample size is 20. This is arithmetic of design, not evidence. To judge “no independent contribution,” one must perform an equivalence test (first defining a minimum meaningful increment, e.g., $\Delta R^2=0.02$, then checking if the confidence interval falls entirely within bounds), rather than relying on p-values. Furthermore, dichotomizing fault distance into “near/far” is a common self-harm—41802432 notes that arbitrary cutpoints cause information loss and reduced power.

    Seventh, ranking feature importance is unstable under collinearity. 41749845 found in breast cancer multi-omics data that removing less than 0.1% of features can flip the Top-20 ranking in supervised models; 41616464 provided a more striking contrast in QSAR data for 1055 chemicals: XGBoost achieved the highest accuracy (0.8569) but showed the least stability in ranking, whereas unsupervised methods had slightly lower accuracy (0.819–0.849) yet were completely stable. The mechanism is clearly outlined in 19460890 and 20187966: Gini importance is biased under correlated predictors, while permutation importance is unbiased under the null hypothesis but distributes importance across the group when correlation exists. Thus, statements like “lithology’s SHAP value exceeds electrical conductivity” cannot be treated as conclusions in collinear designs.

    This layer precisely mirrors the structural approach of traditional methodologies. The Xiang Luan school discusses “Dragon, Cave, Sand, Water, and Direction,” never treating these five as five independent variables mutually controllable. The mnemonic “Seek the Dragon first by looking at its ancestors and descendants, then examine the narrow passes and qi-binding” refers to sequence and hierarchy, not parallelism and weighting. The Dragon sets the bone, Sand sets the posture, Water sets the boundary; ancient masters understood the upstream-downstream order within their minds; modern statistics labels such ordering as “causal hierarchy,” while treating collinear variables lacking such order as “collinearity.” Therefore, this specific question would not arise in traditional thought: one first asks “Who is the Dragon here? Who is the Sand?” establishing primacy and subordination before discussing proportions. The nuance lies in the fact that this is a debatable hypothesis (different schools define “Dragon” differently), not a verified causal graph; incorporating it as a prior into a DAG and then testing with data is the only way to proceed forward.

    What to do? Four things: Avoid residualization; use joint modeling with commonality decomposition, reporting the three components (“Unique Rating,” “Unique Lithology,” “Shared”). Model fault distance non-linearly and explicitly include an interaction term (“Fault Distance × Lithology”). Use spatial block cross-validation, employing the two-stage spatial trend removal (DSR approach from 40705491) when necessary, otherwise spatial structure in residuals will inflate independent contributions. Report feature importance only based on ranking consistency under resampling, not single-run SHAP rankings.

    Evidence and Limitations

    Verified: Collinearity causes coefficient standard errors to inflate and coefficients to flip signs; this is a definitive statistical conclusion, supported by 42807114 which provides a complete field record with $r=0.928$ and VIF=11.7. Shared variance is unattributable in variance decomposition, confirmed by 25495950 and 39280972 from two directions. Feature importance rankings are unstable under correlated predictors, independently replicated across four studies: 41749845, 41616464, 19460890, and 20187966.

    Partially Verified: Threshold effects of fault distance on radon have empirical support (26630035, 640 basement measurement points), though the 150 m cutpoint is data-driven rather than a physical constant; the existence of interaction between “fault distance and lithology” is inferred from two stratified sets of numbers (22.8%/15.2% vs 30.7%/17.5%), but the original study did not report an interaction term test. The two-stage estimation method (40705491) was verified through methodological simulations and has not yet been implemented in the radon-geological scenario.

    To Be Verified: The actual proportion of splitting via commonality decomposition for the specific variable group “Rating vs. Geophysical” currently lacks empirical data; whether the “equivalence test threshold” should be 0.02 or 0.05 requires determination based on decision consequences; the values provided in this text are examples rather than standards.

    Related Reading

    Evidence Level: B
    References: PMID 26630035, PMID 36497982, PMID 25495950, PMID 39280972, PMID 41749845, PMID 41616464, PMID 19460890, PMID 20187966, PMID 42807114, PMID 40705491, PMID 30649165, PMID 41802432, PMID 36323796


    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.

  • 同一条山脊上四十个点,不算四十个样本:凶地评级模型的过拟合边界

    Evidence level: B

    上一个模型说”把凶地评级和四个物探变量一起喂给机器学习”。这一篇问的是那台机器最少要喂多少个点才不说谎——答案是:先看你数的是点,还是山。

    正文

    凶地评级模型的设计图纸已经有了:风水师盲法评级做因变量,断层距离、岩性、土壤氡、地表电导率做自变量,训练一个能预测”凶不凶”的模型。剩下最硬的一关是样本量——用四十个点评级、四个变量,模型能不能站住?直觉会给出一个比例,比如”一个变量至少十个样本”。这个直觉在这一条线上会连错三次。

    第一错:横轴画错了。 四十个评级点如果落在同一条山脊、同一个村、同一片花岗岩体上,它们不是四十个独立观测。德国做全国地生氡潜势图时手上有 4448 个实测点,仍然不敢按随机方式切分,而是把数据按 40 公里见方的空间块轮流留出来做交叉验证(PMID 33254926)——因为相邻点的自变量和因变量都高度相关,随机切分等于让模型在考场上遇见复习过的题。四十个点如果聚成三簇,有效样本量可能只有个位数。所以”样本量-特征数比”这条曲线的横轴不该是点数,而该是去除空间相关后的有效点数,这个数通常比表格里的行数小一个量级。

    第二错:交叉验证会给你假安慰。 用交叉验证的误差去挑模型参数,再用同一个误差报性能,这是双重使用。Varma 与 Simon 用完全无信号的数据(两类特征分布相同)做过这件事:优化后的分类器交叉验证误差低于 30% 的数据集占 18.5%(收缩质心)和 38%(支持向量机),而它们在独立测试集上的表现与抛硬币无异;只有嵌套交叉验证能给出接近无偏的估计(PMID 16504092)。更麻烦的是,在合并的训练加测试数据上做特征筛选,制造的偏倚比调参更大,而且样本量到 1000 时 K 折交叉验证的偏倚仍然存在(PMID 31697686)。四十个点的规模,正好落在偏倚最严重的那一档。

    第三错:变量是风水师圈出来的。 四个物探变量不是预先登记的,而是先看过评级、觉得”这几个大概有关”才纳入的。这就是逐步选择。GUSTO-I 试验(40830 例)的仿真显示:每变量 3 个事件时,8 变量模型里有 7 个变量的回归系数被高估超过 25%;每变量 10 个事件时有 3 个;而完全不做筛选时全部低于 20%(PMID 10513756)。筛选这个动作本身,就会把系数推向”看起来更显著”。

    “一个变量十个样本”这条规则本身也站不住。 支持它的三项仿真里只有一项支持(PMID 27881078),另一项大样本仿真直接表明每变量事件数与预测性能(校准、区分度、概率误差)没有强关系,真正决定样本需求的是变量数、总样本量与事件比例三者(PMID 29966490)。Riley 等人给的三准则——收缩因子不低于 0.9、表观与校正后的 R² 差不超过 0.05、总体风险估得够准——同一个方法算出来,一个诊断模型只需每变量 4.8 个事件,一个复发静脉血栓模型要 23 个(PMID 30357870)。差四倍。所以正确做法不是查表,而是先说出你预期这个模型能解释多少方差,再反算样本量。

    这就逼出本文最锋利的一刀:天花板已经被别人量过了。 法国做过一万多户室内氡测量,把地生氡潜势、建筑材料、建造年份、基础类型、楼层、生活习惯全部放进去,模型只解释了氡浓度对数值总变异的约 20%(PMID 24056050)。罗马尼亚 100 个案例的研究也说,地质基础能”合理解释”多数记录值——但这是定性表述,不是方差比例(PMID 31897872)。而爱尔兰那篇报”能解释约 75% 变异”的模型,解释的是地生氡潜势这个由土壤氡与渗透率合成的中间量,不是室内氡(PMID 37343860)。这两个数字经常被混着引用,但它们隔着一整层建筑与通风。推论很直接:如果凶地评级模型报出 R² 大于 0.4,它解释的几乎一定不是室内氡,而是它自己泄漏进来的某个中间量。

    所以判据不能靠经验比例,得靠零信号基线。 把结局标签随机打乱五十次,每次重跑整条流程,取平均 AUC;凡超过 0.5 的部分全是过拟合。用这个方法实测,要让仅仅 10 个特征的 AUC 膨胀控制在 0.05 以内,样本量至少要 200(PMID 38614825)。影像组学的领域共识也是同一条:数据驱动建模必须预注册分析路径、必须有独立数据验证(PMID 33492473)。凶地评级这条线,应当先跑这个置换基线,再说自己的 R² 有多少是真的。

    最后,该报的不是 R²,是增量。 只喂四个物探变量的模型是一个必须存在的基线;凶地评级模型的全部价值在于它比这个基线多解释了多少(ΔR²),以及这个增量的置信区间是否跨零。样本量曲线也该这样画:固定特征数,扫样本量,看 ΔR² 什么时候走平。现成的数字范式可以参考:随机森林在 500 例以下严重过拟合,1000 例配 3 个特征勉强可比,20 个以上特征就再度过拟合,3000 例以上才稳定(PMID 42394051)。四十个点配四个变量,落在哪一格,不用猜。

    证据与局限

    已验证: 空间块交叉验证可抑制空间自相关造成的性能虚高(PMID 33254926);用同一交叉验证误差既调参又报性能会显著低估真实误差,嵌套交叉验证近无偏(PMID 16504092);小样本下 K 折交叉验证偏倚在 n=1000 仍存在,且特征筛选的偏倚贡献大于调参(PMID 31697686);逐步选择会系统性高估入选变量系数(PMID 10513756);每变量事件数规则缺乏一致依据(PMID 27881078、29966490),最小样本量须按预期 R² 反算且结果可差四倍(PMID 30357870);地质与建筑变量合计只解释室内氡对数变异约 20%(PMID 24056050);结局置换可量化过拟合幅度,200 例方能承载 10 个特征(PMID 38614825)。

    部分验证: “凶地评级”作为因变量的可行性尚未有实测数据,其与四个物探变量的关联强度未知;有效样本量的折算方法(如按空间块数计)在本领域尚无统一标准,只有邻域借鉴。

    待验证: 四十点/四变量配置下置换基线 AUC 的实际膨胀量;ΔR² 相对纯物探基线的点估计与区间;跨地区(岭南/江南/华北)时样本需求是否随地质复杂度上升;评级点位聚簇程度与有效样本量的定量关系曲线。

    相关阅读

    证据等级: B(过拟合机制、样本量准则、方差天花板均有 PubMed 直接支撑;”凶地评级作为因变量”本身仍属研究设计提案,无实测)
    参考文献: PMID 33254926, PMID 16504092, PMID 31697686, PMID 10513756, PMID 27881078, PMID 29966490, PMID 30357870, PMID 24056050, PMID 37343860, PMID 31897872, PMID 38614825, PMID 33492473, PMID 42394051


    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。

    Forty Points on the Same Ridge Do Not Count as Forty Samples: The Overfitting Boundary in凶地 Rating Models

    Evidence level: B

    The previous model proposed “feeding the machine learning algorithm the adverse land rating alongside four geophysical variables.” This article asks: how many data points must be fed into that machine before it stops lying? The answer depends on whether you are counting points or mountains.

    Main Text

    The blueprint for the adverse land rating model is already in place: Feng Shui blind ratings serve as the dependent variable, while fault distance, lithology, soil radon, and surface electrical conductivity serve as independent variables. The goal is to train a model capable of predicting “adverse or not.” The most critical remaining hurdle is sample size: can a model stand firm with forty ratings and four variables? Intuition suggests a ratio, such as “at least ten samples per variable.” However, this intuition fails three times along this specific line.

