Evidence level: B
上一篇把 C-A 差推成候选预测指标,这一篇问一个更靠地基的问题:那个 20 dB 到底是一条线,还是一段坡。
正文
上一篇论证了 C-A 差(dBC 与 dBA 之差)可能比绝对声压级更能预测办公楼烦扰。但它脚下踩着一块没人验过的地基:20 dB 这个切点,究竟是物理断点,还是被反复抄写的经验整数?
一、同一个现象,五个切点
如果 20 dB 是物理断点,它不该是个可以商量的数。可工程界并不统一:DIN 45680 与 Alberta 监管口径取 20 dB,澳大利亚新南威尔士的工业噪声政策取 15 dB,Acoustics Australia 的综述列完全部出处后建议”至少 20 dB”,同时提到有人主张 25 dB;上海交大对住宅固定设施的研究给出 15 dB。同一物理量,切点跨度 10 dB。 这不像阈值的常态,更像判据的常态——20 dB 是一个被抄写的整数,不是被检验出的界线。
二、两种形状,两种政策
这是”函数形式”之争,而两个候选模型对世界的说法相反。阈值模型说:20 dB 以下无效应,越过才有——管线把楼压到 19 dB 就交差。连续梯度模型说:效应随差值平滑上升,每一分贝都算——最划算的减排恰恰在线以下。
空气污染领域早打过这一仗。NMMAPS 在 20 个美国城市用对数线性、样条、阈值三模型对比,结论是无阈值、线性够用;法国 18 城(PMID 36602605)把四种函数形式放在同一批数据上,得到”超线性”——低浓度端更陡,中段趋平。那不是断点,是弯。 而这两种形状在同一张散点图上都能拟合得像模像样。
三、分段回归的致命陷阱
即便有人”检验过” 20 dB,也不等于有证据。断点在零假设下不可识别——它只在”确实存在断点”的备择假设里才有定义,常规 t 检验与 Wald 检验在此失效,必须用 Davies 检验或 Muggeo 伪分数检验校正零分布;而 segmented 类算法对初值极敏感,换个起点就收敛到另一个局部极小。没做这一步而报出的”阈值位于 X dB”,报的是一个没有零分布的 p 值。
四、现有证据更像斜坡
PMID 39147186 用生态瞬时评估加便携声级计追踪真实时段的烦恼,非线性拐点出现在实时 60 dB、回忆 70 dB,且随活动情境漂移——拐点位置不是声学常数,是活动类型的函数。PMID 36586852 研究飞机起飞低频声,把室外 C-A 差作第二个连续预测因子,解释方差直接到 98%,全程未引入任何断点。连续变量已经够用,切成二值只是丢信息。
五、低频端本来就是陡的
方向上也对不上。Møller 与 Pedersen 的综述(PMID 15273024)指出低频烦扰随声级上升得比中高频更快。若曲线在低频端是凸的,”低端平坦、越线才起效”的图像在方向上就是反的——低端不是平台,是陡坡。
六、但”没测到”不等于”不存在”
必须留退路。PMID 32018132 系统展示了代理暴露的测量误差如何结构性抹平低端非线性;NMMAPS 的专家委员会也专门警告:暴露测量误差可能掩盖真实阈值。所以本轮判决只有一条——20 dB 目前不可判定,而不是已被证伪。当已验证的因果阈值是越界,当已被推翻同样是越界。
七、判决实验
选 8-12 栋频谱形状差异大的办公楼,每栋 20-40 人;C-A 差实测分布须横跨 15-40 dB,每 5 dB 单元不少于 20 人——只在 20 dB 附近堆点的样本永远判不出来。同一批数据平行拟合四个模型:线性、先验 20 dB 分段、无先验断点分段(segmented + Davies)、限制性立方样条。结局用 ICBEN 连续评分,不二分。
核心判据:Davies p>0.05 且样条在 20 dB 附近斜率变化不足 30% ⇒ 判连续梯度,20 dB 降为工程筛查线、不得作因果阈值;Davies p<0.05、估计断点落在 18-22 dB 且 bootstrap 区间宽小于 4 dB ⇒ 断点成立、可写进标准;断点漂到 30 dB 以外或区间宽过 10 dB ⇒ 判断点不可识别、只报连续剂量-反应;样条显示低端更陡 ⇒ 干预优先级翻转,先救低端而非先压超标楼。
风水线上,”煞”本来就并存两套直觉:一套门槛式(”路冲不过三丈”),一套坡度式(”煞有轻重,日积月累”)。本文不站队,只把它变成能被数据裁决的统计问题——这是”化煞”从口诀走向剂量学的必经一步。
证据与局限
已验证: 20/15/25 dB 三个切点在同一工程领域并存,无统一实证来源;阈值模型与连续模型在污染流行病学中已被系统对撞过(PMID 36602605 四模型对比得超线性而非断点);噪声烦恼的非线性拐点随活动情境漂移(PMID 39147186);C-A 差作连续预测因子解释方差达 98%、无需断点(PMID 36586852);低频烦扰随声级上升比中高频更快(PMID 15273024)。
部分验证: C-A 差对烦扰解释力”显著但很小”(Kjellberg 1997,经 PMID 37818147 转述)——提示其单独解释力有限,增量可能来自函数形式误设而非指标本身;断点方法学(Davies 检验、Muggeo 线性化、bootstrap 区间)已成熟,但尚无任何低频噪声研究正式套用。
待验证: 20 dB 在多楼宇样本上是否对应真实断点;断点估计的置信区间宽度是否足以支撑一个监管数字;C-A 差与烦扰的关系究竟是超线性、阈值还是线性——目前无直接研究,属 P2 池子空白。
相关阅读
证据等级: B
参考文献: PMID 36602605, PMID 39147186, PMID 36586852, PMID 15273024, PMID 32018132, PMID 37818147
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。
20 dB is a breakpoint or a slope: adjudicating the functional form of the criterion for low-frequency \”spectral imbalance\
Evidence level: B
