Evidence level: C
例数口径的代价不是 9%,而是 2% 到 3%;它不是由相关性强弱决定,而是由例数多少决定——把区间做窄的最快办法是加例数,不是标定折减系数。
正文
上一轮 #1559 给出了一个让人安心的结论:气功外气判读的群体均值检验,从「按例数报自由度」这个保守口径起步,置信区间覆盖率在所有会话内相关水平上都稳稳停在 0.95,标定错也不会伤到结论。但保守是有价格的:把每例的重复遍数只当成一个数来用,那些遍数携带的信息是不是被白白扔掉了?backlog 的预设估算是「当会话内相关只有 0.1、折减系数 0.83 时,CI 比标定后折减口径宽约 9%」。本轮就用解析推导加蒙特卡洛仿真,把这个价格精确量出来,并给出「标定实验到底值不值得做」的量化门槛。
先看价格的结构。两个口径估计的是同一个量——群体均值的标准误,而标准误的方差标度两侧完全相同:例数口径用每例均值,其方差自动携带设计效应 (1+(r−1)ρ)/r;折减口径用全部读数加设计效应校正,方差标度也是同一个式子。也就是说,保守口径的「浪费」不在方差标度上,只在一个地方:自由度。例数口径 df = n−1,折减口径 df = n·r/(1+(r−1)ρ) − 1,两者的 t 分位数不同,宽度比几乎就等于两个 t 分位数的比值。这是一个可以手算的结论,仿真只是确认它。
仿真网格覆盖 r = 2/3/5/10、ρ = 0.05/0.1/0.2/0.3、n = 5 到 50,每组 20 万次重复,另加一层「覆盖率校平」的公平比较——把折减口径的区间统一缩放到实际覆盖 0.95 后再比宽度,消除任何虚胖精度。结果分三层。第一层,主工作区(n=30、r=3、ρ<0.3)的代价只有 2.0% 到 2.2%,即使把遍数拉到 10、相关压到 0.05,也才 3.0%;n=20 时不超过 5%。backlog 预设的「约 9%」明显高估了——9% 那个数字其实是设计效应相对「把遍数当独立观测」这一错误名义口径的膨胀率,不是保守口径相对标定口径的代价,两者被混成了一条线。
第二层,代价几乎与 ρ 无关,只与 n 有关。同一个 (n, r) 下,ρ=0.05 和 ρ=0.3 的宽度比相差不足 1 个百分点。原因还是结构性的:ρ 以同样的方式进入两侧口径的方差标度,互相抵消,剩下的只有它对折减自由度的微弱拉动。换句话说,「高遍数低相关场景代价最大」这个直觉只说对了一半——高遍数确实放大代价(r=10 比 r=3 贵约 0.8 个百分点),但低相关本身几乎不抬价。真正让代价起飞的是例数太少:n=10 时代价跳到 8% 到 11%,n=7 时 12% 到 22%,n=5 时高达 21% 到 37%。
第三层,代价超过 25% 的区域确实存在,但只出现在 n 不超过 5 的极端小样本(r≥5 时 n=5 超线,r=3 时 n=4 超线)。这个区域里最便宜的解决方案不是标定折减系数,而是加例数:n 从 5 加到 7,代价立刻从 25% 以上掉到 12% 到 20%,加两个例子的收益远超任何一次标定实验。而且要注意,上述对比还假设标定后的 ρ 是已知常数——真实标定实验里 ρ 本身带估计误差,折减口径的实际收益还要再打折扣,保守口径的相对优势只会更大。
落到立项判断上,backlog 的两条判据有明确答案。常用工作区(r=3、ρ<0.3、n≥20)实测代价 2% 到 5%,远低于 10% 门槛——「例数口径起步规则全场景可用」成立,标定实验正式降级为研究性工作,不设立项优先级。代价超 25% 的区域存在但极其狭窄(n≤5),且该区域的正确解是先把例数补到 7 以上而不是去标定。把两条合起来:标定实验从「最好有」降为「研究性」,新研究一律例数口径起步,样本量按 n(例数)而不是 n×r(例数乘遍数)计算。信度样本量规划领域确有成熟精度框架(PMID 24338600,2014),样本量对重测信度估计的实质影响也有 2025 年的直接证据(PMID 40119099),这两个方法恰好可以反过来用于回答「多少例才够」:把仿真里的 n 换成招募成本,就是每个项目自己的标价表。
这个结论有一层更稳的底。本系列 #1559 已经证明例数口径的覆盖率不依赖 ρ,本轮又证明它的宽度代价也几乎不依赖 ρ——两个性质加在一起,意味着这套规则对「判读者的会话内相关到底是多少」这个未知数完全免疫。在一个连 ρ 取值范围都只能靠猜的研究领域(#1559 引用的压力痛阈文献报过 0.73 的会话内重复相关,PMID 33767329),能有一把对未知参数天然免疫的尺子,本身就是最值钱的资产的来源。剩下的唯一缺口是:例数口径把判读遍数的贡献压缩在每例均值里,代价以 t 分位数的形式存在——这个缺口是数学结构,不是工程缺陷,补它只能靠例数,这也是全系列最终给出的一致处方。
证据与局限
已验证: 蒙特卡洛(20 万次/组,覆盖率校平后公平比较)给出全网格代价:n≥20 全部低于 10%,主工作区 2% 到 5%;代价随 n 单调上升、随 ρ 近似不变;超 25% 区仅存于 n≤5。解析层确认宽度比≈两口径 t 分位数之比,方差标度两侧相同。
部分验证: 折减口径假设 ρ 已标定且为常数;真实标定带估计误差,实际折减收益只会更小,方向性结论模型无关。
待验证: 气功外气判读场景的会话内相关实测值仍未取得(同 #1559/#1564);若未来实测 ρ 显示判读者间折减差异悬殊,「折减随读者经验漂移」分支仍需单独裁决。
相关阅读
L1 网络状态:该主题(例数口径 vs 标定折减口径的精度代价)在 CORRESPONDENCE_NETWORK / EDGE_TABLE 中无对应边;检索「折减/有效自由度/例数/精度代价/标定」零命中。最近邻为 E-060 标签前提边(conflict,讲人际一致性低,PMID 30383398)与 KB-510(深呼吸 HRV 静息 ICC 0.50–0.83,讲读数稳定性),均为「读数信度」邻近议题而非本命题的会话内结构命题。该主题尚未进入 L1 对应网络,本文证据等级封顶 C。
证据等级: C
参考文献: PMID 24338600, PMID 40119099, PMID 33767329
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 does the conservative limit cost: reporting degrees of freedom by count, and how much wider is the confidence interval compared to the calibrated reduction
Evidence level: C
The cost of the sample-count approach is not 9%, but rather 2% to 3%; it is determined by the number of samples, not by the strength of correlation—the fastest way to narrow the interval is to add samples, not to apply a calibration discount factor.
Body
The previous post #1559 provided a reassuring conclusion: the population mean test for Qigong external air judgment, starting from the conservative “sample-count” approach, maintained a confidence interval coverage rate of 0.95 across all levels of within-session correlation, and calibration errors did not harm the conclusion. However, conservatism comes at a price: treating the repetition count per sample merely as a number means that the information carried by those repetitions is seemingly wasted. The backlog’s preset estimate was “when within-session correlation is only 0.1 and the discount factor is 0.83, the CI is about 9% wider than the calibrated-discounted approach.” This round uses analytical derivation plus Monte Carlo simulation to precisely quantify this cost and provide the quantitative threshold for “whether calibration experiments are actually worth doing.”
First, let’s examine the structure of the cost. Both approaches estimate the same quantity—the standard error of the population mean—and the variance scaling factor is identical on both sides: the sample-count approach uses the mean per sample, whose variance automatically carries the design effect (1+(r−1)ρ)/r; the calibrated-discount approach uses all readings plus design effect correction, and the variance scaling factor is the same formula. Thus, the “waste” of the conservative approach does not lie in the variance scaling factor, but solely in one aspect: degrees of freedom. For the sample-count approach, df = n−1; for the calibrated-discount approach, df = n·r/(1+(r−1)ρ) − 1. Since the t-distribution quantiles differ, the width ratio is almost exactly the ratio of these two t-quantiles. This is a conclusion that can be calculated by hand; the simulation merely confirms it.
The simulation grid covers r = 2/3/5/10, ρ = 0.05/0.1/0.2/0.3, n = 5 to 50, with 200,000 repetitions per group, plus an additional layer of “coverage leveling” for fair comparison—scaling the intervals of the calibrated-discount approach uniformly to actually cover 0.95 before comparing widths, eliminating any apparent precision inflation. The results fall into three layers.
Layer 1: In the primary working area (n=30, r=3, ρ<0.3), the cost is only 2.0% to 2.2%; even when pulling repetitions up to 10 and correlation down to 0.05, it is only 3.0%; for n=20, it does not exceed 5%. The backlog’s preset “about 9%” clearly overestimates—the 9% figure actually represents the inflation rate of the design effect relative to the erroneous nominal approach of treating repetitions as independent observations, not the cost of the conservative approach relative to the calibrated-discount approach; these two were conflated into a single line.
Layer 2: The cost is almost independent of ρ and depends primarily on n. Under the same (n, r), the width ratio difference between ρ=0.05 and ρ=0.3 is less than 1 percentage point. The reason remains structural: ρ enters the variance scaling factor of both approaches in the same way, canceling each other out, leaving only its weak pull on the discounted degrees of freedom. In other words, the intuition that “high repetitions with low correlation incur the highest cost” is only half right—high repetitions do amplify the cost (r=10 costs about 0.8 percentage points more than r=3), but low correlation itself hardly raises the price. What truly drives the cost up is too few samples: at n=10, the cost jumps to 8%–11%; at n=7, it ranges from 12% to 22%; at n=5, it reaches as high as 21%–37%.
Layer 3: Regions where the cost exceeds 25% do exist, but only in extreme small-sample cases where n ≤ 5 (when r≥5, n=5 crosses the line; when r=3, n=4 crosses the line). In this region, the cheapest solution is not calibrating the discount factor, but adding samples: increasing n from 5 to 7 immediately drops the cost from over 25% down to 12%–20%; the benefit of adding two samples far exceeds any single calibration experiment. Moreover, note that the above comparison assumes ρ is a known constant after calibration—in real calibration experiments, ρ itself carries estimation error, so the actual gain of the calibrated-discount approach is further discounted, making the relative advantage of the conservative approach even greater.
Turning to project initiation decisions, the backlog’s two criteria have clear answers. In the common working area (r=3, ρ<0.3, n≥20), the measured cost is 2%–5%, far below the 10% threshold—”the sample-count starting rule is applicable across all scenarios” holds true, and calibration experiments formally downgrade to research-level work, not setting project priority. The region where cost exceeds 25% exists but is extremely narrow (n≤5), and the correct solution there is to first supplement samples to more than 7 rather than calibrate. Combining these two points: calibration experiments drop from “best to have” to “research-level,” and all new studies should start with the sample-count approach, calculating sample size based on n (number of samples) rather than n×r (number of samples multiplied by repetitions per sample).
The reliability sample size planning field indeed has mature precision frameworks (PMID 24338600, 2014), and there is direct evidence from 2025 on the substantive impact of sample size on retest reliability estimation (PMID 40119099); these two methods can be turned around to answer “how many samples are enough”: replacing n in the simulation with recruitment cost yields each project’s own price list.
This conclusion has an even more solid foundation. Post #1559 in this series already proved that the coverage rate of the sample-count approach does not depend on ρ; this round further proves that its width cost is also almost independent of ρ—these two properties combined mean this set of rules is completely immune to the unknown parameter “what is the within-session correlation for the judge.” In a research field where even the range of ρ can only be guessed (the pressure pain threshold literature cited in #1559 reported within-session repetition correlation of 0.73, PMID 33767329), having a ruler naturally immune to unknown parameters is itself the source of the most valuable asset. The only remaining gap is: the sample-count approach compresses the contribution of judgment repetitions into the mean per sample, with the cost manifesting as t-distribution quantiles—this gap is a mathematical structure, not an engineering defect; filling it can only be done by adding samples, which is also the consistent prescription given throughout this entire series.
Evidence and Limitations
Verified: Monte Carlo (200,000 per group, with coverage leveling for fair comparison) yields full-grid costs: n≥20 all below 10%, primary working area 2%–5%; cost increases monotonically with n and remains approximately constant with ρ; regions exceeding 25% exist only for n≤5. Analytical layer confirms width ratio ≈ ratio of t-quantiles, variance scaling factors identical on both sides.
Partially Verified: The calibrated-discount approach assumes ρ is already calibrated and constant; real calibration carries estimation error, so actual discounted gain will be even smaller, making the directional conclusion model-independent.
Pending Verification: Measured within-session correlation values for Qigong external air judgment scenarios have not yet been obtained (same as #1559/#1564); if future measured ρ shows large differences in discounting between judges, the “discount varies with reader experience” branch still requires separate adjudication.
Related Reading
- How much does within-session correlation eat into effective degrees of freedom?
- Can the discount factor transfer across judgment configurations?
- Research Pool 1: Resonance Consciousness Field
L1 Network Status: This topic (cost of sample-count vs. calibrated-discount approaches) has no corresponding edge in CORRESPONDENCE_NETWORK / EDGE_TABLE; search for “discount/effective degrees of freedom/sample-count/precision cost/calibration” yields zero hits. Nearest neighbors are E-060 label premise edge (conflict, discussing low interpersonal consistency, PMID 30383398) and KB-510 (deep breathing HRV resting ICC 0.50–0.83, discussing reading stability), both concerning “reading reliability” rather than this topic’s within-session structural proposition. This topic has not yet entered the L1 corresponding network; evidence level capped at C.
Evidence Level: C
References: PMID 24338600, PMID 40119099, PMID 33767329
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