    First Error: The X-axis is drawn wrong. If forty rating points fall along the same ridge, in the same village, or within the same granite body, they are not forty independent observations. When Germany produced a national map of terrestrial radon potential, despite having 4,448 measured points, they dared not split the data randomly. Instead, they reserved blocks of 40×40 km for cross-validation (PMID 33254926) because the independent and dependent variables are highly correlated between adjacent points. Random splitting is equivalent to giving the model a review of past exam questions on the same day. If forty points cluster into three groups, the effective sample size may be only single-digit. Therefore, the x-axis of the “sample size-to-feature ratio” curve should not be the raw number of points, but rather the effective number of points after removing spatial correlation, a figure typically an order of magnitude smaller than the row count in a table.

    Second Error: Cross-validation gives false comfort. Using cross-validation error to tune model parameters and then reporting performance based on the same error constitutes double dipping. Varma and Simon demonstrated this using completely noise-free data (where feature distributions are identical across classes): optimized classifiers showed cross-validation errors below 30% in 18.5% of datasets (shrinkage centroid) and 38% (support vector machines), yet their performance on independent test sets was no better than coin flipping; only nested cross-validation provided estimates close to unbiased (PMID 16504092). Complicating matters further, performing feature selection on the combined training and testing data creates bias larger than hyperparameter tuning, and even with K-fold cross-validation, bias persists when sample sizes reach 1,000 (PMID 31697686). A scale of forty points falls squarely into the category where bias is most severe.

    Third Error: Variables are selected by the Feng Shui circle. The four geophysical variables were not pre-registered; they were included only after reviewing the ratings and concluding “these seem related.” This is stepwise selection. A simulation from the GUSTO-I trial (4,083 cases) showed that with 3 events per variable, 7 of the 8 variables in a model had regression coefficients overestimated by more than 25%; with 10 events per variable, 3 were affected; whereas without any selection, all remained below 20% (PMID 10513756). The act of selection itself pushes coefficients toward “appearing more significant.”

    The rule of “ten samples per variable” also stands unsupported. Of the three simulations supporting it, only one does (PMID 27881078); another large-scale simulation directly indicates that the number of events per variable has no strong relationship with predictive performance (calibration, discrimination, probability error). What truly determines sample requirements are the number of variables, total sample size, and event proportion (PMID 29966490). Riley et al. provided three criteria—shrinkage factor not below 0.9, difference between apparent and corrected R² not exceeding 0.05, and accurate overall risk estimation—where the same method yielded a requirement of only 4.8 events per variable for a diagnostic model, but 23 for a recurrent venous thromboembolism model (PMID 30357870). That is a fourfold difference. Therefore, the correct approach is not to consult a table, but to first state how much variance you expect this model to explain, then back-calculate the sample size.

    This forces the sharpest blade of this article: the ceiling has already been measured by others. France conducted indoor radon measurements in over 10,000 households, incorporating terrestrial radon potential, building materials, construction year, foundation type, floor level, and lifestyle habits; the model explained only about 20% of the total variance in radon concentration (PMID 24056050). A Romanian study of 100 cases also stated that geological foundations could “reasonably explain” most recorded values—but this is a qualitative statement, not a variance proportion (PMID 31897872). Meanwhile, the Irish study reporting a model that explains “about 75% of the variance” refers to terrestrial radon potential, an intermediate variable synthesized from soil radon and permeability, not indoor radon (PMID 37343860). These two figures are frequently cited together, yet they are separated by an entire layer of building and ventilation. The inference is direct: if the adverse land rating model reports an R² greater than 0.4, it almost certainly does not explain indoor radon, but rather some intermediate variable that leaked into the model.

    Therefore, the criterion cannot rely on empirical ratios but must depend on a zero-signal baseline. Randomly shuffle the outcome labels fifty times, re-run the entire pipeline each time, and take the average AUC; any portion exceeding 0.5 indicates overfitting. Applying this method empirically, to keep the AUC inflation of a model with only 10 features within 0.05, the sample size must be at least 200 (PMID 38614825). The field consensus in radiomics is the same: data-driven modeling must have a pre-registered analysis path and independent data validation (PMID 33492473). For the adverse land rating line, one should first run this permutation baseline before claiming how much of the R² is genuine.

    Finally, what should be reported is not R², but the increment. A model fed only with four geophysical variables constitutes a necessary baseline; the entire value of the adverse land rating model lies in how much additional variance it explains compared to this baseline (ΔR²) and whether the confidence interval of this increment crosses zero. The sample size curve should also be drawn this way: fix the number of features, sweep the sample size, and observe when ΔR² flattens. Existing numerical paradigms can be referenced: random forests suffer severe overfitting below 500 cases, are barely comparable with 3 features at 1,000 cases, become severely overfit again with more than 20 features, and only stabilize above 3,000 cases (PMID 42394051). With forty points and four variables, one does not need to guess which grid it falls into.

    Evidence and Limitations

    Verified: Spatial block cross-validation suppresses inflated performance caused by spatial autocorrelation (PMID 33254926); using the same cross-validation error for both tuning and reporting significantly underestimates true error, while nested cross-validation is nearly unbiased (PMID 16504092); K-fold cross-validation bias persists at n=1,000 in small samples, and feature selection contributes greater bias than tuning (PMID 31697686); stepwise selection systematically overestimates coefficients of selected variables (PMID 10513756); the rule of events per variable lacks consistent basis (PMID 27881078, 29966490), and minimum sample size must be back-calculated based on expected R² with results potentially differing by fourfold (PMID 30357870); geological and building variables combined explain only about 20% of the log-variance in indoor radon (PMID 24056050); outcome permutation quantifies overfitting magnitude, requiring 200 cases to support 10 features (PMID 38614825).

    Partially Verified: The feasibility of “adverse land rating” as a dependent variable lacks empirical data; the strength of association between it and the four geophysical variables is unknown; methods for calculating effective sample size (e.g., counting spatial blocks) lack a unified standard in this field, relying only on neighborhood borrowing.

    To be Verified: The actual inflation of permutation baseline AUC with forty points and four variables; the point estimate and interval for ΔR² relative to a pure geophysical baseline; whether sample requirements rise with geological complexity across regions (Lingnan/Jiangnan/Northern China); the quantitative relationship curve between rating point clustering and effective sample size.

    Related Reading

    Evidence Level: B (Overfitting mechanisms, sample size criteria, and variance ceilings have direct PubMed support; “adverse land rating as a dependent variable” remains a research design proposal without empirical data)
    References: PMID 33254926, PMID 16504092, PMID 31697686, PMID 10513756, PMID 27881078, PMID 29966490, PMID 30357870, PMID 24056050, PMID 37343860, PMID 31897872, PMID 38614825, PMID 33492473, PMID 42394051


    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.

  • 疼了九个月,是热还没散,还是免疫已经改了?

    Evidence level: B

    “短期有效、长期也有效”不是同一个东西量了两次,是两条衰减速度差两个数量级的曲线叠在一起——分不开它们,谈机制就是猜。

    正文

    先给结论:氡浴研究里被反复引用的两个读数——疗程结束时疼痛下降、九个月后还有优势——从来不是同一件事的两个快照。它们是两条曲线叠在一张图上:一条以小时为半衰期,一条以月为半衰期。当成一条曲线读,必然得出”疗效在缓慢消退”的错觉;真正的任务是把两条曲线的衰减常数各自拟合出来。

    第一层:热量的是小时级的钟。 热疗镇痛起效快、消退也快,这是常识,但”多快”一直没被严肃拟合过。方法学上有个提醒:评估热适应的时程不能靠固定时点前后比对,因为基线选择与衰减模型形状会直接改变结论(PMID 38872315)。同一个坑在热镇痛上更深——单次热暴露的效应可能一天内归零大半,而你只在疗程前后各测一次,中间那段衰减完全看不见。

    第二层:免疫量的是周级的钟。 小鼠吸入氡 2,000 Bq/m³,2 周时免疫指标没动,4 周才激活 Th1/Th2/Th17,且三者平衡未被打破,同时脾脏 SOD 降、CAT 升(PMID 35400822)。”2 周不动、4 周才动”的分量在于:免疫通道有启动延迟,而临床疗程普遍只有 3 到 4 周,正好压在起效边缘。更麻烦的是方向:同一实验里正常鼠被激活、关节炎模型鼠(SKG)被拉回正常——调控是双向的,直接否掉了”某个标志物越高越好”的追踪思路。另一条 K/BxN 模型给出的信号又换了对象:临床评分改善伴随外周血 B 细胞与 IL-5 升高,不是 Treg(PMID 35203348)。不同模型指向不同细胞,这本身就是”通道还没定”的证据。

    第三层:一个算术就能定生死。 设热效应半衰期 1 天、免疫效应半衰期 90 天。停疗第 30 天,热残余是 0.5³⁰,约十亿分之一,等于零;免疫残余 0.5^(30/90),还有 79%。第 180 天免疫剩 25%,第 270 天剩 12.5%。IMuRa 试验 681 人随访九个月,疼痛与镇痛药用量仍显著优于对照(p=0.032 / p=0.007,PMID 23864139)——九个月时热残余已是十亿分之一,”还显著”本身就把热效应排除了,不需要任何中介分析。疗程结束那刻测到的总效应里含热的大量份额,这部分本来就不该留到第九个月。效应从 0.7 个标准差掉到 0.2,正确读法是”热的那部分按时退场”,不是”疗效在衰减”。

    第四层:最锋利的一刀。 83 名强直性脊柱炎患者在 Bad Gastein 接受 3 至 4 周含低剂量氡的矿坑疗法,活性 TGF-β1 从 77 升到 1,096 pg/ml;按疼痛是否缓解分组,应答者升 17 倍(96→1,654),非应答者升 7 倍(53→402,PMID 16810484)。非应答者也涨了七倍。若 TGF-β1 真是”氡→免疫→镇痛”链条上的中介,疼痛没改善的人不该出现同样方向。它涨了,说明它更像伴随反应,不是传话的那一环。把免疫标志物当替代终点,会在这里翻车。

    第五层:双臂设计根本拆不开。 “氡浴 vs 热浴”看起来干净,但氡必须靠热浴递送——热不是可剔掉的对照,而是氡的载体。当一个因子的水平依赖另一因子的存在,两者是交互,不是中介。中介分析假设路径可加,交互则明确说不可加。用双臂数据做中介,等于默认了正在被质疑的前提。

    第六层:可拆的结构已经出现了。 RAD-ON02 试验里,热浴与氡浴都减少了破骨细胞分化与骨片吸收面积,氡浴更明显;但只有氡浴组出现 RANKL 显著下降与 Treg 显著上升,Th17 不变(PMID 38292488)。正确表述是“共有的热通道 + 氡独有的免疫增量”,不是”氡有效、热无效”。要检验它,得用热×氡的析因结构,只有析因才能把主效应与交互项分开估。

    第七层:真要做中介,规矩是先说清时点。 报告规范要求写出因果假设、中介与结局各自的测量时点、以及两者间是否还有未测混杂(PMID 34546296);未测混杂的敏感性可用贝叶斯方法给出”要多强的混杂才能把间接效应抹平”的定量边界(PMID 28882092)。25 人的膝骨关节炎试验给了盆冷水:加氡与不加氡,两组疼痛都持续下降、内源性大麻素 anandamide 都下降,氡无额外效应(PMID 33649972)——效应量可能小到中介分析根本做不出来。2026 年综述也承认,方案异质性与热、CO₂ 等共模因素让归因至今不成立(PMID 42340546)。

    判决四件事:① 中介变量必须多点测(前、中、疗程末、随访),只测一次等于只测一个点;② “共有通道”与”氡增量”分开报,并报交互项;③ 免疫标志物不得当替代终点,非应答者也涨七倍就是反例;④ 先拟合衰减常数再看 p 值——两个指数项叠成的曲线,单点比较永远读错。

    易学侧的一次对齐

    中医讲”急则治标、缓则治本”,通常被理解成两种治法、两套药。放到这个场景里,它其实是在描述同一种干预在两个时间尺度上的两种读数:热是”标”,来得快、去得也快;免疫调节是”本”,启动慢、退得也慢。《素问·至真要大论》有一句更狠的话——”久而增气,物化之常也;气增而久,天之由也”,说的是同一种刺激长期同向施加,会从”调节”滑向”偏性”。这句话和上面第二层的双向调控是同一个意思:正常鼠被激活、病鼠被拉回正常,说明系统要的是回到中间,不是一路加码。古人对”疗程越长越好”的警惕,不是玄学禁忌,是对”调节系统在长期同向输入下会失去调节能力”的观察。

    证据与局限

    已验证:氡浴对疼痛的效应可持续至九个月(IMuRa,681 人);氡浴组 RANKL 下降与 Treg 上升为氡特有,热浴无此效应(RAD-ON02);含氡矿坑疗法后活性 TGF-β1 显著上升且应答者升幅更大;小鼠免疫激活需 4 周、2 周不显现。

    部分验证:热效应半衰期、免疫效应半衰期均未在氡浴人群中被直接拟合,本文的双指数算术用的是量级假设而非实测参数;TGF-β1 是否参与镇痛链条,仅有相关性而无中介检验。

    待验证:① 用限制性设计把两条衰减曲线拆开的实测;② 热×氡析因试验中交互项的方向与大小;③ Treg 在停疗后的回落时程,决定”长期效应”能否由它承载。

    相关阅读

    证据等级: B
    参考文献: PMID 23864139, PMID 38292488, PMID 16810484, PMID 35400822, PMID 35203348, PMID 33649972, PMID 34546296, PMID 28882092, PMID 38872315, PMID 42340546


    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。

    Nine Months of Pain: Is It Still Heat or Has the Immune System Changed?

    Evidence level: B

    “Short-term efficacy and long-term efficacy” are not two snapshots of the same variable measured twice; they represent two decay curves differing by two orders of magnitude superimposed on one another. Without disentangling them, discussing mechanisms amounts to speculation.

    Main Text

    Conclusion First: In radon bath research, the two repeatedly cited metrics—the reduction in pain at the end of the course and the advantage persisting for nine months—are never two snapshots of the same event. They represent two curves overlaid on a single graph: one with an hourly half-life, the other with a monthly half-life. Reading them as a single curve inevitably leads to the illusion that “efficacy is slowly fading”; the true task is to fit the decay constants for each curve separately.

    Layer 1: Heat Operates on an Hourly Clock. Thermal therapy acts quickly and wears off quickly; this is common sense, but “how quickly” has rarely been rigorously modeled. Methodologically, a caution is warranted: assessing the timeline of thermal adaptation cannot rely on pre-post comparisons at fixed time points, as baseline selection and decay model shapes directly alter conclusions (PMID 38872315). This pitfall is deeper in thermal analgesia—the effect of a single thermal exposure may vanish by half within a day, yet measurements are often taken only before and after the course, rendering the intermediate decay invisible.

    Layer 2: Immunity Operates on a Weekly Clock. In mice inhaling 2,000 Bq/m³ radon, immune markers remained unchanged at 2 weeks but were activated at 4 weeks (Th1/Th2/Th17 balance remained intact), while splenic SOD decreased and CAT increased (PMID 35400822). The significance of “no change at 2 weeks, activation at 4 weeks” lies in the fact that immune pathways have a latency delay, while clinical courses typically last only 3 to 4 weeks, precisely straddling the onset threshold. Complicating matters further is the direction: in the same experiment, normal mice were activated while arthritis model mice (SKG) were pulled back to normal—regulation is bidirectional, directly negating the tracking logic that “higher levels of a specific marker are always better.” Another signal from the K/BxN model shifts focus: clinical score improvement accompanied by increased peripheral blood B cells and IL-5, not Tregs (PMID 35203348). Different models point to different cell types; this itself is evidence that the pathway is not yet defined.

    Layer 3: One Arithmetic Calculation Can Determine Survival. Assume a half-life of 1 day for the thermal effect and 90 days for the immune effect. On day 30 post-treatment, thermal residual is $0.5^{30}$, approximately one billionth, effectively zero; immune residual is $0.5^{(30/90)}$, remaining at 79%. By day 180, immunity remains at 25%; by day 270, it drops to 12.5%. The IMuRa trial followed 681 patients for nine months, with pain and analgesic usage still significantly superior to controls (p=0.032 / p=0.007, PMID 23864139)—at nine months, thermal residual is already one billionth; the mere fact that it remains “significant” excludes the thermal effect without requiring any mediation analysis. The total effect measured at the end of the course contains a substantial share attributable to heat; this portion should not be expected to persist until the ninth month. The drop from 0.7 standard deviations to 0.2 should be correctly read as “the thermal component exiting on schedule,” not “efficacy is decaying.”

    Layer 4: The Sharpest Cut. In 83 patients with ankylosing spondylitis treated at Bad Gastein for 3 to 4 weeks with low-dose radon mineral spring therapy, active TGF-β1 rose from 77 to 1,096 pg/ml. Stratifying by pain response, responders showed a 17-fold increase (96→1,654), while non-responders showed a 7-fold increase (53→402, PMID 16810484). Non-responders also increased seven-fold. If TGF-β1 were truly a mediator in the “radon → immunity → analgesia” chain, those without pain improvement should not show a similarly directional increase. Since it did rise, it suggests it is more of a concomitant response than the messenger link. Treating immune markers as surrogate endpoints will lead to failure here.

    Layer 5: The Two-Arm Design Cannot Be Disassembled. The “radon bath vs. thermal bath” design appears clean, but radon must be delivered via thermal bath—heat is not a removable control but the carrier for radon. When the level of one factor depends on the existence of another, they constitute an interaction, not a mediation. Mediation analysis assumes additivity of paths; interaction explicitly states non-additivity. Performing mediation analysis on two-arm data defaults to the premise being questioned.

    Layer 6: A Disassemblable Structure Has Emerged. In the RAD-ON02 trial, both thermal and radon baths reduced osteoclast differentiation and bone resorption area, with radon showing a more pronounced effect; however, only the radon bath group showed significant RANKL reduction and significant Treg increase, with no change in Th17 (PMID 38292488). The correct formulation is “shared thermal channel + radon-specific immune increment,” not “radon effective, thermal ineffective.” To test this, a factorial design of heat × radon is required; only factorial analysis can separate main effects from interaction terms.

    Layer 7: True Mediation Requires Clarifying Timing First. Reporting standards require specifying causal hypotheses, measurement time points for mediators and outcomes, and whether unmeasured confounding exists between them (PMID 34546296); sensitivity to unmeasured confounding can be quantified using Bayesian methods to define the strength of confounding needed to wash out the indirect effect (PMID 28882092). A 25-person knee osteoarthritis trial offered a cold water basin: adding radon or not, pain declined continuously in both groups and endogenous anandamide decreased in both, with no additional effect from radon (PMID 33649972)—the effect size may be too small for mediation analysis to function. A 2026 review also acknowledges that protocol heterogeneity and common-mode factors like heat and CO₂ render attribution unproven to date (PMID 42340546).

    Verdict on Four Matters: ① Mediators must be measured at multiple time points (pre, mid, end of course, follow-up); measuring once equals measuring only one point; ② Report “shared channel” and “radon increment” separately, including interaction terms; ③ Immune markers should not serve as surrogate endpoints, as the seven-fold rise in non-responders serves as a counterexample; ④ Fit decay constants before examining p-values—a curve composed of two exponential terms cannot be correctly read via single-point comparison.

    Alignment from the Perspective of Yijing Studies

    Traditional Chinese medicine posits “treat the manifestation in emergencies, treat the root in delays,” often understood as two distinct therapies and two sets of medicines. In this context, it actually describes two readings of the same intervention across two time scales: heat is the “manifestation,” arriving quickly and departing quickly; immune regulation is the “root,” starting slowly and receding slowly. Su Wen · Zhi Zhen Da Yao Lun contains a sharper statement—”prolonged application increases qi; this is the normal transformation of things; when qi increases and persists, this follows the way of heaven”—indicating that prolonged unidirectional application of a single stimulus shifts from “regulation” to “bias.” This aligns with the bidirectional regulation in Layer 2: normal mice were activated, while sick mice were pulled back to normal, indicating the system seeks a return to equilibrium rather than continuous escalation. The ancient caution against “the longer the course, the better” is not an esoteric taboo but an observation that “regulatory systems lose their capacity under long-term unidirectional input.”

    Evidence and Limitations

    Verified: Radon bath effects on pain persist up to nine months (IMuRa, 681 patients); RANKL reduction and Treg increase in the radon group are unique to radon, absent in thermal bath (RAD-ON02); active TGF-β1 significantly increased after radon mineral spring therapy with greater magnitude in responders; mouse immune activation requires 4 weeks, not apparent at 2 weeks.

    Partially Verified: Half-lives for thermal and immune effects were not directly fitted in radon bath populations; the dual-exponential arithmetic used here relies on magnitude assumptions rather than empirical parameters; TGF-β1’s role in the analgesic chain shows correlation but lacks mediation testing.

    To Be Verified: ① Empirical disentanglement of the two decay curves using restricted designs; ② Direction and magnitude of interaction terms in heat × radon factorial trials; ③ The回落 (rebound) timeline of Tregs post-treatment cessation, determining whether the “long-term effect” can be carried by them.

    Related Reading

    Evidence Level: B
    References: PMID 23864139, PMID 38292488, PMID 16810484, PMID 35400822, PMID 35203348, PMID 33649972, PMID 34546296, PMID 28882092, PMID 38872315, PMID 42340546


    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.

  • 氡浓度从低到高,镇痛是不动、缓升,还是先升后降?