The previous article promoted the C-A difference as a candidate predictive metric; this one asks a more foundational question: is that 20 dB figure a straight line or a segment of a slope.
Main Text
The previous article argued that the C-A difference (the gap between dBC and dBA) may predict office annoyance better than absolute sound pressure levels. However, it stands on unverified ground: Is the 20 dB cutoff a physical discontinuity or a repeatedly copied empirical integer?
I. One Phenomenon, Five Cutoffs
If 20 dB were a physical discontinuity, it should not be negotiable. Yet engineering standards diverge: DIN 45680 and Alberta regulations use 20 dB; New South Wales, Australia’s industrial noise policy uses 15 dB; Acoustics Australia’s review lists all sources and suggests “at least 20 dB,” while noting some advocate for 25 dB; Shanghai Jiao Tong University’s study on residential fixed installations yields 15 dB. The same physical quantity spans a 10 dB range for the cutoff. This resembles the norm for criteria rather than thresholds—20 dB is a copied integer, not an empirically verified boundary.
II. Two Shapes, Two Policies
This is a debate over “functional form,” where two candidate models offer contradictory accounts of reality. The threshold model states: no effect below 20 dB; effects only emerge after crossing the line—compliance is achieved if a building stays at 19 dB. The continuous gradient model states: effects rise smoothly with the difference; every decibel counts—the most cost-effective reduction occurs precisely below the line.
This battle has been fought in air pollution. NMMAPS compared log-linear, spline, and threshold models across 20 US cities; the conclusion was no threshold, linear suffices. France’s study of 18 cities (PMID 36602605) fitted four functional forms to the same dataset, finding “super-linear”—steeper at low concentrations, flattening in the mid-range. That is not a cliff, but a curve. Yet both shapes fit the same scatterplot with equal plausibility.
III. The Fatal Trap of Piecewise Regression
Even if someone claims to have “verified” 20 dB, that does not equate to evidence. A discontinuity is unidentified under the null hypothesis—it only has definition under the alternative hypothesis that a discontinuity indeed exists; conventional t-tests and Wald tests fail here, requiring Davies testing or Muggeo pseudo-score testing to correct the null distribution; segmented algorithms are extremely sensitive to initial values, converging to different local minima with a changed starting point. Reporting a “threshold at X dB” without this step reports a p-value with no null distribution.
