Evidence level: B
仿真里算出来的红利是对着真分布算的;实践中你手里只有估计值——用 ĝ 收窄区间之前,先问估计噪声会不会把红利吃掉甚至反咬一口。
正文
先回顾这笔账从哪来。前面那篇 g 因子文章(95% 档高置信自由度校正的尾部函数)报告过一个反直觉结果:95% 水平下 72 个格子的校正因子 g 全部落在 0.94 到 1.02 之间,其中 w=0.75、ν=3 那一格 g=0.936——意思是修正后的区间比实际需要的还宽了约 6%。这是重尾分布的一个副产品:截尾把尾巴削掉了,但肩膀也变轻了,于是在 95% 这个不算极端的水平上,保守性反而多出了一点。我们管这叫「反向富余」——不是要加宽区间,而是有机会把区间收窄一点、让判读更锐利,覆盖还守得住 0.95。
问题在于:这个 g<1 是仿真里对着真分布算出来的。实践中没有人知道真分布,g 只能靠标定样本去估计。一旦从「真 g」换成「估计的 ĝ」,故事就完全变了——ĝ 有抽样误差,误差方向不定:这一批数据估出来 0.91,下一批可能估出 0.97。按 ĝ 收窄区间,估小了就把区间切得过窄,覆盖直接掉地板;估大了则红利缩水。红利是否可兑现、兑现几成,从未被测过。
这正是统计学里一个老问题的翻版:用估计出来的修正量去改进区间,覆盖取决于修正量本身的估计精度。Kenward-Roger 型修正之所以在小样本混合模型里被广泛采用,就是因为它对修正量(自由度与协方差的调整)做了高阶处理,使得「修正本身引入的偏差」不至反噬覆盖;而对纵向数据删失场景的对比研究也显示,协方差估计量的偏差校正方式不同,推断的覆盖表现随之不同。换句话说,「有修正」不等于「修正可白拿」——修正量的估计噪声本身就是新的误差源。另一条平行证据来自宽度与覆盖的权衡:为小样本量身定制的区间方法表明,想把区间压窄,必然要在覆盖上付出代价,关键是代价曲线的形状;少研究个数下的似然法元分析模拟同样显示,窄幅收益越大,覆盖被侵蚀的斜率越陡。
所以可行的检验设计很直接:在 95% 档对每个格子取 g 的点估计,再叠加估计噪声——ĝ 的抽样方差按标定样本量 10、30、100 三档模拟——然后按 ĝ 收窄区间,重测覆盖与平均窄幅。三档样本量对应三种物理情形:标定样本 10 个时 ĝ 噪声很大,收窄很可能切过头,覆盖掉到 0.93 以下;标定 100 个时噪声小,红利接近全额兑现,平均窄幅接近那 6%;中间档 30 个是分水岭——统计修正文献的经验是,修正量的相对标准误要压到修正幅度的一半以下,修正才净赚不赔。
还有一层不对称性值得点明:保守(区间偏宽)的代价是判读钝一点,永远不会出错;激进(收窄过头)的代价是假阳性,会直接污染结论库。所以红利兑现的举证责任在收窄方——不是「没证据说不行就收」,而是「有证据说行才收」。这与整个判读规范的设计哲学一致:宁可钝,不可错。
判读口径事先写死,分三档:若标定样本 30 以上时收窄后覆盖仍不低于 0.945、且平均窄幅至少 3%,则反向富余可写进规范,95% 档按 g 收窄;若收窄后覆盖掉到 0.93 以下,则红利不可兑现,95% 档维持原闭式不动;若窄幅超过 5% 但覆盖贴着 0.945 的地板,说明收益与风险都偏大,收窄只对 ν≥10 的小富余格子开放。
这事对气功外气实验为什么重要:外气判读依赖谐波幅值与遍数口径的 95% 区间,重尾是常态,截尾是标配。如果反向富余可兑现,同样的标定成本能换来更锐的判读边界——信号与噪音的分界线提前一格;如果不可兑现,硬收窄就是拿覆盖换好看,判出来的「显著」会是假阳性富集。账要算在兑现之前,不能算在发表之后。
证据与局限
已验证:宽度与覆盖的权衡关系有直接文献支撑;修正量(自由度、协方差)的估计精度决定推断覆盖,Kenward-Roger 类修正与偏差校正协方差估计均有系统性模拟证据。
部分验证:g 因子化框架与 72 格 g 值来自本项目内部仿真,非独立复现。
待验证:ĝ 估计噪声叠加后按 ĝ 收窄的覆盖表现(三档标定样本量、三档判据),为未做的仿真;反向富余在真实脉象/气功数据上的兑现率未知。
相关阅读
证据等级: B
参考文献: PMID 20231867, PMID 30634920, PMID 37399845, PMID 26265765
证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索
L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 与 CORRESPONDENCE_NETWORK 中无「反向富余/g 收窄/覆盖兑现」对应边;最近邻 E-060 标签信度,状态 conflict,KB-317:加权系数在「中低一致性+大样本+类别数多」时精度显著下降)。
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 \”reverse surplus\” of the 95% band be realized: can coverage still hold if narrowed by the estimated g interval?
Evidence level: B
95% Confidence Interval “Reverse Surplus”: Can It Be Realized? Narrowing the Interval Using Estimated g — Is Coverage Still Maintainable?
Simulations calculate expected gains against the true distribution; in practice, only estimated values are available. Before narrowing the interval using ĝ, one must first ask whether estimation noise will consume the expected gain—or even reverse it.
Body
First, we review the origin of this calculation. The previous article on the g-factor (the tail function with 95% confidence level, high-confidence degrees-of-freedom correction) reported a counterintuitive result: at the 95% level, all 72 grid correction factors g fell between 0.94 and 1.02, with the cell where w=0.75 and ν=3 yielding g=0.936—meaning the corrected interval became approximately 6% wider than actually required. This is a byproduct of heavy-tailed distributions: truncating the tail lightens the shoulders, so at this non-extreme 95% level, conservatism actually increases slightly. We call this “reverse surplus”—not to widen the interval, but to have the opportunity to narrow it slightly, sharpening interpretation while still maintaining coverage above 0.95.
The problem is: this g<1 was calculated in simulation against the true distribution. In practice, no one knows the true distribution; g can only be estimated via calibration samples. Once we switch from “true g” to “estimated ĝ”, the story changes completely—ĝ has sampling error, with uncertain direction: one batch of data might estimate 0.91, the next 0.97. Narrowing the interval based on ĝ risks cutting it too narrow if underestimated, causing coverage to drop; or too wide if overestimated, wasting potential gain. Whether this gain is realizable and to what extent has never been tested.
This mirrors a classic statistical problem: using an estimated correction factor to improve interval coverage depends entirely on the precision of that correction estimate itself. Kenward-Roger-type corrections are widely adopted in small-sample mixed models precisely because they apply higher-order adjustments to the correction factor (degrees of freedom and covariance), ensuring that “bias introduced by the correction itself” does not undermine coverage; comparative studies on censored longitudinal data also show that different bias-correction methods for covariance estimators lead to different inferential coverage performance. In other words, “having a correction” does not mean “the correction is free”—estimation noise in the correction factor itself becomes a new error source. A parallel line of evidence comes from the trade-off between width and coverage: interval methods tailored for small samples indicate that narrowing the interval inevitably incurs a cost in coverage, with the shape of the cost curve being critical; likelihood-based meta-analysis simulations similarly show that under fewer studies, the steeper the slope of coverage erosion as narrowness gains increase.
Thus, the feasible test design is straightforward: at the 95% level, take point estimates of g for each cell, then layer on estimation noise—simulating ĝ’s sampling variance across calibration sample sizes of 10, 30, and 100—then narrow the interval based on ĝ and retest coverage and average narrowness. These three sample sizes correspond to three physical scenarios: with only 10 calibration samples, ĝ noise is large, narrowing likely cuts too much, dropping coverage below 0.93; with 100 samples, noise is minimal, gain approaches full realization, and average narrowness nears that 6%; the middle tier of 30 samples serves as the threshold—statistical correction literature suggests that for a correction to net positive, the relative standard error of the correction factor must be less than half the magnitude of the correction itself.
One additional asymmetry deserves emphasis: the cost of conservatism (interval too wide) is merely duller interpretation, never wrong; the cost of aggression (narrowing too much) is false positives, which directly pollute the conclusion pool. Therefore, the burden of proof for realizing gain lies with those proposing narrowing—not “narrow unless evidence says no,” but “narrow only if evidence says yes.” This aligns with the overall design philosophy of interpretation norms: better dull than wrong.
Interpretation criteria are fixed in advance across three tiers: if calibration samples ≥30 and post-narrowing coverage remains ≥0.945 with average narrowness ≥3%, reverse surplus can be incorporated into the norm, and the 95% level uses g-narrowing; if post-narrowing coverage drops below 0.93, the gain is unrealizable, so the 95% level retains its original closed-form; if narrowness exceeds 5% while coverage hovers just above the 0.945 floor, both gain and risk are excessive, so narrowing is opened only for cells with ν≥10 exhibiting slight surplus.
This matters critically for qigong external qi experiments: external qi interpretation relies on the 95% interval of harmonic amplitude and cycle count, where heavy tails are normal and truncation is standard. If reverse surplus is realizable, the same calibration cost yields sharper interpretation boundaries—the signal-noise threshold advances by one grid; if unrealizable, forced narrowing trades coverage for aesthetics, making declared “significance” a false-positive enrichment. The account must be settled before publication, not after.
Evidence and Limitations
Verified: Direct literature supports the trade-off between width and coverage; estimation precision of correction factors (degrees of freedom, covariance) determines inference coverage, with systematic simulation evidence supporting Kenward-Roger-type corrections and bias-corrected covariance estimation.
Partially Verified: The g-factor framework and the 72-cell g-values originate from internal simulations within this project, not independent replication.
Pending Verification: Coverage performance after ĝ estimation noise superposition and subsequent narrowing by ĝ (calibrated across three sample size tiers and three decision criteria) remains un-simulated; the realization rate of reverse surplus on real pulse/qigong data is unknown.
Related Reading
- Tail function for 95% confidence interval height correction with degrees-of-freedom adjustment: source of negative surplus
- Stratum spacing design for correction factor stratification: Pareto frontier
- Research pool: resonant consciousness field
Evidence level: B
References: PMID 20231867, PMID 30634920, PMID 37399845, PMID 26265765
Evidence level: A = direct PubMed support; B = scientific analogy or indirect evidence; C = traditional consensus or preliminary exploration
L1 network status: This topic has not yet entered the L1 corresponding network (no edges for “reverse surplus / g narrowing / coverage fulfillment” exist in EDGE_TABLE and CORRESPONDENCE_NETWORK; the nearest-neighbor E-060 label has low credibility, status conflict, KB-317: weighted coefficient accuracy decreases significantly under conditions of moderate-to-low consistency plus large sample size plus many categories).
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