    Evidence level: B

    氡温泉”有没有效”已经被测过很多次,”多浓才有效”几乎一次都没被测过——两档设计画不出曲线,而曲线形状决定了行业该用哪个浓度。

    正文

    先给结论:几乎所有氡浴试验只设两个档——有氡、无氡。两个点连不出曲线,只能连出一条直线,而直线恰恰是三种可能形状里最不该默认的那一种。

    第一层:两档设计能给什么、不能给什么。 IMuRa 试验 681 人、四类风湿病、随访 9 个月,氡浴组在疼痛(p=0.032)和镇痛药用量(p=0.007)上优于对照(PMID 23864139)。这是一份扎实的”差值”。但差值回答的是”两组不一样吗”,不回答”浓度每升一档,疼痛评分降多少”。前者的统计量是组间均值差,后者要的是斜率——而斜率至少需要三个浓度点才估得出来。所以现在所有”氡越浓越好”的说法都没有数据支撑,它只是把”有比没有好”外推了一层。

    第二层:三种形状,三个完全不同的世界。 阈值型:低于某浓度完全无效应,越过后效应突然出现,含义是”修复系统需要被激活到某个门槛”。单调型:浓度越高效应越大,含义是”效应与 α 粒子命中数成正比”。非单调型:低浓度促进、高浓度抑制,即 hormesis 的双相曲线(PMID 32564900)。三者对实操的指示正好相反——阈值型说”够浓度就行,再浓没用”;单调型说”越浓越好”;倒 U 型说”浓过头会反噬”。同一批数据在不同假设下能推出三条相反的养生建议,这就是形状必须先判的理由。

    第三层:为什么至今判不了——不是数据少,是数据结构不对。 塞尔维亚六处温泉实测水氡在 1.5 至 60.7 Bq/L 之间(PMID 31671300),欧洲氡浴场常高一到两个数量级。跨研究看浓度跨度几十倍,但每个研究内部浓度是一个固定值。这叫研究间变异,不叫剂量-反应。跨研究的浓度差同时混着水温、CO₂、疗程长度与病种,无法把效应归给氡。要画曲线,必须让浓度在同一个试验内部变化。

    第四层:更麻烦的是,你的”零剂量”不是零。 RAD-ON02 安慰剂对照试验发现,纯热浴也能减少破骨细胞分化,氡浴只是更明显;而只有氡浴组出现 RANKL 显著下降与 Treg 上升(PMID 38292488)。这句话的分量在于:对照组接受的不是”没有干预”,而是同样温度、同样时长、同样水质的热浴。所以曲线左端点不是零,是热浴基线。想估氡的独立贡献,对照必须是同源水除氡(脱气),而不是普通热水。

    第五层:真要画,怎么做。 三个最低要求:① 至少三档浓度、跨一个数量级,全部同源水、同温、同时长;② 同水除氡臂作零点锚;③ 分析用限制性立方样条看非线性(剂量-反应 meta 的标准做法,PMID 36261077;分段阈值示例见 PMID 42883787 的 6.76 METs 断点),再用似然比与 AIC 在单调模型与双相模型之间做比较(双相拟合的两种标准模型见 PMID 37645052)。判据可以写死:样条非线性检验显著且拐点置信区间不含最低最高浓度,判为阈值型或倒 U;非线性不显著而线性项显著,判为单调型;两者皆不显著,结论只能是”在此浓度范围内无可检出梯度”,而不是”氡无效”。

    第六层:易学这一侧该怎么接。 “太过不及皆病”在剂量轴上就是一条倒 U,这不是后人附会——已有综述专门讨论阴阳学说与 hormesis 双相曲线在曲线形状、对立成分关系、时间依赖性上的同构(PMID 32564900)。这条桥是真桥。但必须说清:同构是启发,不是证据。它告诉我们”该去找倒 U”,不能告诉我们”倒 U 就在这里”。判形状只能靠数据。

    第七层:现有证据的天花板。 2026 年的综述把话说得很直:疗程方案、研究质量、热与 CO₂ 等混杂因素异质性大,氡浴的独立疗效与最优方案仍未充分表征(PMID 42340546)。换句话说,这个领域缺的不是又一篇”氡有效”,而是一个多浓度臂。

    证据与局限

    已验证: 氡浴相对对照的镇痛与减少镇痛药用量效应(681 人 RCT,PMID 23864139);纯热浴亦有部分效应、氡效应更明显(PMID 38292488);阴阳与 hormesis 的同构关系已有专门综述论证(PMID 32564900)。

    部分验证: 剂量-反应形状的方法学(限制性样条、断点回归、双相模型)已经成熟,但尚未用于氡浴(PMID 36261077, 42883787, 37645052);温泉浓度实测跨度大,而单一研究内部固定(PMID 31671300)。

    待验证: 氡浴镇痛的曲线究竟属阈值型、单调型还是倒 U 型;拐点位置在哪;同水除氡臂下曲线左端点落在何处。

    局限: 本文的”三形状”框架是方法学推演,不是实测结论;所引文献以欧洲氡浴人群为主,外推到亚洲温泉需谨慎。

    相关阅读

    证据等级: B
    参考文献: PMID 23864139, PMID 38292488, PMID 32564900, PMID 31671300, PMID 36261077, PMID 42883787, PMID 37645052, PMID 42340546


    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 radon concentration increase pain relief by remaining static, gradually rising, or rising then falling?

    Evidence level: B

    The efficacy of “radon hot springs” has been tested many times, but the question of “what concentration is effective” has almost never been measured. Two-point designs cannot draw a curve, yet the shape of that curve determines which concentration tier the industry should adopt.

    Main Text

    Conclusion First: Almost all radon bath trials utilize only two tiers—radon present vs. radon absent. Two points cannot generate a curve; they can only form a straight line, which is precisely the least defensible default among the three possible shapes.

    Layer 1: What Two-Tier Designs Provide and Fail to Provide. The IMuRa trial involved 681 patients with four types of rheumatism, followed up for 9 months. The radon bath group showed superiority over controls in pain scores ($p=0.032$) and analgesic consumption ($p=0.007$) (PMID 23864139). This provides a robust “difference.” However, a difference answers “Are the two groups different?” not “How much does pain score decrease per unit increase in concentration?” The former statistic is an inter-group mean difference; the latter requires a slope—which cannot be estimated with fewer than three concentration points. Consequently, all claims that “higher radon is better” lack data support; they merely extrapolate the finding that “radon is better than none.”

    Layer 2: Three Shapes, Three Distinct Worlds.
    * Threshold Type: No effect below a certain concentration; the effect appears abruptly after crossing it. This implies “the repair system requires activation above a specific threshold.”
    * Monotonic Type: Higher concentration yields higher effect. This implies “the effect is proportional to the number of alpha particle hits.”
    * Non-monotonic Type: Low concentrations promote, high concentrations inhibit; i.e., a biphasic curve known as hormesis (PMID 32564900).

    These three shapes dictate opposite operational strategies: the threshold model suggests “sufficient concentration is enough; more is not better”; the monotonic model suggests “the higher, the better”; and the inverted U-model suggests “excess concentration causes backlash.” The same dataset can yield three opposing wellness recommendations depending on the assumed shape, necessitating a prior determination of form.

    Layer 3: Why the Shape Remains Undetermined—Not Lack of Data, but Incorrect Data Structure. Measured radon levels in six Serbian hot springs ranged from 1.5 to 60.7 Bq/L (PMID 31671300), while European radon bath facilities often range one to two orders of magnitude higher. Across studies, concentration spans span decades; however, within each study, the concentration is fixed. This is inter-study variability, not dose-response. The concentration differences across studies are confounded by water temperature, CO₂ content, treatment duration, and disease type, preventing attribution of effects to radon alone. To plot a curve, concentration must vary within the same trial.

    Layer 4: The Complication of a Non-Zero “Zero Dose”. The RAD-ON02 placebo-controlled trial found that pure thermal baths also reduce osteoclast differentiation, with radon baths showing a more pronounced effect; only the radon group showed significant RANKL reduction and Treg increase (PMID 38292488). The weight of this statement lies in the fact that the control group received not “no intervention,” but a thermal bath of identical temperature, duration, and water quality. Therefore, the left endpoint of the curve is not zero, but the thermal bath baseline. To estimate radon’s independent contribution, the control must be de-radonized water (degassing), not ordinary hot water.

    Layer 5: How to Plot It If You Must. Three minimum requirements: ① At least three concentration tiers spanning one order of magnitude, using identical water sources, same temperature, and same duration; ② Use a de-radonized arm of the same water as an anchor for the zero point; ③ Analyze using restricted cubic splines to detect non-linearity (standard practice for dose-response meta-analysis, PMID 36261077; see breakpoint example at 6.76 METs in PMID 42883787), then compare monotonic vs. biphasic models using likelihood ratio and AIC (two standard models for biphasic fitting see PMID 37645052). The criteria can be hardcoded: if spline non-linearity is significant and the confidence interval for the inflection point excludes the minimum and maximum concentrations, classify as threshold or inverted U; if non-linearity is insignificant but linear terms are significant, classify as monotonic; if neither is significant, the conclusion must be “no detectable gradient within this concentration range,” not “radon is ineffective.”

    Layer 6: How Yijing Studies Should Engage. “Excess and deficiency both cause disease” maps to an inverted U on the dose axis. This is not a later attribution—reviews exist specifically discussing the isomorphism between Yin-Yang theory and the biphasic hormesis curve regarding shape, opposing component relationships, and time dependence (PMID 32564900). This bridge is genuine. However, clarity must be maintained: isomorphism is heuristic, not evidence. It tells us “to look for an inverted U,” but cannot tell us “the inverted U is here.” Shape determination relies solely on data.

    Layer 7: The Ceiling of Current Evidence. A 2026 review states bluntly that high heterogeneity exists in treatment protocols, study quality, and confounding factors such as heat and CO₂; thus, the independent efficacy and optimal regimen of radon baths remain insufficiently characterized (PMID 42340546). In other words, this field lacks not another “radon is effective” paper, but a multi-tiered arm.

    Evidence and Limitations

    Verified: Analgesic effects and reduced analgesic consumption relative to controls in radon baths (681 patients, RCT, PMID 23864139); partial effects in pure thermal baths with radon effects being more pronounced (PMID 38292488); isomorphism between Yin-Yang and hormesis has been argued in dedicated reviews (PMID 32564900).

    Partially Verified: Methodologies for dose-response shapes (restricted splines, breakpoint regression, biphasic models) are mature but have not yet been applied to radon baths (PMID 36261077, 42883787, 37645052); hot spring concentrations show wide measured ranges, while single studies maintain fixed values (PMID 31671300).

    To Be Verified: Whether the radon bath pain relief curve is threshold, monotonic, or inverted U; the location of the inflection point; and where the left endpoint falls under a de-radonized water arm.

    Limitations: The “three-shape” framework in this text is a methodological deduction, not an empirical conclusion; cited literature focuses primarily on European radon bath populations, requiring caution when extrapolating to Asian hot springs.

    Related Reading

    Evidence Level: B
    References: PMID 23864139, PMID 38292488, PMID 32564900, PMID 31671300, PMID 36261077, PMID 42883787, PMID 37645052, PMID 42340546


    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.

  • 温泉员工的低剂量能不能直接套矿工模型:斜率衔接的三道关卡

    Evidence level: B

    决定温泉员工风险的从来不是”剂量更低”,而是”同样的剂量,以更慢的速度、更湿的方式、更长的年限吸进去”——而矿工模型恰恰在这一点上最不确定。

    正文

    先给结论:把矿工的 ERR/WLM 直接乘到温泉员工的累积剂量上,算术上做得出来,证据上站不住。问题不是”外推太远”,而是温泉员工刚好落在矿工模型最不确定的那一格——点估计最高、置信区间最宽的那一格。

    第一层:先算温泉员工一年到底吃多少 WLM。 WLM 量的是暴露量,不是浓度。换算式为 WLM = 氡浓度(Bq/m³) × 平衡因子F × 工时(h) ÷ 629000(因 1 WLM 定义为 3700 Bq/m³ 平衡态下暴露 170 小时)。取温泉现场常见值:室内氡 1000 Bq/m³、F=0.4、年工时 2000 小时 → 1.27 WLM/年;低氡温泉(200 Bq/m³)约 0.25;伊朗 Dehloran 那种未改造浴场(实测私汤 1880–2450 Bq/m³,PMID 32642584)可到 3.2。干满 40 年累积约 50 WLM。对照纽芬兰萤石矿工队列的个体均值 380.9 WLM(PMID 38649728),差约 7.5 倍。

    第二层:七倍,在流行病学里叫”外推”,不叫”跨界”。 PUMA 汇总五国 57,873 名 1960 年后入职矿工,1.9 百万人年、1,217 例肺癌死亡,ERR/100 WLM = 1.33(95%CI 0.89–1.88)(PMID 35604341)。按此斜率,50 WLM 对应 ERR ≈ 0.67,即相对风险 +67%;PUMA 的终生超额绝对风险模型给出每 WLM 5.4×10⁻⁴–7.7×10⁻⁴(PMID 38172372),乘 50 得 2.7%–3.8% 的终生肺癌死亡绝对增量——两条路算出来互相对得上。到此为止,外推成立。

    第三层:但那个 1.33 是”平均暴露率”下的斜率,不是温泉员工的斜率。 同一份 PUMA 分析里,取”确诊年龄 <55 岁、35 岁后首次暴露、年暴露率 <0.5 WL”这一格,ERR/100 WLM 跳到 8.38(95%CI 3.30–18.99);德国 Wismut 1960+ 亚组同格里是 6.92,置信区间下界直接贴零(<0; 16.59)(PMID 37695330)。温泉员工的暴露率是多少?1000 Bq/m³、F=0.4 → 0.11 WL,正好落进”<0.5 WL”这一格——而这一格的点估计是全套模型里最高的,区间也最宽。取 1.33 还是 8.38,结论差六倍。这不是精度问题,是选择问题。

    第四层:”慢慢吸”在辐射生物学里不等于”更安全”。 高 LET 的 α 暴露有个反直觉特性——逆剂量率效应:同样的累积剂量,拉得越长,致癌效应越强(PMID 16523346)。温泉员工八小时班、断断续续、常年累月,正是”极慢”的一端。但同一篇模型也指出,在低剂量端斜率会收敛回与剂量率无关的线性。两条结论方向相反,交界处在哪里,没人测过。 温泉行业恰好就在这个交界处。

    第五层:一个 WLM 值几毫希,也没有唯一答案。 用流行病学方法反推:职业暴露 14 mSv/WLM、居室 3.6 mSv/WLM(PMID 28916937);更新的估算给职业 1 WLM/年 → 6 mSv/年、居室 100 Bq/m³ → 1.1 mSv/年(PMID 40367955),并明确低于 ICRP 剂量学模型值。同一个 WLM,选职业系数还是居室系数,差 2–4 倍。温泉员工的尴尬在于:场所写在职业清单上,暴露形态却像居室(长时间、低浓度、非连续)。系数选哪个,先决定了结论。

    第六层:现场实测的巨大离散度,大半是系数和工况造成的,不是人群差异。 西班牙 54 家温泉,水温氡 <2 至 775×10³ Bq/m³、空气氡 <10 至 5,200 Bq/m³,估出员工年有效剂量 1–44 mSv,上限超 20 mSv 职业限值(PMID 10086601)——用上面的换算式复算:5,200 Bq/m³、F=0.4、2000 小时 → 6.6 WLM/年 × 6 mSv ≈ 40 mSv,与实测上限吻合。波兰 9 家地热温泉多数 <0.6 mSv/年,仅 1 家 >3(PMID 26489952);希腊案例中水氡需 >300 kBq/m³ 才越线(PMID 15051238);意大利 Ischia 20 家温泉做完通风与结构改造后浓度平均降 74%(PMID 35427836)。这些数字互相不可比,因为换算系数不同、平衡因子没报、通风工况没报。

    第七层:风水这边,说的其实是同一件事。 “气乘风则散,界水则止”——温泉池是”界水”的极端形态:一层热水把气锁在水面上方。锁住的若是湿润暖汽,人觉得舒服;锁住的若是氡子体,员工就在里面站八小时。所以温泉的风水问题不在水,在池面上方那层不流动的气——挑高、排风、池面气流才是可操作的变量。这与”藏风=换气率”的口径完全一致(见地下水氡→室内空气转移系数)。

    一句话收:矿工模型能给温泉员工一个数量级,给不了那个具体的数。要拿那个数,得先测暴露率,再决定用哪一格斜率、哪一个换算系数。

    证据与局限

    已验证: 矿工队列的 ERR/100 WLM 与终生超额绝对风险有大规模汇总分析支撑(PUMA 57,873 人 / 1.9 百万人年 / 1,217 例肺癌死亡);斜率随确诊年龄、暴露年龄、年暴露率显著修饰,低暴露率格点估计显著更高(1.33 → 8.38);温泉场所室内氡浓度与员工剂量有多个国家的实测数据;换算系数的职业/居室差异有流行病学方法学的量化比较。

    部分验证: 温泉员工的年 WLM 是用现场氡浓度、平衡因子 F、工时三个参数估算出来的,其中 F 与工时在多数温泉研究里并未同步实测,故 1.27 WLM/年只能算量级;”温泉员工落在 <0.5 WL 格”这一判断依赖 F=0.4 的默认值,若实际 F 更低(水汽环境下未结合态比例高),暴露率会进一步下移,落点更深。

    待验证: 温泉员工尚无带肺癌结局的个体化前瞻队列,”低暴露率段斜率是否与矿工外推一致”从未被直接检验;逆剂量率效应在 0.01–0.1 WL 段的实际走向未知;温泉员工该套职业系数还是居室系数,目前是规范选择而非实证结论;水汽环境下个人剂量计的读数可靠性(湿度/凝结偏移)未做系统校准。

    相关阅读

    证据等级: B
    参考文献: PMID 35604341, PMID 38172372, PMID 37695330, PMID 38649728, PMID 16523346, PMID 28916937, PMID 40367955, PMID 10086601, PMID 26489952, PMID 15051238, PMID 35427836, PMID 32642584


    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。

    Can the low-dose model for hot spring workers be directly applied to miners: a three-stage slope-connection approach

    Evidence level: B

    The factor determining risk for hot spring workers is never “lower dose,” but rather “the same dose absorbed at a slower rate, in a wetter manner, over longer years”—where the miner model is precisely most uncertain at this point.

    Main Text

    Conclusion First: Directly multiplying the miner’s ERR/WLM by the cumulative dose of hot spring workers is arithmetically feasible but unsupported by evidence. The issue is not “extrapolation too far,” but that hot spring workers fall exactly into the miner model’s most uncertain cell—the one with the highest point estimate and widest confidence interval.

    Layer 1: First, calculate exactly how much WLM hot spring workers consume annually. WLM measures exposure quantity, not concentration. The conversion formula is WLM = Radon Concentration (Bq/m³) × Equilibrium Factor F × Working Hours (h) ÷ 629,000 (since 1 WLM is defined as exposure for 170 hours at equilibrium of 3,700 Bq/m³). Using common hot spring site values: indoor radon 1,000 Bq/m³, F=0.4, annual working hours 2,000 → 1.27 WLM/year; low-radon springs (200 Bq/m³) approx 0.25; unmodified bathhouses like Iran’s Dehloran (measured private baths 1,880–2,450 Bq/m³, PMID 32642584) reach 3.2. Cumulative over a full 40 years approx 50 WLM. Comparing against the Newfoundland fluorite miner cohort’s individual mean 380.9 WLM (PMID 38649728), the difference is approx 7.5 times.

    Layer 2: Seven times in epidemiology is called “extrapolation,” not “crossing boundaries.” PUMA aggregated data from 57,873 miners hired after 1960 across five countries: 1.9 million person-years, 1,217 lung cancer deaths, ERR/100 WLM = 1.33 (95%CI 0.89–1.88) (PMID 35604341). At this slope, 50 WLM corresponds to ERR ≈ 0.67, i.e., relative risk +67%; PUMA’s lifetime excess absolute risk model gives 5.4×10⁻⁴–7.7×10⁻⁴ per WLM (PMID 38172372), multiplied by 50 yields a lifetime lung cancer death absolute increment of 2.7%–3.8%—both paths yield consistent results. Up to here, extrapolation holds.

    Layer 3: But that 1.33 is the slope under “average exposure rate,” not hot spring workers’ slope. In the same PUMA analysis, taking the cell “diagnosis age <55, first exposure after 35, annual exposure rate <0.5 WL” jumps ERR/100 WLM to 8.38 (95%CI 3.30–18.99); the German Wismut 1960+ subgroup in the same cell is 6.92, with the confidence interval lower bound touching zero (<0; 16.59) (PMID 37695330). What is the hot spring worker’s exposure rate? 1,000 Bq/m³, F=0.4 → 0.11 WL, falling exactly into the “<0.5 WL” cell—where this cell’s point estimate is the highest across all models and its interval is widest. Choosing 1.33 vs 8.38 results in a six-fold difference in conclusion. This is not a precision issue, but a selection issue.

    Layer 4: “Slow absorption” in radiation biology does not equal “safer.” High LET alpha exposure has a counter-intuitive characteristic—the inverse dose-rate effect: for the same cumulative dose, the longer it is stretched, the stronger the carcinogenic effect (PMID 16523346). Hot spring workers, with eight-hour shifts, intermittent work, and years of continuous labor, are precisely at the “extremely slow” end. However, the same model notes that at the low-dose end, the slope converges back to a linear relationship unrelated to dose rate. The two conclusions point in opposite directions; where the boundary lies, no one has measured it. The hot spring industry sits exactly at this boundary.

    Layer 5: One WLM value equals several millisieverts, with no single answer. Using epidemiological methods to reverse-calculate: occupational exposure 14 mSv/WLM, residential 3.6 mSv/WLM (PMID 28916937); updated estimates give occupational 1 WLM/year → 6 mSv/year, residential 100 Bq/m³ → 1.1 mSv/year (PMID 4036955), explicitly lower than ICRP dosimetric model values. For the same WLM, choosing an occupational coefficient vs. a residential coefficient yields a 2–4 fold difference. The awkwardness of hot spring workers lies in: the location is listed under occupational categories, yet the exposure morphology resembles residential (long duration, low concentration, non-continuous). Which coefficient is selected first determines the conclusion.

    Layer 6: The huge dispersion in field measurements stems largely from coefficients and working conditions, not population differences. Across 54 Spanish hot springs, water temperature radon <2 to 775×10³ Bq/m³, air radon <10 to 5,200 Bq/m³, estimating employee annual effective dose 1–44 mSv, with the upper limit exceeding the 20 mSv occupational limit (PMID 10086601)—re-calculating with the above formula: 5,200 Bq/m³, F=0.4, 2,000 hours → 6.6 WLM/year × 6 mSv ≈ 40 mSv, matching the measured upper limit. In Poland, 9 geothermal springs mostly <0.6 mSv/year, only 1 exceeding 3 (PMID 26489952); in the Greek case, water radon must exceed 300 kBq/m³ to cross the line (PMID 15051238); in Italy’s Ischia, after ventilation and structural modifications, concentrations averaged down 74% (PMID 35427836). These numbers are incomparable due to different conversion coefficients, unreported equilibrium factors, and unreported ventilation conditions.

    Layer 7: Feng Shui here refers to the same matter. “Qi scatters when winded, stops when bounded by water”—the hot spring pool is an extreme form of “bounded by water”: a layer of hot water locks the qi above the surface. If what is locked is moist warm vapor, humans feel comfortable; if it is radon progeny, employees stand inside for eight hours. Therefore, the Feng Shui issue with hot springs is not in the water, but in that non-moving layer of qi above the pool surface—height adjustment, ventilation, and pool surface airflow are the actionable variables. This is fully consistent with the “hiding wind = air exchange rate” metric (see Radon transfer coefficient from groundwater to indoor air).

    One Sentence Summary: The miner model provides hot spring workers with an order of magnitude, but not that specific number. To obtain that number, one must first measure the exposure rate, then decide which slope cell to use and which conversion coefficient applies.

    Evidence and Limitations

    Verified: The miner cohort’s ERR/100 WLM and lifetime excess absolute risk are supported by large-scale meta-analyses (PUMA 57,873 individuals / 1.9 million person-years / 1,217 lung cancer deaths); the slope is significantly modified by diagnosis age, exposure age, and annual exposure rate, with low exposure rate grid estimates being significantly higher (1.33 → 8.38); indoor radon concentrations at hot spring sites and employee doses have measured data from multiple countries; differences between occupational and residential conversion coefficients have quantitative comparisons via epidemiological methodology.

    Partially Verified: Annual WLM for hot spring workers is estimated using three parameters: site radon concentration, equilibrium factor F, and working hours; where F and working hours are not simultaneously measured in most hot spring studies, thus 1.27 WLM/year only calculates the magnitude; the judgment that “hot spring workers fall into the <0.5 WL grid” relies on the default value F=0.4; if actual F is lower (higher unattached fraction in humid environments), the exposure rate would shift further down, placing the point deeper.

    To Be Verified: There is currently no individualized prospective cohort for hot spring workers with lung cancer outcomes; whether “low exposure rate segment slopes align with miner extrapolation” has never been directly tested; the actual direction of the inverse dose-rate effect in the 0.01–0.1 WL segment remains unknown; whether hot spring workers should use this set of occupational or residential coefficients is currently a normative choice rather than an empirical conclusion; the reliability of personal dosimeter readings in humid environments (humidity/condensation offsets) has not undergone systematic calibration.

    Related Reading

    Evidence Level: B
    References: PMID 35604341, PMID 38172372, PMID 37695330, PMID 38649728, PMID 16523346, PMID 28916937, PMID 4036955, PMID 10086601, PMID 26489952, PMID 15051238, PMID 35427836, PMID 32642584


    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.

  • 湿度不是让剂量计读错,是让你记错了变量

    Evidence level: B

    温泉现场那句”湿度让剂量计读不准”,方向对了一半,归因全错了:真正被系统性漏掉的是气压,而湿度只在”凝结”这一个事件上才真的动手。

    正文

    先给结论:把”湿度偏移”当成一个可以减掉的常数,是这件事上最贵的误会。同一个湿度,在三种主流探测原理上的作用方向、量级、甚至数学形式都不一样——在 CR-39 上它几乎没有连续效应,只在一个二值事件上起作用;在静电场收集上是真实的连续衰减,但决定量是绝对含水量;在活性炭上是 2 倍量级的共犯。真正被忽略的是气压。

    第一层:先把”读数偏移”拆成三件事。 探测器读错,可能是灵敏度变了(斜率)、本底变了(截距)、或者响应慢了(相位)。对累积型探测器,前两者直接改结果;对连续监测仪,第三件最致命——响应 15 小时的仪器摆在池边,测的是昨天的平均,而员工今天的暴露被记进了昨天。三种错误统称”湿度偏移”,等于把三个不同的数加成一个。

    第二层:CR-39 上,湿度几乎无罪,凝结才是开关。 蒙特卡洛模拟 5 种 α 径迹探测器:气压降 35.5 kPa(约升高 3500 m),校正常数升 31%;温度 5→35℃ 升 12%;而相对湿度几乎不影响,除非探测器内部发生水汽凝结(PMID 24593337)。2026 年的实测同向:RH 从 <20% 到 >98%,检出限只升约 10%(PMID 41175455)。这两句话的结构很关键——不是”湿度影响小”,是”湿度的影响是阶跃的”:不凝结接近零,一凝结就是另一回事。温泉恰恰最容易凝结:池边 40℃ 湿空气遇到走廊、清晨的窗,露点差一越界就凝。

    第三层:于是替罪羊结构出现了——你担心湿,漏了压。 中国、日本及中国台湾地区的温泉大量在山地。海拔 2000 m 相对海平面气压约低 20 kPa,高原路线可到 35 kPa 以上,正对上那 31% 的区间。同一批探测器在平地校准、在山区温泉使用,差出的量比”湿度偏移”大一个数量级。湿度被反复点名,是因为它看得见(墙在滴水);气压看不见,因为没人给房间装气压计。

    第四层:静电场收集上湿度真有罪,但判据是绝对含水量。 RAD7 这类仪器用水分子中和 ²¹⁸Po 离子,含水量越高,硅探测器相对闪烁室的效率越低;含水量超过 0.01 g 后效应大幅增强,整体效率误差 1.8%–7.5%(PMID 27253897)。另一组实验给出钍灵敏度随绝对湿度呈负指数下降(PMID 36603217)。注意是绝对湿度:20℃、RH 60% 的空气含水约 10.4 g/m³;浴室 40℃、RH 90% 约 46 g/m³——RH 只从 60 写到 90,含水量差 4.4 倍;拿冬天室外 0℃、RH 80%(约 3.9 g/m³)来比,差 12 倍。用 RH 记日志,等于拿一把随温度缩放的尺子量长度。

    第五层:活性炭罐上,湿度是双倍量级的共犯,还带记忆。 活性炭的吸附效率在室内温度区间内可变 2 倍(PMID 20937546),必须温湿度双参数校正;Polanyi 势理论解释水分子为何抢占吸附位(PMID 7860304),开放面罐与扩散屏障罐行为又不同(PMID 8253580)。更麻烦的是它对暴露期间浓度剧变敏感(PMID 17077091)——温泉”有人就升、没人就降”的工况,正打在它最不擅长的点上。

    第六层:互比实验指向的实体是外壳,不是空气。 MetroRADON 项目在天然辐射实验室做被动与主动同场比对,>80% 结果落在参考值 ±20%/±10% 带内,而多数偏差是被动探测器高估,归因到探测器外壳材料的脱气时间(PMID 32182944)。这就是”记忆”的物理形态:外壳先吸湿,再把水汽慢慢放进探测腔,上一段高湿被记进了下一段。同批低成本监测仪与参考仪的差 5%–16%、缺失记录 1%–19%(PMID 38544099)。这给出一道门槛:若湿度偏移小于仪器之间的固有离散,那”校准湿度”只是给噪声换个更精细的名字。

    第七层:就算浓度测准了,湿度还在改剂量。 三种未结合态氡子体监测仪在真实环境与氡室比对:平衡等效浓度多在 ±3.0% 内,但未结合态份额的相对差达 3.3%–39.5%,且”极高湿 + 低气溶胶”环境里出现异常高的份额(PMID 32365338)。温泉池边正是这种环境。未结合态份额直接决定剂量换算——同一浓度下剂量可差数倍。室内平衡因子还呈 V 形日变化,受温度、湿度、降雨、湍流共同驱动(PMID 40056770)。所以湿度对”剂量”的影响可能大于对”浓度”的影响,而所有人都在校准浓度。

    判决五件事:① 日志记绝对含水量(g/m³)与露点差,不记 RH;”凝结事件”当布尔变量单独记 ② 分原理校准:CR-39 只查凝结与气压,静电场按月含水量分档算效率,活性炭温湿度双校正 ③ 被动 + 主动双原理并行互比,不互替——失效模式不重叠 ④ 先算”效应量 / 仪器不确定度”,小于 1 就不必校准,否则校的是噪声 ⑤ 剂量侧必报平衡因子与未结合态份额,不只报浓度;顺手给房间挂一台气压计。

    风水这边:传统风水讲湿,从来不讲”湿度百分之几”,讲的是”墙壁出汗””地气上返””回南天”——这些全是凝结事件的语言,不是平均湿度的语言。古人没有湿度计,只能记现象,反而记对了对象:对 CR-39 这类探测器,决定性变量恰好就是”凝结没凝结”这个事件,不是月均相对湿度。现代人记一个平均数,以为更精确,其实把真正起作用的那件事抹平了。两边少见地对上了:判据都该是事件式的。错位在别处——风水把湿当作”气的性质”(气不清),仪器只认水与空气的比值。

    一句话收:高湿现场需要的不是更准的校准系数,是更对的日志字段。

    证据与局限

    已验证: 气压与温度对 α 径迹探测器校正常数的影响有蒙特卡洛量化(31% / 12%)并被 2026 年实测(RH <20%→>98% 检出限仅升约 10%)同向支持;静电场收集受水分子中和影响的机制与量级有对照实验(效率误差 1.8%–7.5%,含水量 >0.01 g 后显著增强);活性炭吸附效率的温度依赖性有 2 倍量级的舱室实测;被动/主动同场互比中”被动探测器高估、归因外壳脱气”有欧洲多实验室数据;未结合态份额在极高湿低气溶胶环境的异常升高有实测。

    部分验证: 温泉现场的湿度偏移量本身缺少”同一批探测器跨湿度梯度同场比对”的直接数据,多数推论由氡室与外场互比外推而来;”外壳脱气滞后”是互比偏差的归因解释,尚未通过更换外壳材料的前后对照独立证实;气压效应针对高原温泉的定量外推依赖海拔-气压换算,尚未在温泉场景实测。

    待验证: 温泉现场”凝结事件”与探测器读数偏差的逐夜对应关系从未建立;绝对含水量分档校正能否把静电场类仪器的读数离散压进仪器固有噪声以内,缺少验证;把日志字段从 RH 改为绝对含水量 + 露点差之后,员工年 WLM 估计的方差是否实际下降,需要前瞻性对照。

    相关阅读

    证据等级: B
    参考文献: PMID 24593337, PMID 41175455, PMID 27253897, PMID 36603217, PMID 20937546, PMID 7860304, PMID 8253580, PMID 17077091, PMID 32182944, PMID 38544099, PMID 32365338, PMID 40056770


    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。

    Humidity Does Not Cause Dosimeters to Read Incorrectly; It Causes You to Forget the Variable

    Evidence level: B

    The statement at the hot spring site, “humidity causes the dosimeter to read inaccurately,” got the direction right but the attribution entirely wrong: what was systematically missed is atmospheric pressure, while humidity only truly acts on the single event of “condensation.”

    Main Text

    Conclusion First: Treating “humidity offset” as a constant to be subtracted is the most expensive misconception in this matter. The same humidity level exerts different directional effects, magnitudes, and even mathematical forms across three mainstream detection principles: on CR-39 it has almost no continuous effect, acting only on a binary event; in electrostatic collection it causes real continuous attenuation, but the determining factor is absolute water content; on activated carbon it is an accomplice at a two-fold magnitude. What was truly ignored is atmospheric pressure.

    Layer 1: Decompose “Reading Offset” into Three Distinct Events. A detector reading error may stem from changed sensitivity (slope), changed background (intercept), or slowed response (phase). For cumulative detectors, the first two directly alter results; for continuous monitors, the third is most lethal—an instrument left by the poolside measuring yesterday’s average while today’s exposure is recorded as yesterday’s. All three errors are collectively termed “humidity offset,” which amounts to adding three different numbers into one sum.

    Layer 2: On CR-39, Humidity Is Almost Innocent; Condensation Is the Switch. Monte Carlo simulations of five alpha-track detectors show that a pressure drop of 35.5 kPa (approx. +3500 m elevation) increases the calibration constant by 31%; a temperature rise from 5→35°C increases it by 12%; while relative humidity (RH) has almost no effect unless water vapor condenses inside the detector (PMID 24593337). 2026 field measurements align with this: RH rising from <20% to >98% increases the detection limit by only ~10% (PMID 41175455). The structure of these statements is critical—it is not “humidity has a small effect,” but “the effect of humidity is step-function“: near zero without condensation, completely different with it. Hot springs are precisely where condensation is most likely: warm, moist air at 40°C near the pool meets corridors or morning windows; a small difference in dew point crosses the threshold for condensation.

    Layer 3: Thus, a Structure of Guilt Appears—You Worry About Humidity, Miss the Pressure. Many hot springs in China, Japan, and Taiwan are located in mountainous regions. At 2000 m altitude, atmospheric pressure is approx. 20 kPa lower than sea level; high-altitude routes can exceed 35 kPa, matching the range for that 31% shift. The same batch of detectors calibrated at sea level but used in mountain hot springs shows discrepancies an order of magnitude larger than “humidity offset.” Humidity is repeatedly named because it is visible (water dripping on walls); pressure is invisible because no one installs a barometer in the room.

    Layer 4: On Electrostatic Collection, Humidity Is Guilty, But the Criterion Is Absolute Water Content. Instruments like RAD7 use water molecules to neutralize ²¹⁸Po ions; higher water content lowers the efficiency of the silicon detector relative to the scintillation chamber; effects significantly enhance once water content exceeds 0.01 g, causing overall efficiency errors of 1.8%–7.5% (PMID 27253897). Another study shows thoron sensitivity decreases exponentially with absolute humidity (PMID 36603217). Note the absolute humidity: air at 20°C with RH 60% contains approx. 10.4 g/m³; a bathroom at 40°C with RH 90% contains approx. 46 g/m³—RH only changes from 60 to 90, but water content differs by 4.4 times; compared to winter outdoor air at 0°C with RH 80% (approx. 3.9 g/m³), the difference is 12 times. Using RH for logging is like using a ruler that scales with temperature to measure length.

    Layer 5: On Activated Carbon Canisters, Humidity Is a Double-Magnitude Accomplice with Memory. Adsorption efficiency on activated carbon varies by 2-fold within indoor temperature ranges (PMID 20937546) and requires dual-parameter correction for temperature and humidity; Polanyi potential theory explains why water molecules抢占 (seize) adsorption sites (PMID 7860304), and open-face canisters behave differently from diffusion-barrier canisters (PMID 8253580). More麻烦 (troublesome) is its sensitivity to sharp concentration changes during exposure (PMID 17077091)—the hot spring condition of “rising when people enter, falling when they leave” hits its least capable point.

    Layer 6: Comparative Experiments Point to the Housing, Not the Air. The MetroRADON project performed passive and active co-location comparisons in a natural radiation laboratory; >80% of results fell within the reference ±20%/±10% band, while most deviations were overestimates by passive detectors attributed to outgassing time of the detector housing material (PMID 32182944). This is the physical form of “memory”: the housing absorbs moisture first, then slowly releases it into the detection chamber, carrying forward high humidity from one segment to the next. Differences between low-cost monitors and reference instruments ranged from 5%–16%, with missing records of 1%–19% (PMID 385434099). This sets a threshold: if the humidity offset is smaller than the intrinsic dispersion between instruments, then “calibrating for humidity” merely renames noise with greater precision.

    Layer 7: Even If Concentration Is Measured Accurately, Humidity Still Alters Dose. Three unattached radon progenitor monitors compared in real environments and radon chambers showed equilibrium equivalent concentrations mostly within ±3.0%, but relative differences in the unattached fraction reached 3.3%–39.5%, with anomalously high fractions appearing in “very high humidity + low aerosol” environments (PMID 32365338). Hot spring poolside is precisely such an environment. The unattached fraction directly determines dose conversion—dose can differ by orders of magnitude at the same concentration. Indoor equilibrium factors also show V-shaped diurnal variations driven jointly by temperature, humidity, rainfall, and turbulence (PMID 40056770). Thus, the effect of humidity on “dose” may exceed its effect on “concentration,” yet everyone calibrates for concentration.

    Verdict on Five Matters: ① Log absolute water content (g/m³) and dew point difference, not RH; record “condensation events” as a separate boolean variable ② Calibrate by principle: check condensation and pressure for CR-39; calculate efficiency tiers by monthly water content for electrostatic collection; apply dual correction for temperature and humidity on activated carbon ③ Run passive and active parallel comparisons without substitution—failure modes must not overlap ④ Calculate “effect size / instrument uncertainty” first; if <1, no calibration needed, otherwise you are calibrating noise ⑤ On the dose side, always report equilibrium factors and unattached fractions, not just concentration; hang a barometer in the room.

    On Feng Shui: Traditional Feng Shui discusses “dampness” but never “humidity percentage”; it speaks of “walls sweating,” “earth qi rising,” and “Hainan days”—all are languages of condensation events, not average humidity. Ancient people had no hygrometers and could only record phenomena, yet they recorded the correct objects: for detectors like CR-39, the decisive variable is exactly whether “condensation occurred,” not the monthly average RH. Modern people record an average number, thinking it is more precise, yet they flatten the very event that actually acts. Rarely do both sides align correctly: the criterion should be event-based. The misalignment lies elsewhere—Feng Shui treats dampness as a “nature of Qi” (unclear qi), while instruments only recognize the ratio of water to air.

    One Sentence Summary: High-humidity sites do not need more accurate calibration coefficients; they need correct log fields.

    Evidence and Limitations

    Verified: The effects of pressure and temperature on alpha-track detector calibration constants have been quantified via Monte Carlo (31% / 12%) and supported by 2026 field measurements (RH <20%→>98% detection limit increased by only ~10%); mechanisms and magnitudes of electrostatic collection affected by water molecule neutralization have been verified via comparative experiments (efficiency errors 1.8%–7.5%, significant enhancement after water content >0.01 g); temperature dependence of activated carbon adsorption efficiency has been verified via chamber measurements at a two-fold magnitude; data from multiple European laboratories confirms “passive detector overestimation attributed to housing outgassing” in passive/active co-location comparisons; anomalous elevation of unattached fractions in very high humidity-low aerosol environments has been measured.

    Partially Verified: Direct data lacking “co-location comparison of the same batch of detectors across humidity gradients” for the magnitude of humidity offset at hot spring sites; most inferences extrapolated from radon chamber and field comparisons; “housing outgassing lag” is an attribution explanation for comparison deviations but has not been independently verified via pre-post control studies changing housing materials; quantitative extrapolation of pressure effects to high-altitude hot springs relies on altitude-pressure conversion and has not been measured in the hot spring scenario.

    To Be Verified: The nightly correspondence between “condensation events” and detector reading deviations at hot spring sites has never been established; whether tiered calibration by absolute water content can reduce reading dispersion in electrostatic-class instruments within the instrument’s intrinsic noise lacks verification; whether changing log fields from RH to absolute water content + dew point difference actually reduces the variance of annual WLM estimates for employees requires a prospective control study.

    Related Reading

    Evidence Level: B
    References: PMID 24593337, PMID 41175455, PMID 27253897, PMID 36603217, PMID 20937546, PMID 7860304, PMID 8253580, PMID 17077091, PMID 32182944, PMID 385434099, PMID 32365338, PMID 40056770


    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.

  • 「一比一万」是折算率,不是常数:气水比的三个数量级,长在房间里不在地区上

    Evidence level: B

    地下水氡→室内空气的转移系数,从来不是一个测出来的常数,而是”用水量×释放率÷(房间体积×换气率)”的一个折算结果。它把两个自变量吸收进一个数字,于是”地区差异”里混进了大量”你测的到底是哪间屋子”。

    正文

    同一个州、同一批研究者、同一套协议、同一种探测器,测出的水→气转移系数能从 9.6×10⁻⁶ 跨到 2.0×10⁻²(PMID 15083149)。这是两千倍。当一种”物理系数”在同一个州的七所学校之间就能跨三个数量级,讨论”地区异质性”之前,得先问这个系数到底是什么。

    第一层:这个比值可以写成一条式子。 一区稳态质量平衡给出 f = C_空气/C_水 = Q_水·η /[V·(AER+λ)],其中 Q_水 是单位时间用水量、η 是释放率、V 是房间体积、AER 是换气率,λ=0.0076 h⁻¹(氡半衰期 3.82 天)可忽略。做个量级核对:一户 400 L/日、η≈0.5、V=300 m³、AER=0.5 h⁻¹,得 f≈5.6×10⁻⁵,即 1/18000,与 1:10000 同量级。这说明”一比一万”不是测量结果,是一组典型住宅参数下的折算结果——参数一换,它跟着换。 把它当常数写进指南,等于把四个自变量锁死在某个 1980 年代的参考住宅上。

    第二层:同一个州内跨两千倍,已经把”地区异质性”的叙事打掉。 缅因州七所学校的实测里,水氡 10–960 Bq/L、换气率 0.012–0.066 min⁻¹(即 0.72–3.96 h⁻¹)、释放率 0.04–0.98——只释放率一项就有 25 倍跨度(PMID 15083149)。若真存在占主导的”地区效应”,地区内部就该高度同质;这里恰恰相反:变异主要长在同城、同州、同一种房子的房子之间。 地区差异要成立,得先证明组内变异小于组间,而这类实测从没给出过这个前提。

    第三层:”地区差异”里混着”你测的是哪间屋子”。 同一口井、同一栋房,全屋口径是 1/10000,浴室口径是 1/2300,差的 4.3 倍纯粹来自浴室体积小(PMID 24803512)。再看两组常被当作”地区偏高”引用的数据:广东温泉酒店浴室 6.2×10⁻⁴–4.1×10⁻³(PMID 21382658)、保加利亚康复医院 4.5×10⁻⁴–8.4×10⁻³(PMID 34709547)。它们相对 10⁻⁴ 高出 6 到 84 倍,听起来像地域特征;可与 Harley 的浴室/全屋 4.3 倍一比,这些”偏高”里很大一块只是”测的是浴室而不是全屋”。广东浴室值 6.2×10⁻⁴ 与 Harley 浴室 4.35×10⁻⁴ 只差 40%,两者相隔一个太平洋。

    第四层:分母本身测不准。 示踪气体法在多区建筑里,即使方法得当,精度也只到 33% 以内,且区数 ≥2 时信息不足、得不到准确换气率(PMID 21318005)。要归因必须独立测”氡进入率”与”换气率”两条通道(PMID 21460366)——只测总浓度,两者会被混成一个数。节能建筑正是靠压低换气率把氡抬上去的(PMID 25977350),这提示:气候区对气水比的作用,几乎全部经由 AER 这一条路走,它不是独立的地理变量。

    第五层:分子的取样点决定一切。 希腊 Arnea 实测:龙头 22–28.9%(流量 2.7–7 L/min),淋浴 45–48.3%(6–8 L/min)(PMID 37276723)。Harley 的数据更刺眼:冷水 60、热水 38 Bq/L,而淋浴过程在浴缸底收集到的只剩 5 Bq/L——92% 已在水柱下落途中跑掉(PMID 24803512)。水温还改变气液分配系数,忽略盐度会造成 10%–20% 的高估(PMID 22385122)。所以”水氡”这个分子,必须先交代取样点、水温与冷热端,否则同一口井能给出三个不同的数。

    第六层:合并这一步最容易出系统性错。 环境与生物监测数据几乎全是对数正态,跨研究合并必须换成几何均值与几何标准差,算术均值±SD 在分布右偏时不可比(PMID 25333994)。而 I² 只说明”有异质性”,不回答”下一个房子会落在哪”——要报预测区间;Cochrane 479 个显著且带异质性的 meta 分析里,72.4% 的预测区间含零甚至反向(PMID 27406637)。把这套用到气水比上,结论是:能报的是”新房子可能落在哪个区间”,不是”某地区平均是多少”。

    第七层(也是最该改口的一层):分母也许根本不该是空气。 科罗拉多 37 户的实测说明,水贡献是否可测,取决于水氡/土壤气氡之比,而不是水氡绝对值;土壤气高时,水氡到 750 kBq/m³ 也被掩盖;一旦把土壤气治理到 150 Bq/m³ 以下,洗衣房的水贡献立刻升到 20%–40%(PMID 8056590)。中国跨气候带 171 户的年均值 81 Bq/m³、冬夏比 1.68、季节修正因子 0.80–1.43(PMID 41626976),说明背景本身随季节大幅摆动。“气水比”这个提法预设了”空气是唯一背景”,可真实背景是土壤气——分母选错,比值的含义就变了。

    要做事,就做四件。① 报转移系数时四件套同报:房间体积、换气率、单位时间用水量、取样点与水温,缺一项即不可跨研究比较。② 用几何均值加预测区间,不用算术均值。③ 把 f 拆成可复算的分项(Q·η / V / AER),逐项给不确定度,别只给一个总比值。④ 分母改用”水氡/土壤气氡”比;气候与季节只通过 AER 进入模型,修正因子加在 AER 上而不是加在 f 上。

    风水这边,《葬书》讲”气乘风则散,界水则止”——它把水当聚气的界面,判据是空间、尺度是年。而 1:10000 在风水里没有对应物,它是辐射防护为了做剂量筛查造出来的折算率。真正能与风水对话的变量是”藏风”,它的物理名字就叫换气率。所以”哪个地区的水更危险”这个问法本身就偏了:该问的是”哪种房子把水关得更紧”。风水一直在讲房子,从来不在讲地区。

    证据与局限

    已验证: 同一州内转移系数跨 3.3 个数量级(9.6×10⁻⁶–2.0×10⁻²),换气率与释放率是主要来源(PMID 15083149);浴室与全屋口径差 4.3 倍,源于体积(PMID 24803512);龙头与淋浴释放率分别 22%–28.9%、45%–48.3%(PMID 37276723);示踪气体测换气率不确定度约 33%(PMID 21318005);换气率与氡浓度负相关(PMID 25977350);水贡献可测性取决于水/土壤气比(PMID 8056590)。
    部分验证: 广东 6.2×10⁻⁴–4.1×10⁻³(PMID 21382658)与保加利亚 4.5×10⁻⁴–8.4×10⁻³(PMID 34709547)的区域差异,可被”测的是浴室”解释掉大部分,但两地的用水强度与水温差异尚未逐项分解;季节修正因子 0.80–1.43 已跨气候带标定(PMID 41626976),但未与转移系数联合建模。
    待验证: 是否存在在控制体积、AER、用水量、取样点之后仍然显著的”地区剩余效应”,从未有研究做过这个四变量校正后的方差分解;气水比的地区异质性中,房间参数的贡献占比缺直接估计。

    相关阅读

    证据等级: B
    参考文献: PMID 15083149, PMID 24803512, PMID 37276723, PMID 21382658, PMID 34709547, PMID 21318005, PMID 21460366, PMID 25977350, PMID 22385122, PMID 25333994, PMID 27406637, PMID 8056590, PMID 41626976


    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。

    \”1 to 10,000\” is a conversion factor, not a constant: three orders of magnitude in gas-to-water ratio depend on the room, not the region

    Evidence level: B

    The transfer coefficient from groundwater to indoor air is never a measured constant; it is a derived result calculated as “(water usage rate × release rate) ÷ (room volume × air exchange rate).” It absorbs two independent variables into a single number, thereby conflating “regional differences” with the sheer variability of “which specific room was measured.”

    Main Text

    Within the same state, among the same researchers, under the same protocol, with the same detector, the water-to-air transfer coefficient ranges from $9.6 \times 10^{-6}$ to $2.0 \times 10^{-2}$ (PMID 15083149). This is a two-thousand-fold difference. When a “physical coefficient” spans three orders of magnitude across seven schools in a single state, one must question the nature of this coefficient before discussing “regional heterogeneity.”

    Layer 1: This ratio can be expressed as an equation. A steady-state mass balance for Zone 1 yields $f = C_{air}/C_{water} = Q_{water} \cdot \eta / [V \cdot (AER + \lambda)]$, where $Q_{water}$ is the water usage rate, $\eta$ is the release rate, $V$ is room volume, and $AER$ is the air exchange rate. $\lambda = 0.0076 \text{ h}^{-1}$ (radon half-life of 3.82 days) is negligible. A magnitude check confirms: for a household with $400 \text{ L/day}$, $\eta \approx 0.5$, $V = 300 \text{ m}^3$, and $AER = 0.5 \text{ h}^{-1}$, we get $f \approx 5.6 \times 10^{-5}$, or $1/18,000$, which is of the same order as $1:10,000$. This indicates that “one-to-ten-thousand” is not a measurement result, but a conversion factor derived from typical residential parameters; the value shifts if the parameters change. Encoding it as a constant in guidelines effectively locks four independent variables to a reference residence from the 1980s.

    Layer 2: A two-thousand-fold variation within the same state already dismantles the narrative of “regional heterogeneity.” In the Maine study across seven schools, water radon ranged from 10–960 Bq/L, air exchange rates from $0.012–0.066 \text{ min}^{-1}$ (i.e., $0.72–3.96 \text{ h}^{-1}$), and release rates from 0.04–0.98. The release rate alone spans a 25-fold range (PMID 15083149). If a dominant “regional effect” truly existed, internal homogeneity should be high; here, the opposite is true: variation occurs primarily between houses in the same city, state, and housing type. For regional differences to hold, within-group variance must be smaller than between-group variance, a premise never established by such empirical data.

    Layer 3: “Regional differences” are conflated with “which room was measured.” For the same well and building, the whole-house aperture yields $1/10,000$, while the bathroom aperture yields $1/2,300$—a 4.3-fold difference purely due to smaller volume (PMID 24803512). Consider two datasets often cited as evidence of “higher regional values”: Guangdong spa hotel bathrooms ($6.2 \times 10^{-4}–4.1 \times 10^{-3}$) (PMID 21382658) and Bulgarian rehabilitation hospitals ($4.5 \times 10^{-4}–8.4 \times 10^{-3}$) (PMID 34709547). These are 6 to 84 times higher relative to $10^{-4}$, sounding like regional characteristics. Yet, compared to Harley’s bathroom/whole-house 4.3-fold difference, a large portion of these “elevated” values simply reflects measuring a bathroom rather than the whole house. The Guangdong bathroom value of $6.2 \times 10^{-4}$ differs from Harley’s bathroom value of $4.35 \times 10^{-4}$ by only 40%, despite being separated by an ocean.

    Layer 4: The denominator itself is measured with uncertainty. Tracer gas methods in multi-zone buildings, even when applied correctly, achieve precision only within 33%, and information is insufficient or inaccurate for air exchange rates when the number of zones $\ge 2$ (PMID 21318005). Attribution requires independent measurement of both “radon entry rate” and “air exchange rate” (PMID 21460366); measuring only total concentration conflates the two. Energy-efficient buildings raise radon levels precisely by suppressing air exchange rates (PMID 25977350), suggesting that the role of climate zones in the air-water ratio operates almost entirely through the AER pathway, not as an independent geographic variable.

    Layer 5: The sampling point in the numerator determines everything. Greek Arnea measurements show tap water release rates of 22–28.9% (flow 2.7–7 L/min) versus shower water at 45–48.3% (6–8 L/min) (PMID 37276723). Harley’s data is even more striking: cold water at 60 Bq/L, hot water at 38 Bq/L, yet only 5 Bq/L is collected at the bottom of the tub during showering—92% lost during water column descent (PMID 24803512). Water temperature also alters the gas-liquid partition coefficient, and ignoring salinity causes overestimation by 10%–20% (PMID 22385122). Therefore, the “water radon” numerator must specify sampling points, water temperature, and cold/hot ends; otherwise, the same well can yield three different numbers.

    Layer 6: Merging data introduces systematic error most easily. Environmental and biological monitoring data are nearly all log-normally distributed; cross-study merging must use geometric means and geometric standard deviations, as arithmetic mean $\pm$ SD is incomparable under right-skewed distributions (PMID 25333994). $I^2$ only indicates “heterogeneity exists,” not where the next house will fall—one must report prediction intervals. Among 479 significant meta-analyses with heterogeneity in Cochrane reviews, 72.4% of prediction intervals include zero or even reverse effects (PMID 27406637). Applying this framework to the air-water ratio concludes that one can report “which interval a new house might fall into,” not “what the regional average is.”

    Layer 7 (and the layer most requiring a correction): The denominator may not even be air. Colorado measurements across 37 households indicate whether water contribution is measurable depends on the ratio of water radon to soil gas radon, not absolute water radon levels; high soil gas masks water radon even up to $750 \text{ kBq/m}^3$; once soil gas is mitigated below $150 \text{ Bq/m}^3$, water contribution in laundry rooms immediately rises to 20%–40% (PMID 8056590). China’s cross-climate zone study of 171 households shows an annual mean of $81 \text{ Bq/m}^3$, winter/summer ratio of 1.68, and seasonal correction factors of 0.80–1.43 (PMID 41626976), indicating that background levels swing significantly with seasons. The term “air-water ratio” presupposes “air is the sole background,” yet the true background is soil gas; selecting the wrong denominator changes the meaning of the ratio entirely.

    To proceed correctly, perform four actions: 1) Report the transfer coefficient alongside a quartet of parameters: room volume, air exchange rate, water usage rate, and sampling point with water temperature; omit any one, and cross-study comparison becomes invalid. 2) Use geometric means with prediction intervals, not arithmetic means. 3) Decompose $f$ into calculable components ($Q \cdot \eta / V / AER$), assigning uncertainty to each term rather than providing a single total ratio. 4) Replace the denominator with the “water radon/soil gas radon” ratio; let climate and season enter the model only via AER, adding correction factors to AER rather than to $f$.

    Regarding Feng Shui, the Zang Shu states “Qi scatters when wind blows; it stops at water”—treating water as a boundary for gathering Qi, with criteria based on spatial configuration and annual scales. The $1:10,000$ ratio has no counterpart in Feng Shui; it is a conversion factor created by radiation protection for dose screening. The variable truly capable of dialogue with Feng Shui is “wind retention,” whose physical name is air exchange rate. Therefore, the question “which region’s water is more dangerous” is fundamentally misplaced; one should ask “which house keeps water more tightly sealed.” Feng Shui has always discussed houses, never regions.

    Evidence and Limitations

    Verified: Transfer coefficients span 3.3 orders of magnitude within the same state ($9.6 \times 10^{-6}–2.0 \times 10^{-2}$), driven primarily by air exchange rates and release rates (PMID 15083149); the 4.3-fold difference between bathrooms and whole houses stems from volume (PMID 24803512); tap and shower release rates are 22%–28.9% and 45%–48.3% respectively (PMID 37276723); tracer gas air exchange rate uncertainty is approx. 33% (PMID 21318005); air exchange rate correlates negatively with radon concentration (PMID 25977350); water contribution measurability depends on the water/soil gas ratio (PMID 8056590).
    Partially Verified: Regional differences in Guangdong ($6.2 \times 10^{-4}–4.1 \times 10^{-3}$) (PMID 21382658) and Bulgaria ($4.5 \times 10^{-4}–8.4 \times 10^{-3}$) (PMID 34709547) can be largely explained by “measuring bathrooms,” though differences in water intensity and temperature between regions remain un-decomposed; seasonal correction factors of 0.80–1.43 are calibrated across climate zones (PMID 41626976) but not yet jointly modeled with transfer coefficients.
    To Be Verified: Whether a significant “residual regional effect” exists after controlling for volume, AER, water usage, and sampling points has never been addressed via variance decomposition after four-variable correction; the contribution share of room parameters within regional heterogeneity of the air-water ratio lacks direct estimation.

    Related Reading

    Evidence Level: B
    References: PMID 15083149, PMID 24803512, PMID 37276723, PMID 21382658, PMID 34709547, PMID 21318005, PMID 21460366, PMID 25977350, PMID 22385122, PMID 25333994, PMID 27406637, PMID 8056590, PMID 41626976


    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.