IV. Existing Evidence Suggests a Slope
PMID 39147186 used ecological instantaneous assessment with portable sound level meters to track annoyance during real-time periods; the non-linear inflection point appeared at 60 dB real-time and 70 dB recall, drifting with activity context—the inflection point is not an acoustic constant but a function of activity type. PMID 36586852 studied low-frequency noise from aircraft takeoffs, using outdoor C-A difference as a second continuous predictor, explaining 98% of variance without introducing any discontinuity. Continuous variables suffice; dichotomizing discards information.
V. The Low-Frequency End Is Naturally Steep
The direction also mismatches. Møller and Pedersen’s review (PMID 15273024) notes that low-frequency annoyance rises faster with sound level than mid-to-high frequencies. If the curve is convex at low frequencies, an image of “flat at the bottom, effective only after crossing” is reversed—the low end is not a plateau but a steep slope.
VI. “Not Measured” Does Not Equal “Does Not Exist”
A retreat must be left open. PMID 32018132 systematically demonstrates how measurement error in proxy exposure structurally flattens low-end non-linearity; the NMMAPS expert committee also specifically warns that exposure measurement error may mask true thresholds. Thus, this round’s verdict has only one rule: 20 dB is currently indeterminate, not refuted. When a verified causal threshold is crossing the line, and when a debunked one is also crossing the line.
VII. Judgment Experiment
Select 8–12 office buildings with large spectral shape differences, each housing 20–40 people; the measured C-A difference distribution must span 15–40 dB, with at least 20 people per 5 dB unit—samples only piling up near 20 dB will never yield a verdict. Fit four models in parallel on the same data: linear, prior 20 dB piecewise, no-prior-discontinuity piecewise (segmented + Davies), restricted cubic spline. Outcomes use ICBEN continuous scoring, not binary.
Core criteria: If Davies p > 0.05 AND spline slope change near 20 dB is <30% ⇒ judge continuous gradient, reduce 20 dB to an engineering screening line, do not treat as causal threshold; If Davies p < 0.05, estimated discontinuity falls within 18–22 dB AND bootstrap interval width is <4 dB ⇒ discontinuity confirmed, can be written into standards; If discontinuity drifts beyond 30 dB OR interval width exceeds 10 dB ⇒ judge discontinuity unidentified, report only continuous dose-response; If spline shows steeper low-end ⇒ intervention priority flips, save the low end before pressuring over-compliant buildings.
In Feng Shui, “Sha” inherently coexists with two intuitions: one threshold-based (“road collision not exceeding three zhang”), one gradient-based (“Sha has weight, accumulating over time”). This article does not take sides but transforms it into a statistical problem adjudicated by data—this is the necessary step for “neutralizing Sha” moving from mnemonic rules to dosimetry.
Evidence and Limitations
Verified: The coexistence of 20/15/25 dB cutoffs within the same engineering domain lacks unified empirical sources; threshold models and continuous models have been systematically clashed in pollution epidemiology (PMID 36602605 found super-linear rather than discontinuity); non-linear inflection points in noise annoyance drift with activity context (PMID 39147186); C-A difference as a continuous predictor explains 98% variance without needing discontinuities (PMID 36586852); low-frequency annoyance rises faster with sound level than mid-to-high frequencies (PMID 15273024).
Partially Verified: C-A difference’s explanatory power for annoyance is “significant but small” (Kjellberg 1997, recounted via PMID 37818147)—suggesting limited standalone explanatory power, with incremental gains possibly arising from misspecified functional forms rather than the metric itself; discontinuity methodology (Davies testing, Muggeo linearization, bootstrap intervals) is mature, yet no low-frequency noise study has formally applied it.
To Be Verified: Whether 20 dB corresponds to a true discontinuity across multi-building samples; whether the confidence interval width of discontinuity estimates is sufficient to support a regulatory number; whether the C-A difference–annoyance relationship is super-linear, threshold-based, or linear—currently no direct research exists, representing a P2 pool gap.
Related Reading
- Low-Frequency “Spectral Imbalance”: A More Accurate Office Annoyance Predictor Than Absolute Sound Pressure
- Subthreshold Building-Vibration Dose-Response Curves for Office Populations
- Systematic Comparative Measurement of Building Form Sha Ratings vs. Indoor Infrasound Frequency-Intensity Spectra
- Research Pool: Environment-Life Frequency Interface
Evidence Level: B
References: PMID 36602605, PMID 39147186, PMID 36586852, PMID 15273024, PMID 32018132, PMID 37818147
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply