Evidence level: C
先粗扫定位、再在候选区加密,这个补救方案默认了「粗估的点足够准」——而变点估计偏偏在最需要加密的边界处最不可靠,漏窗时加密采到的全是同侧数据,结果比均匀铺点更差。
正文
测量「功夫上身」的时间,本质是在估计一条随时间变化的曲线上的拐点:练功多少个月,胸腹相位角才开始出现可检测的位移。上一轮(#1184)我们已经知道,这个拐点的最小可检测位移不是常数,而是拐点位置的函数——拐点落在随访区间两端时,检测能力系统性变差。面对这个困境,一个看起来很聪明的补救方案是两阶段设计:先用几个稀疏的点粗扫一遍,大概定位拐点在哪个区间,再回到那个区间加密采样,把点数花在刀刃上。
这个方案的底气来自一条统计学的经典结论:对本质非线性的函数,在曲率最大处加密采样,参数恢复最好(PMID 26609742)。问题在于,拐点恰恰是我们要估的东西——曲率最大处在哪,事先并不知道。于是两阶段设计必须依赖一个隐含前提:粗估的点足够准,加密窗能圈住真拐点。
这个前提恰恰是最不可靠的。变点估计有一个已知的病理:当真拐点靠近边界、或样本量小时,点估计与区间估计会一起失效(PMID 40585262)。也就是说,粗扫阶段最容易出错的地方,正是边界附近——而边界附近又恰是最需要加密的那一段。两件事叠在一起,构成了一个「越补救越糟」的结构:
如果粗估把真拐点漏在了加密窗外,那么加密采样采到的全部是真拐点同一侧的数据。对这样一个单侧数据堆做拟合,拐点会被牢牢钉在加密窗的边界上——不是因为它真的在那,而是因为窗外没有任何数据能把它拉回来。此时两阶段设计的误差,会比同样点数均匀铺开还要大。
这不是理论上的杞人忧天。测量频率与随访设计的研究早就提醒过:增加测量点数带来的收益是递减的(PMID 33644405),所以两阶段设计必须胜过「同样点数的均匀铺点」才有意义——否则你就是用更复杂的方案、花同样的钱、拿到更差的结果。而自适应设计的统计文献也反复强调一个核心教训:设计规则里嵌进的估计(比如用中期数据选剂量、选亚组),会把自己的误差带进最终估计(PMID 26459506);变点模型驱动的自适应设计在药理学里早已被检验过,它的可操作特性(operating characteristics)必须靠仿真逐场景标定(PMID 27221807),没有任何理由相信「粗估 → 加密」在气功纵向数据上会自动成立。
那么怎么裁决?这是一个纯仿真就能回答的问题,不需要额外招募一个人。用 #1184 同一套实测噪声底,把三种设计并排摆开:
- 现行 5 点均匀(0/1/3/6/12 月);
- 均匀 9 点(与两阶段总点数对齐);
- 两阶段:粗扫 3 点等距,按粗估拐点 ±1 个间隔设加密窗、窗内再取 4 点。
然后让真拐点位置在 1-12 月之间扫描、幅度在 1-10 度之间扫描,各重复 ≥500 次,记录三件事:拐点区间半宽、点估计偏差、以及漏窗率(真拐点落在加密窗之外的比例)。
判据分三档:
- 若漏窗率 >20%,且 6-12 月段区间半宽反而比均匀 9 点更差 → 「加密」这条补救路线在该设计下作废,改均匀铺点或连续记录;
- 若漏窗率 <5%,且半宽优于均匀 9 点 → 两阶段可用,但加密窗必须按粗估的区间上界宽度(而不是点估计)来定,给粗估的误差留出余量;
- 若两阶段只在拐点落于窗内(2-8 月)时优于均匀、在两端劣于均匀 → 写成条件处方:先做均匀粗扫判定「是否值得进两阶段」,而不是一开始就跳进两阶段。
这一判的核心不是「两阶段好不好」,而是「补救方案的有效性本身是有条件的」——任何把粗扫当作免费午餐的设计,都必须先证明粗扫的误差不会传染给加密阶段。对研究「练功多久见功」这类问题,这个裁决直接决定后续所有随访协议:是花两倍点数赌一个可能漏窗的两阶段,还是老老实实均匀铺点。
证据与局限
已验证: 变点估计在真拐点靠近边界或样本量小时点估计与区间估计共同失效(PMID 40585262);在曲率最大处加密采样对本质非线性函数参数恢复最好(PMID 26609742);测量频率与随访时长对非线性趋势项估计的改善存在收益递减(PMID 33644405);自适应设计中的选择规则会把估计误差带进最终估计(PMID 26459506);变点模型自适应设计的操作特性须仿真逐场景标定(PMID 27221807)。
部分验证: 两阶段(先粗扫后加密)设计在剂量反应试验中已有损失函数框架支撑(PMID 12036265),但其收益依赖中期估计的准确性,该依赖在纵向随访场景下的表现尚无直接证据。
待验证: ① 以 #1184 同一套 φ 噪声底代入,两阶段设计的漏窗率在 6-12 月段是否真的 >20%;② 加密窗按「区间上界宽度」而非点估计来定,能否把漏窗率压到 5% 以下;③ 两阶段是否只在拐点落于窗内时优于均匀铺点(条件处方的成立范围)。
相关阅读
- 一个数字骗人:最小可检测拐点位移不是常数,是拐点位置的函数
- 「稳而无用」能不能靠换形态救回来:φ-SD 与变异时间结构指标头对头
- Δ 的方差里有多少是隔天再测抖出来的:方差成分拆分
- 研究池子:生物频率密码
证据等级: C
参考文献: PMID 26609742, PMID 40585262, PMID 33644405, PMID 26459506, PMID 27221807, PMID 12036265L1 网络状态: 该主题(变点估计的两阶段采样设计是否受粗扫误差污染)尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK 节点表生理侧仅列「呼吸率 / 呼吸模式 | 可穿戴 | 与节律耦合最直接」,无「随访采样设计 / 变点估计 / 加密窗 / 粗扫误差」类测量学节点或边;EDGE_TABLE_science-yixue.md 反抽概念表无「变点 / 采样设计 / 加密窗」条目。该主题已登记至 science-yixue-l1-gaps.md。
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。
The more you patch, the worse it gets: Two-stage encryption windows perform worse than uniform point placement when coarse scanning is misaligned.
Evidence level: C
The proposed remedy of a coarse scan followed by encryption assumes that “the rough estimate is accurate enough.” Yet, change-point estimation is precisely least reliable at the boundaries where encryption is most needed. When the window leaks, the encrypted sample captures data only from one side of the true turning point. Fitting such unilateral data forces the estimated turning point to the window boundary—not because it is there, but because no data outside the window can pull it back. Consequently, this approach yields worse results than uniform spacing.
Body
Estimating the time when “cultivation takes effect” fundamentally involves estimating the inflection point on a curve that changes over time: after how many months does the thoraco-abdominal phase angle begin to show a detectable shift. As established in the previous round (#1184), the minimum detectable displacement at this inflection point is not constant; it is a function of the inflection point’s location itself—detection capability systematically degrades when the inflection point lies near the boundaries of the follow-up interval.
Faced with this dilemma, a seemingly clever remedy is the two-stage design: first, use a few sparse points for a rough scan to roughly locate which interval the inflection point falls into, then return to that interval for dense sampling, spending the points where they matter most.
The confidence in this approach stems from a classic statistical conclusion: for inherently nonlinear functions, dense sampling at the point of maximum curvature yields the best parameter recovery (PMID 26609742). The problem is that the inflection point is precisely what we are trying to estimate—the location of maximum curvature is unknown beforehand. Thus, the two-stage design must rely on an implicit assumption: the rough estimate is accurate enough that the encryption window will capture the true inflection point.
This assumption is precisely the most unreliable part. Change-point estimation has a known pathology: when the true inflection point is near the boundary or the sample size is small, both the point estimate and the interval estimate fail simultaneously (PMID 40585262). In other words, the coarse scan stage is most prone to error exactly at the boundaries—the very region that needs encryption the most. These two factors combine to create a “the more you try to fix it, the worse it gets” structure:
If the rough estimate misses the true inflection point outside the encryption window, then all data collected during the dense sampling phase comes from only one side of the true turning point. Fitting such unilateral data forces the estimated inflection point to the edge of the encryption window—not because it is truly there, but because no data outside the window can pull it back. Under these conditions, the error of the two-stage design exceeds that of uniform spacing with the same number of points.
This is not theoretical worry. Research on measurement frequency and follow-up design has long warned that the benefit of increasing the number of measurements diminishes (PMID 33644405). Therefore, the two-stage design must outperform “uniform spacing with the same number of points” to be meaningful—otherwise, you are using a more complex scheme, spending the same money, and getting worse results. Statistical literature on adaptive design also repeatedly emphasizes a core lesson: estimation rules embedded in the design (e.g., selecting dose or subgroups using interim data) carry their own error into the final estimate (PMID 26459506); adaptive designs driven by change-point models in pharmacology have long been tested, and their operating characteristics must be calibrated via simulation across scenarios (PMID 27221807). There is no reason to believe that “rough estimate → encryption” would automatically hold for Qigong longitudinal data.
So how do we decide? This is a question answerable entirely by simulation, requiring no additional recruitment of a single person. Using the same measured noise floor from #1184, we compare three designs side by side:
- Current 5-point uniform (0/1/3/6/12 months);
- Uniform 9-point (aligned with the total number of points in the two-stage design);
- Two-stage: coarse scan with 3 equally spaced points, setting the encryption window at ±1 interval around the rough estimate of the inflection point, then taking 4 additional points within the window.
We then scan the true inflection point location across 1–12 months and its amplitude across 1–10 degrees, repeating each scenario ≥500 times, recording three metrics: half-width of the inflection point interval, bias of the point estimate, and window leak rate (the proportion of true inflection points falling outside the encryption window).
The criteria are tiered:
- If the window leak rate >20%, and the half-width of the 6–12 month interval is worse than uniform 9-point → the “encryption” remedy route is invalid under this design; switch to uniform spacing or continuous recording;
- If the window leak rate <5%, and the half-width is better than uniform 9-point → the two-stage design is viable, but the encryption window must be set based on the interval upper-bound width (not the point estimate) of the rough estimate, leaving room for the error of the rough scan;
- If the two-stage design is superior only when the inflection point falls inside the window (2–8 months), but worse at both ends → formulate a conditional prescription: first perform uniform coarse scanning to determine “whether it is worth entering the two-stage design,” rather than jumping straight into the two-stage design from the start.
The core of this judgment is not “is the two-stage design good,” but rather “the validity of the remedy itself is conditional”—any design that treats the coarse scan as a free lunch must first prove that the error of the coarse scan does not propagate into the encryption stage. For research questions like “how long until cultivation takes effect,” this verdict directly determines all subsequent follow-up protocols: whether to gamble twice the number of points on a two-stage design that might leak the window, or to stick with uniform spacing.
Evidence and Limitations
Verified: Change-point estimation fails simultaneously for both point and interval estimates when the true inflection point is near the boundary or sample size is small (PMID 40585262); dense sampling at maximum curvature optimally recovers parameters of inherently nonlinear functions (PMID 26609742); increasing measurement frequency yields diminishing returns for estimating nonlinear trend terms (PMID 33644405); selection rules in adaptive designs embed their own estimation error into the final estimate (PMID 26459506); operating characteristics of change-point model-driven adaptive designs must be calibrated via scenario-specific simulation (PMID 27221807).
Partially Verified: The two-stage (coarse scan then encryption) design has a loss function framework supporting it in dose-response trials (PMID 12036265), but its benefit depends on the accuracy of interim estimates, for which there is no direct evidence in longitudinal follow-up scenarios.
To Be Verified: ① With the same φ noise floor from #1184, does the two-stage design’s window leak rate truly exceed 20% in the 6–12 month interval? ② If the encryption window is set by “interval upper-bound width” rather than point estimate, can the leak rate be suppressed below 5%? ③ Does the two-stage design outperform uniform spacing only when the inflection point falls within the window (the range of validity for the conditional prescription)?
Further Reading
- A Number Deceives: Minimum Detectable Inflection Point Displacement Is Not Constant, but a Function of the Inflection Point Location
- Can “Steady but Useless” Be Saved by Changing Form: φ-SD vs. Variation Time Structure Endpoint Head-to-Head
- How Much of the Variance of Δ Comes from Day-to-Day Fluctuations: Variance Component Decomposition
- Research Pool: Biological Frequency Codes
Evidence Level: C
References: PMID 26609742, PMID 40585262, PMID 33644405, PMID 26459506, PMID 27221807, PMID 12036265L1 Network Status: This topic (whether the two-stage sampling design for change-point estimation is contaminated by coarse scan error) has not yet entered the L1 corresponding network—the CORRESPONDENCE_NETWORK node table on the physiological side lists only “breathing rate / breathing pattern | wearable | most directly coupled with rhythm,” lacking measurement nodes or edges such as “follow-up sampling design / change-point estimation / encryption window / coarse scan error.” The EDGE_TABLE_science-yixue.md concept extraction table also lacks entries for “inflection point / sampling design / encryption window.” This topic has been registered in science-yixue-l1-gaps.md.
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