当坐标模糊到和网格一样大:文化点 × 地生氡潜力叠加检验的定位误差敏感性分析

Written by

in

Language

Evidence level: B

一个”县”级坐标落在 10 公里网格的哪一格,很可能由测绘本身体决定,而不是由风水决定。

正文

上一篇把实验摆上了桌:中国风水穴、穴位地名、名医故里数字化成点,投到欧洲地生氡潜力(GRP)栅格上做蒙特卡洛置换检验。那篇留了个尾巴——方志里的”县”只有公里级精度,而 GRP 图是 10 公里网格,两者同量级。这一篇专治这个尾巴。

先说结论:如果文化点的定位误差与栅格尺度同量级,”落在哪一格”这个动作本身就成了随机抽取;置换检验做得再漂亮,检验的也可能只是”我们怎么测坐标”,而不是”古人怎么选址”。

一、定位误差有多大,有硬数字

不是”可能有点偏”。街道级地理编码实测误差中位数 41 米、99 分位 273 米,而且这些误差不是随机的——250 米以内它系统性高估暴露人数,假阳假阴同时大量出现(PMID 17367533)。农村更差:不做地址规范化,25% 的点偏离真位超过 1000 米(PMID 20406495);法国 E3N 队列 4247 条记录里,城市地址能落到门牌或街道段一级的占 74.4%,农村只有 10.5%(PMID 28235407);美国农业健康研究两轮复测,农村与低密度区误差系统性更大(PMID 33572119)。

换算一下:方志”县治”级坐标精度通常几公里到十几公里,村一级一到三公里。欧洲 GRP 图是 10×10 公里网格(PMID 23295644、25063783)。误差半径约等于网格边长。 一个点相当大概率落在它”应该”在的那一格之外。

二、为什么这次不能用”大样本”救

经典测量误差会衰减效应量,样本够大仍能检出衰减后的真效应。空间定位误差是另一种东西。系统综述把它分成四类,从单纯的点位误差一路排到协变量与结局的空间错位(PMID 27368370)。误差一旦非随机,就会造出关联,而不只是削弱关联。而地理编码误差恰恰是结构性的:越偏远越糙,越靠路越容易被”吸附”到路段中点。这与”名医故里集中于江南水乡”这类聚集结构同向——误差方向和被检验的空间结构耦合,样本放大只会让偏倚更”显著”,不会让它消失。

三、抖动检验:四步,两只臂

第一步,给每个点加随机扰动,半径取 1/5/10/20 公里,每档重复千次。第二步,每次重跑完整的置换检验——不只重算 p,而要重跑零分布,因为点位变了,聚簇结构也跟着变。第三步,记录 p 值、效应方向,以及 GRP 值的实际变化率。第四步最容易被跳过——同时开一只”零效应臂”:把点位随机撒在同一研究区内(真实关联人为设零),走完全相同的流程,看它有多大比例在抖动后变显著。这是假翻转率,减掉它剩下的才是真损失。

判据写死:若 5 公里扰动下翻转率超过 20%,或零效应臂假显著率超过 10%,则该检验在此坐标精度下不可判,须降级为村级子样本探索分析,或改做区域级聚合(并接受 MAUP 的另一套代价);若 5 公里扰动下 GRP 值变化率中位数低于 10% 且翻转率低于 5%,方可报告主结论。

四、三个容易骗自己的地方

其一,”抖动后还显著”常常是假稳健。 GRP 是空间自相关场,相邻格高度相似。抖动 5 公里往往还落在同一高值斑块里,GRP 值几乎不变,检验当然不翻。所以要报的是 GRP 值变化率,不是 p 值稳不稳。地统计里早有对应做法:用条件模拟生成一组等概率风险图,逐张送进同一套聚类算子,看分类在多大比例的模拟里成立——结果某些”显著低风险聚集”只有少数县在九成以上模拟中显著,另一处高值聚集则在六成模拟图中出现(PMID 16469095)。这才叫”结论的稳定度”。

其二,尺度本身会改结论。 空间聚合的尺度与分区(MAUP)能让回归结果整体偏移:同一份渥太华 NO₂ 与呼吸系统疾病数据,换三种空间结构,OLS 估计显著不同,而三种结构最终都测不到效应(PMID 22040001);”对 MAUP 有多敏感”本身也已可做成统计检验(S-maup,PMID 30481193),给出”最多聚合到哪一级还不失真”的答案。县级坐标是一种聚合,10 公里网格是另一种聚合,两层叠在一起,误差不是相加,而是相乘。

其三,地图的”分辨率”不等于”精度”。 同一地区换数据源做 GRP 图,预测力差别很大:挪威航空伽马能谱把解释方差做到约 70%,纯地质图只有约 40%,且前者能给每个估计配上 95% 置信限(PMID 27105766);爱尔兰的详细制图同时输出预测不确定性(PMID 36497982);另一项研究里仅用地物理数据的模型无法准确预测土壤氡活度,必须整合地质信息才解释约 75% 的变异(PMID 37343860)。这些不确定性图层正是抖动检验最好的输入——扰动幅度应按每格的预测方差抽样,而非全局固定半径。

还有地形。三维地表投到平面坐标时面积与距离都不守恒,这种投影失真会系统性把点判成”聚集”(PMID 28452047)。名医故里扎堆江南,可能只是江南本来就点密。

收尾:这个检验最诚实的表述,不是”古人避开了高氡带”,而是”在我们能测到的坐标精度之内,古人的选址与高氡带的关系是可判或不可判”。把”不可判”当结论写出来,比一个漂亮的 p 值更有价值——它明确告诉你下一步该补的是族谱与方志的村级定位,而不是再换个更花的模型。

证据与局限

已验证: 地理编码位置误差的量级与城乡差异有实测数据(PMID 17367533、20406495、28235407、33572119);空间测量误差可系统分为四类并被统一建模(PMID 27368370);MAUP 的尺度与分区敏感性可量化、可检验(PMID 22040001、30481193、30917821);GRP 制图的分辨率与预测不确定性有成熟方法(PMID 23295644、25063783、27241368、36497982、37343860、27105766);条件模拟能把空间不确定性传播到聚类分类结论上(PMID 16469095)。

部分验证: “抖动-翻转率”作为稳健性指标,在生态点过程(环形位移零模型、地形校正零模型,PMID 28452047)与疾病制图(PMID 16469095)中已有成熟工具,空间分辨率对时空聚集检测的影响也做过系统模拟(PMID 23110167);但它从未被用在”跨文化坐标叠加”这类场景里。

待验证: 中国文化点县/村级坐标的真实误差分布(须以 GPS 实测抽样标定,不能直接套用欧美地理编码数据);误差与本底空间聚集的耦合方向(是系统抹平还是系统造出信号);以及”20% 翻转率”这类阈值本身,仍需通过零效应模拟标定后才能定稿。

相关阅读

证据等级: B(定位误差、MAUP、GRP 制图不确定性各有 PubMed 直接支撑;”抖动-翻转率”用于本场景属方法移植,无先例)
参考文献: PMID 17367533, PMID 20406495, PMID 28235407, PMID 33572119, PMID 27368370, PMID 22040001, PMID 30481193, PMID 30917821, PMID 16469095, PMID 23295644, PMID 25063783, PMID 27241368, PMID 36497982, PMID 37343860, PMID 27105766, PMID 28452047, PMID 23110167


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。

Positioning Error Sensitivity Analysis When Coordinates Blur to Grid Scale: A Superposition Test of Cultural Sites and Hydrogen-Rich Groundwater Potential

Evidence level: B

A “county-level” coordinate falling into which cell of a 10 km grid is likely determined by the cartographer’s body, not by Feng Shui.

Main Text

The previous article placed the experiment on the table: Chinese Feng Shui sites, acupoint place names, and hometowns of famous physicians were digitized into points and subjected to Monte Carlo permutation testing against European Geogenic Radon Potential (GRP) raster grids. That article left a tail—county-level gazetteer coordinates have kilometer-level precision, while GRP maps are 10 km grids; the two are of the same order of magnitude. This article addresses that tail specifically.

Conclusion first: If the localization error of cultural points is commensurate with grid scale, the act of “falling into which cell” itself becomes random sampling. No matter how beautiful the permutation test is, it may only test “how we measured coordinates” rather than “how ancients selected sites.”

1. How Large Is the Localization Error? Hard Numbers

Not “possibly slightly off.” In street-level geocoding, the median实测 error is 41 meters with a 99th percentile of 273 meters, and these errors are not random—within 250 meters they systematically overestimate exposure counts, producing both false positives and false negatives (PMID 17367533). Rural areas are worse: without address normalization, 25% of points deviate from the true location by over 1000 meters (PMID 20406495); in the French E3N cohort of 4,247 records, urban addresses could be mapped to the door number or street segment level in 74.4% of cases, whereas rural addresses only achieved this in 10.5% (PMID 28235407); two rounds of re-measurement in the US Agricultural Health Study showed systematically larger errors in rural and low-density areas (PMID 33572119).

Converted: Gazetteer “county seat” coordinates typically have precision of several kilometers to over ten kilometers; village-level precision ranges from one to three kilometers. European GRP maps use a 10×10 km grid (PMID 23295644, PMID 25063783). The error radius approximates the grid side length. A point has a high probability of falling outside the cell it “should” occupy.

2. Why Large Samples Cannot Save This Time

Classic measurement errors attenuate effect sizes; with sufficiently large samples, the attenuated true effect can still be detected. Spatial localization errors are a different category. Systematic reviews categorize them into four types, ranging from simple point errors to spatial misalignment between covariates and outcomes (PMID 27368370). Once errors are non-random, they create associations rather than merely weakening them. Geocoding errors are precisely structural: the more remote and rough, the worse; the closer to a road, the more likely they are “adsorbed” into the segment midpoint. This aligns with aggregation structures such as “famous physician hometowns concentrated in Jiangnan water towns”—the error direction couples with the spatial structure being tested. Amplifying the sample size will only make the bias more “significant,” not cause it to disappear.

3. Jitter Testing: Four Steps, Two Arms

Step 1: Add random perturbation to each point with radii of 1/5/10/20 km, repeating 1,000 times per tier. Step 2: Re-run the complete permutation testing each time—not just recalculating p-values, but re-running the null distribution, because point locations change and cluster structures follow. Step 3: Record p-values, effect directions, and the actual rate of change in GRP values. Step 4 is most easily skipped—open a second “null-effect arm”: randomly scatter points within the same study area (artificially setting true association to zero) and follow the exact same process to see what proportion becomes significant after jittering. This is the false flip rate; subtract it from the remainder to get the true loss.

Rule fixed: If the flip rate exceeds 20% under 5 km perturbation, or if the null-effect arm false-significance rate exceeds 10%, then this test is unjudgable at this coordinate precision and must be downgraded to village-level subsample exploration analysis, or shifted to regional aggregation (accepting another set of MAUP costs); only if the median GRP value change rate is below 10% and flip rate is below 5% under 5 km perturbation may the main conclusion be reported.

4. Three Places That Easily Deceive You

First, “Significance after jitter” is often false robustness. GRP is a spatially autocorrelated field where adjacent cells are highly similar. A 5 km jitter often still falls within the same high-value patch, leaving GRP values nearly unchanged; thus, the test naturally does not flip. Therefore, one must report the GRP value change rate, not whether p-values remain stable. Spatial statistics already have corresponding methods: use conditional simulation to generate a set of equally probable risk maps, feed each into the same clustering operator, and observe in what proportion of simulations a classification holds—results show that certain “significant low-risk clusters” are significant in over 90% of simulations for only a few counties, while another high-value cluster appears in 60% of simulated maps (PMID 16469095). Only then is this called “conclusion stability.”

Second, scale itself can alter conclusions. Spatial aggregation scales and zoning (MAUP) can shift regression results overall: the same Ottawa NO₂ and respiratory disease dataset, with three different spatial structures, yields significantly different OLS estimates, yet none of the three structures detect an effect (PMID 22040001); “sensitivity to MAUP” itself can be made into a statistical test (S-maup, PMID 30481193), providing an answer of “up to which aggregation level does not distort truth.” County-level coordinates represent one type of aggregation; the 10 km grid represents another. Stacked together, errors do not add but multiply.

Third, a map’s “resolution” does not equal “precision.” Switching data sources for GRP maps in the same region yields vastly different predictive powers: Norwegian airborne gamma spectrometry achieves explanatory variance of approximately 70%, while pure geological maps only achieve about 40%, and the former can assign 95% confidence limits to each estimate (PMID 27105766); Ireland’s detailed mapping simultaneously outputs prediction uncertainty (PMID 36497982); in another study, models using only geophysical data cannot accurately predict soil radon activity and must integrate geological information to explain approximately 75% of the variance (PMID 37343860). These uncertainty layers are precisely the best inputs for jitter testing—perturbation amplitude should be sampled from the prediction variance per cell, not a global fixed radius.

There is also terrain. Projecting a three-dimensional surface onto planar coordinates does not conserve area or distance; such projection distortion systematically misclassifies points as “clusters” (PMID 28452047). Famous physician hometowns clustering in Jiangnan may simply reflect that Jiangnan is denser in points to begin with.

Wrap-up: The most honest statement of this test is not “ancients avoided high-radon zones,” but “within the coordinate precision we can measure, the relationship between ancient site selection and high-radon zones is either judgable or not.” Writing “not judgable” as a conclusion holds more value than a pretty p-value—it explicitly tells you the next step is to supplement gazetteer data with village-level positioning from genealogies, not to switch to a more flashy model.

Evidence and Limitations

Verified: Magnitude of geocoding location errors and urban-rural differences have empirical data support (PMID 17367533, PMID 20406495, PMID 28235407, PMID 33572119); spatial measurement errors can be systematically categorized into four types and unified modeling (PMID 27368370); MAUP scale and zoning sensitivity can be quantified and tested (PMID 22040001, PMID 30481193, PMID 30917821); GRP map resolution and prediction uncertainty have mature methodologies (PMID 23295644, PMID 25063783, PMID 27241368, PMID 36497982, PMID 37343860, PMID 27105766); conditional simulation can propagate spatial uncertainty to clustering classification conclusions (PMID 16469095).

Partially Verified: “Jitter-flip rate” as a robustness indicator has mature tools in ecological point processes (circular displacement null model, terrain correction null model, PMID 28452047) and disease mapping (PMID 16469095); the impact of spatial resolution on spatiotemporal clustering detection has also been systematically simulated (PMID 23110167); however, it has never been applied to “cross-cultural coordinate overlay” scenarios.

To Be Verified: True error distribution of Chinese cultural point county/village coordinates (must be calibrated via GPS实测 sampling, cannot directly apply US/European geocoding data); coupling direction between errors and background spatial aggregation (systematically flattening or systematically creating signals); and thresholds like “20% flip rate” themselves still require finalization through null-effect simulation calibration.

Related Reading

Evidence Level: B (Localization errors, MAUP, and GRP mapping uncertainty each have direct PubMed support; “jitter-flip rate” application in this scenario is method transplantation with no precedents)
References: PMID 17367533, PMID 20406495, PMID 28235407, PMID 33572119, PMID 27368370, PMID 22040001, PMID 30481193, PMID 30917821, PMID 16469095, PMID 23295644, PMID 25063783, PMID 27241368, PMID 36497982, PMID 37343860, PMID 27105766, PMID 28452047, PMID 23110167


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *