Evidence level: C
「练得越久 Δ 越窄」这句话里藏着两次减法。前一次是 Δ = 练功态 − 静息态,后一次是 Δ 对年限回归。把第一次减法做对,救不了第二次。
一、两个减法,前一篇只拆了第一个
两态差值 Δ 的测量误差,此前已经单独审过一次:两个成分各自一致,不代表相减之后还一致,MDC95 那条闸门就是为此设的。
但那只是第一道。第二道减法从来没被拆开过:把 Δ 对修炼年限做回归,看斜率是否显著为负。这一步引入两个全新的误差源——场次之间的随机抖动,以及年限分组本身的混杂。前者与”测得准不准”无关(单个 Δ 可以很准而组间差全是噪声),后者与样本量无关(招再多人也消不掉)。一个设计可以在第一道减法上完全合格,却在第二道减法上从一开始就测错了对象。
二、场次噪声恰好和几十年年龄差一个量级
这是整件事最要命的一个数字。一项针对成人知觉—运动表现的重复测量研究给出:场次之间的变异,平均相当于个体间变异的 25%–50%;更关键的是后半句——这个量级与横断面比较中跨越 12 至 27 岁年龄范围所显示的差异相当(PMID 15014087)。
把它换成本题的语言:如果你按”新手 / 1–5 年 / 5–10 年 / 10 年以上”分层,那么同一个人隔天再测一次的抖动,就足以伪造出一个相当于”多练了十几年”的组间差。
这不是”噪声有点大需要多招人”,而是效应量与噪声底重叠。在这种情况下,任何按年限分层的横断面结果,都无法与”场次噪声恰好这么排”相区分。
三、斜率的方差有两个来源,不是一个
方法学上的正确骨架早在三十多年前就写清楚了:个体斜率(对时间回归,取最小二乘估计)的方差包含两个成分——一个来自个体内变异(由测量误差、随访长度与测量频率共同决定),另一个是真实的个体间斜率差异(PMID 2362980)。
这条公式对本题是致命的。随访期越短、次数越少,个体内那一项就越大。而一个”四层年限、每人只测一次”的横断面设计,本质上把个体内成分推到了最大,真实斜率差异那一项被埋在下面。
所以功效计算不能拿”总方差”当分母——必须把两项分开报。只报总方差的设计,等于承认自己不知道算的是哪一项。
四、用横断面替代纵向,是这道题最贵的捷径
为什么大家爱用横断面?因为便宜,但便宜有代价,代价可以量化。同一批影像遗传学数据上,纵向分析与仅用基线的横断面分析相比,在全基因组显著性水平上检出的关联更多(PMID 25098835)。反向推论就是:横断面要达到同等功效,需要的样本更多。而按变化率比较的样本量公式还多要一个参数——暴露的组内相关;时变暴露的研究在比较变化率时普遍不如时不变暴露高效(PMID 19899065)。本题的”修炼年限”恰恰是最典型的时变暴露。所以四层横断面设计的功效问题不是”够不够”,而是分子从一开始就错了。
五、短期和长期斜率,可以指向相反方向
即便改成纵向,还有一层:短期变化与长期变化不是同一个东西。
三层多层模型把二者同时参数化后,出现了拆分才看得见的分裂——简单选择反应时的反应时不一致度短期平稳、跨年上升;而 one-back 任务的不一致度短期显著下降、长期无变化(PMID 35721020)。同一个指标,两个时间尺度的斜率方向相反,合成一个斜率就等于互相抵消。
对应到本题:Δ 随年限的”变化”里,如果混着”第一次测的新鲜感”和”练了十年之后的结构改变”,这两股力的方向可能不同。不先分开参数化,得到的合成斜率没有解释力。
六、训练效应会把趋势读反,不只是读小
更坏的情形是符号错。
一项 4718 人的纵向研究里,认知得分在第一年逐步改善,此后才开始恶化;调整训练效应后,组间差异才达到显著(p = 0.019),而”整体衰退”的表观趋势消失(PMID 11901278)。也就是说,不调整时看到的趋势和调整后的趋势,可以完全不是一回事。
而训练效应与真实变化在常规设计里几乎不可分:等间隔随访中,时间与暴露次数完美共线,二者只有在特定收敛假设下才能分离;假设不成立时,偏倚与一类错误率同时失控(PMID 21639642)。
对本题,这条曲线的形状可以预期:新手头几次 Δ 快速上升(学会了怎么鼓肚子),老手早已平台。这条曲线若被读成年限效应,符号正好是反的。
七、年限与年代,最后一刀切不开
最底层的问题不在测量,在识别。
年龄、时期、世代三者满足”世代 = 时期 − 年龄”的线性依赖,线性成分因此无法被分别估计;各种约束方案仍在争论,模拟显示没有任何一种模型能通吃(PMID 40779571)。随机效应方案也不占便宜——它施加的约束是任意的、且高度隐晦,与其他受约束估计量并无本质差别(PMID 36644415)。
落到本题:一个六十岁、练了三十年的人,他的 Δ 窄,可能因为练得久,也可能因为他出生在那一代人里、那批人本来就不同。四层年限设计对此完全无解。
证据与局限
已验证: 场次间变异可达个体间变异的 25%–50% 且等量于跨 12–27 岁的横断面差(PMID 15014087);斜率方差由个体内与个体间两成分构成(PMID 2362980);纵向分析在同批数据上功效优于横断面(PMID 25098835);短期与长期斜率可反向(PMID 35721020);训练效应可反转趋势读法(PMID 11901278);APC 线性成分不可分离(PMID 40779571;PMID 36644415)。这些均为方法学层面的定论级证据。
部分验证: 上述数字全部来自认知与知觉—运动领域,没有一项是在胸腹相位角或腹区占比上做的。跨领域外推的方向可信(方差成分结构是普适的),但量级不能直接搬——φ 的个体内变异系数在不同人群间跨 11%–123%,本身离散极大。
待验证: 三个关键量在本主题上完全空白——① Δ 的个体内与个体间方差各占多少(决定年限斜率是否存在可测空间);② φ/%AB 上的 retest gain 有多大、几次之后饱和;③ 修炼年限与出生年代在当前修炼人群中的共线强度。这三项不测出来,本文的判据只能算”条件性成立”。
L1 网络状态:该主题尚未进入 L1 对应网络。
CORRESPONDENCE_NETWORK.md节点表按「易学/中医侧」「物理侧」「生理/体验侧」三类域分派,逐词检索「信度 / 重测 / 最小可检测 / MDC / 功效 / 效度 / 胸腹 / 年限 / 年资 / ICC / 样本量」全部 0 命中;边表止于 E-050(「六十甲子纳音 | 对应 | 60 个频率」,层级 L1、状态 open,现代证据栏原文「尚无任何可测来源」),无任何以方差成分、统计功效或年限斜率为源或目标的边。EDGE_TABLE_science-yixue.md仅有概念计数(丹田 51 篇 / 气功 48 篇 / 呼吸 189 篇),无 L1/L2/L3 层级与 active/open/conflict 判定。
相关阅读
- 静息与练功之间那道缝,是越练越窄还是越练越宽:两态差值的年限斜率裁决
- 「两态差值」自己测得准不准,没人验过:Δφ 的 MDC95 与等效界的可行性闸门
- 「老本」到底存不存在:修炼年限分层的停练保留差,可能只是百分比的算术
- 丹田呼吸与频率科学(研究池子)
证据等级: C
参考文献: PMID 15014087, PMID 2362980, PMID 25098835, PMID 19899065, PMID 35721020, PMID 11901278, PMID 21639642, PMID 40779571, PMID 36644415
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。
Session Noise Can Forge Decades of Proficiency: The Two Variance Components of Slope Rates and the Cross-Sectional Longitudinal Efficacy Trap
Evidence level: C
The phrase “the longer one practices, the narrower Δ becomes” conceals two subtractions. The first is Δ = practice state − resting state; the second is the regression of Δ against tenure. Getting the first subtraction right does not save the second.
I. Two Subtractions; the Previous Article Only Decomposed the First
The measurement error of the two-state difference Δ was already scrutinized separately: each component is consistent, yet their difference need not be, and the MDC95 threshold was set precisely for this reason.
That is only the first cut. The second subtraction has never been decomposed: regressing Δ against years of practice to test whether the slope is significantly negative. This step introduces two entirely new error sources—random jitter between sessions and confounding inherent in tenure grouping. The former is unrelated to “how accurately we measure” (a single Δ can be precise while group differences are pure noise); the latter is unrelated to sample size (no matter how many recruits, confounding cannot be eliminated). A design can be fully adequate for the first subtraction yet measure the wrong object from the start for the second.
II. Session Noise Matches the Magnitude of Decades of Age Difference
This is the most critical number in the whole matter. A repeated-measures study of adult perceptual–motor performance reports: between-session variability averages 25%–50% of between-individual variability; more crucially, this magnitude matches the differences observed across a 12–27 year age range in cross-sectional comparisons (PMID 15014087).
Translating to this topic’s language: if you stratify by “novice / 1–5 years / 5–10 years / 10+ years,” then the jitter from retesting the same person the next day is sufficient to fabricate a group difference equivalent to “practicing for an extra decade or more.”
This is not “noise is too large; we need more recruits.” It is effect size overlapping the noise floor. Under these conditions, any cross-sectional result stratified by tenure cannot be distinguished from “session noise happening to align exactly this way.”
III. The Slope Variance Has Two Sources, Not One
The methodological skeleton was laid out over thirty years ago: the variance of individual slopes (regressed on time, estimated via least squares) comprises two components—one from within-individual variation (determined by measurement error, follow-up duration, and measurement frequency), and the other from true between-individual slope differences (PMID 2362980).
This formula is fatal for this topic. Shorter follow-up and fewer measurements inflate the within-individual term, burying the true between-individual slope differences.
Therefore, power calculations cannot use “total variance” as the denominator—the two terms must be reported separately. Designs reporting only total variance admit ignorance of which component they are actually estimating.
IV. Replacing Longitudinal with Cross-Sectional Is the Most Expensive Shortcut
Why do people favor cross-sectional designs? Because they are cheap, but cheapness has a quantifiable cost. On the same imaging-genetics dataset, longitudinal analysis detects more associations at genome-wide significance than cross-sectional analysis using baseline only (PMID 25098835). In reverse inference: cross-sectional designs require larger samples to achieve equivalent power. Moreover, sample size formulas based on rate-of-change comparisons demand an additional parameter—the intra-group correlation of the exposure; studies with time-varying exposures are generally less efficient than those with time-invariant exposures (PMID 19899065). “Years of practice” is precisely the quintessential time-varying exposure. Thus, the power problem of a four-tier cross-sectional design is not “is it enough?” but the numerator is wrong from the start.
V. Short-Term and Long-Term Slopes Can Point in Opposite Directions
Even if switched to longitudinal, another layer remains: short-term change and long-term change are not the same thing.
Three-level multilevel modeling, which parameterizes both simultaneously, reveals a split visible only upon decomposition: simple choice reaction time shows short-term stability with year-over-year increases; whereas the one-back task shows short-term significant decline with no long-term change (PMID 35721020). For the same metric, slopes over two time scales point in opposite directions; combining them into a single slope results in mutual cancellation.
Mapped to this topic: within the “change” of Δ over tenure, if it mixes “freshness from the first measurement” and “structural change after ten years of practice,” these forces may oppose each other. Without separate parameterization, the composite slope lacks interpretability.
VI. Training Effects Can Reverse the Trend, Not Just Diminish It
An even worse scenario involves sign reversal.
In a longitudinal study of 4,718 participants, cognitive scores improved gradually in the first year before beginning to decline; after adjusting for training effects, between-group differences reached significance (p = 0.019), while the apparent trend of “overall decline” vanished (PMID 11901278). That is, the unadjusted trend and the adjusted trend can be entirely different.
In conventional designs, training effects and true change are nearly inseparable: in equidistant follow-ups, time and exposure frequency are perfectly collinear, separable only under specific convergence assumptions; when those fail, both bias and Type I error rates run wild (PMID 21639642).
For this topic, the expected curve shape is: Δ rises rapidly for novices (learning how to expand the abdomen), while experts are already plateaued. If this curve is read as a tenure effect, the sign is exactly reversed.
VII. Tenure and Calendar Year: The Final Cut Cannot Be Made
The deepest issue lies not in measurement but in identification.
Age, period, and cohort satisfy the linear dependency “cohort = period − age”; thus linear components cannot be separately estimated; various constrained approaches remain debated, with simulations showing no model universally applicable (PMID 40779571). Random effects approaches offer no advantage—their constraints are arbitrary and highly implicit, differing in essence from other constrained estimators (PMID 36644415).
Back to this topic: a 60-year-old who has practiced for thirty years may show narrow Δ either because of long practice or because they were born into a cohort with inherently different characteristics. A four-tier tenure design is completely unsolvable for this.
Evidence and Limitations
Verified: Between-session variability can reach 25%–50% of between-individual variability and equals cross-sectional differences across ages 12–27 (PMID 15014087); slope variance comprises within-individual and between-individual components (PMID 2362980); longitudinal analysis outperforms cross-sectional on the same dataset (PMID 25098835); short-term and long-term slopes can oppose each other (PMID 35721020); training effects can reverse trend interpretation (PMID 11901278); APC linear components are inseparable (PMID 40779571; PMID 36644415). These are methodological conclusions.
Partially Verified: All cited numbers come from cognitive and perceptual–motor domains; none address thoraco-abdominal phase angle or abdominal proportion. Cross-domain extrapolation is credible (variance component structure is universal), but magnitude cannot be directly transferred—φ’s within-individual coefficient of variation spans 11%–123% across populations, showing extreme heterogeneity.
To Be Verified: Three key quantities are entirely blank for this topic: ① the split between within-individual and between-individual variance of Δ (determining whether a measurable tenure slope exists); ② the retest gain on φ/%AB and after how many repetitions saturation occurs; ③ the collinearity strength between years of practice and birth cohort among current practitioners. Without these, this article’s criteria can only be deemed “conditionally valid.”
L1 Network Status: This topic has not yet entered the L1 correspondence network. The
CORRESPONDENCE_NETWORK.mdnode table, distributed across “Yijing Studies/TCM side,” “physics side,” and “physiological/experiential side,” yields zero hits for all terms: “reliability / retest / minimum detectable / MDC / power / validity / thoraco-abdominal / tenure / seniority / ICC / sample size.” The edge table stops at E-050 (“Sixty Jiazi Na Yin | Correspondence | 60 Frequencies,” level L1, status open; modern evidence column reads “no measurable source yet“), with no edges sourced or targeted by variance components, statistical power, or tenure slope.EDGE_TABLE_science-yixue.mdcontains only concept counts (Dantian 51 articles / Qigong 48 articles / Breathing 189 articles), lacking L1/L2/L3 levels and active/open/conflict judgments.
Further Reading
- The Gap Between Resting and Practicing: Is It Narrowening or Widening with More Practice? Verdict on the Tenure Slope of Two-State Differences
- “How Accurate Is the Two-State Difference Itself? No One Has Verified It: Feasibility Thresholds for MDC95 of Δφ and Equivalent Boundaries”
- “Does ‘Old Capital’ Actually Exist? The Stopped-Practice Retention Difference Across Tenure Strata May Be Mere Percentage Arithmetic”
- Dantian Breathing and Frequency Science (Research Pool)
Evidence Level: C
References: PMID 15014087, PMID 2362980, PMID 25098835, PMID 19899065, PMID 35721020, PMID 11901278, PMID 21639642, PMID 40779571, PMID 36644415
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