Evidence level: C
一条筛选规则的合格证,不是它省了多少样本,而是换一批数据后它选中的还是不是那批参数。
一、”漏检率”这个指标,缺一个填不进去的数
想算漏检率,得先知道真效应有多大——可真效应正是这场研究要测的东西。拿观测到的效应当基准算漏检率,等于用答案批改答案。能算的只有两件:一是”可检出差异”,即给定样本量与重复次数,这个参数最小能分辨多大的差;二是条件功效,即按现有数据继续做下去,最终能出显著结果的概率。前者是设备与设计的属性,可以事前算;后者才是筛选规则真正动到的量,而且它随数据滚动,不是常数。任何声称给出”漏检率百分比”的筛选规则验证,第一步就该交代它用的是哪一个。
二、阈值卡在估计值上,选出来的集合必然偏高
信噪比是估出来的,不是量出来的。用同一批数据既估信噪比、又按信噪比决定谁入选,就是典型的双重 dipping:选择与检验不独立,描述统计与统计推断同时失真。神经影像里早就把这个坑讲透了——被阈值筛过、又是局部极大的那些峰值,其幅度天然带选择偏倚,必须专门做重采样校正才能还原原始效应量,否则后续功效分析全部建立在虚高的地基上。同样的机理在流行病学里叫效应量放大:只报告”过了线”的结果,报告的效应就不是总体效应,而是条件效应。
三、”≥2″是个比值门槛,比值估计量没有稳定的方差
信噪比的分母是设备不一致带来的误差项。分母趋近于零时,比值爆炸;分母本身还随设备对、体位、标定状态漂移。这意味着”某参数信噪比 2.3″这句话,换一台机器、换一批受试者,完全可能变成 1.6。阈值一旦卡在这种抖动量上,入选集合就是随机集合:真值 2.5 的参数有一半概率被踢出去,真值 1.4 的参数有一半概率混进来。所以该报的第一个数是入选集合的稳定性——比如两次独立分半各自筛选,两个集合的交集占并集的比例。这个数低,后面所有功效讨论都是沙上建塔。
四、降级的代价有两种,被混成了一种
把参数降级为”仅单设备内比较”,损失的不是一件事。其一是功效损失:可用信息变少,可检出差异变大,这是统计问题,加样本、加重复次数能补。其二是范围损失:结论的适用范围从”这套丹田参数”缩成”这台设备上的这套参数”,这是外部效度问题,加多少样本都补不回来——样本再多,也只在那台机器上更精确。多中心研究里这两笔账还互相纠缠:中心数少、入组严重不平衡时,常规渐近近似本身就会失控,p 值与误差控制同时失真,有研究报告过标准近似漏掉真实风险因素、反而制造 II 类错误的实例。降级决策若不区分这两笔账,就会用”加样本”去补一个加样本补不了的洞。
五、阈值挪一格,结论就换一个
阈值 2 没有生理学出处,是约定。约定的东西必须做敏感性分析,否则结论的稳健性无从判断。行为科学 RCT 的多宇宙分析给过刺眼的数字:仅预处理流程的选择就能解释估计效应总方差的 76.9%,而模型规格只占 7.5%;某些标准化或对数变换流程把效应压掉九成以上,另一些保留原始尺度的流程则把效应放大一个数量级。切点选择的研究也是同一结论:挑一个让两组分得最开的”最优切点”,I 类错误严重膨胀、效应绝对值被系统性高估,连置信区间的覆盖率都要靠收缩加自助法才能救回来。筛选阈值与切点同构——它也是从数据里挑出来的最优分割。
六、反转:该先验的不是漏检率,是规则的自我一致性
原命题假设规则是稳的、只是有点保守,于是去量”保守造成多少漏检”。反转在于:这条规则可能根本不稳,它的问题不是漏掉多少,而是每一批数据漏掉的不是同一批。所以验证顺序要倒过来。第一,报入选集合的稳定性(分半 Jaccard、bootstrap 入选频率),低于某个水平就判定规则不可复现,功效问题不必再谈。第二,报双阈值区间(如 1.5 到 3)内结论是否翻转,翻转即说明结论由约定而非数据决定。第三,对未入选参数报”可检出差异”而非”漏检率”,让读者自己判断这个分辨率够不够用。第四,明确降级属于功效损失还是范围损失,前者写样本量方案,后者写适用边界。四件事做完,这条筛选规则才算有了合格证;在此之前,它省下的样本量只是把不确定性挪到了看不见的地方。
证据与局限
已验证: 选择与检验不独立会使描述统计与推断同时失真,在纯噪声数据上也能造出虚假效应(PMID 19396166);被阈值筛选后的峰值幅度带选择偏倚,需重采样校正,且校正值才能用于功效分析(PMID 31866164);只报告过线结果会造成效应量放大(PMID 34434078);最优切点选择导致 I 类错误膨胀与效应高估,需收缩加自助法恢复覆盖率(PMID 15160403);预处理流程选择可解释效应估计方差的 76.9%(PMID 41185384);多中心小中心数、入组失衡时渐近近似失效并可造成 II 类错误(PMID 41569702)。
部分验证: 因无效性提前终止的试验中存在效应误估,系统调查了 52 个此类试验的设计与条件功效,但未给出统一的漏检率估计(PMID 31948397);测量误差校正依赖信度系数,tau 等价假设被违反时校正后的均值相关有偏、覆盖率偏低(PMID 35604700);成人呼吸模式测量工具的证据分散,尚无单一模态达到”临床可用”(PMID 42862511);把多个指标聚合成复合指数会稀释单个指标的预测力(PMID 42645861)。
待验证: ①”信噪比≥2″阈值在丹田参数上的分半 Jaccard 稳定性从未被实测;②双阈值区间内结论翻转率未知;③降级造成的范围损失尚无量化指标;④比值型门槛在小分母参数上的方差行为未在呼吸力学参数上标定。
相关阅读
- 「X5/AX 设备偏倚最大」缺主语:排序是设备对的属性,不是参数的属性
- 跨设备”可互换率”不是参数的属性,是「方法对 × 值域」的函数
- 丹田参数与其修炼者-新手效应幅度的头对头量化
- 时空频率拓扑研究池
L1 网络状态:该主题尚未进入 L1 对应网络——
CORRESPONDENCE_NETWORK.md节点表仅含易学/中医、物理、生理体验三类域,无「信噪比 / 筛选阈值 / 入选集合稳定性 / 试验终点降级」等测量学维度节点;边表仅至 3.4 节(天干地支↔五行),无对应边。据「薄 / 仅文献 / 无 A 级证据」规则,本文证据等级封顶 C。证据等级: C
参考文献: PMID 19396166, PMID 31866164, PMID 34434078, PMID 15160403, PMID 41185384, PMID 41569702, PMID 31948397, PMID 35604700, PMID 42862511, PMID 42645861
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 Reproducibility of the SNR≥2 Filtering Rule Is More Prior Than Its Efficacy in Saving Parameters
Evidence level: C
A certification for a selection rule is not defined by how many samples it saves, but by whether it selects the same parameters when applied to a new batch of data.
I. The Metric of “False Negative Rate” Lacks a Fillable Number
To calculate the false negative rate, one must first know the magnitude of the true effect—which is precisely what this study aims to measure. Using the observed effect as a baseline to calculate the false negative rate is akin to grading an answer with the answer itself. Only two things can be calculated: first, the “detectable difference,” which is the minimum magnitude of difference a given sample size and repetition count can resolve for a specific parameter; second, conditional power, which is the probability of ultimately achieving statistical significance if the study continues with current data. The former is an attribute of equipment and design, calculable a priori; the latter is the quantity truly affected by the selection rule, and it fluctuates with data accumulation rather than being constant. Any validation of a selection rule claiming to provide a “false negative rate percentage” must first specify which definition it employs.
II. When the Threshold Is Fixed on an Estimate, the Selected Set Is Inherently Biased
The signal-to-noise ratio is estimated, not measured. Using the same dataset to both estimate the signal-to-noise ratio and determine inclusion based on that ratio constitutes classic double dipping: selection and testing are not independent, causing simultaneous distortion in descriptive statistics and statistical inference. This pitfall has been thoroughly elucidated in neuroimaging—peaks filtered by thresholds and identified as local maxima naturally carry selection bias in their amplitude; specific resampling corrections are required to restore the original effect size. Without this, all subsequent power analyses rest on a foundation of inflated values. The same mechanism in epidemiology is termed effect size inflation: reporting only results “above the line” means the reported effect size is not the population effect, but a conditional effect.
III. “≥2” Is a Ratio Threshold; Ratio Estimators Lack Stable Variance
The denominator of the signal-to-noise ratio represents error terms arising from equipment inconsistency. As the denominator approaches zero, the ratio explodes; furthermore, the denominator itself drifts with equipment pairs, positioning, and calibration states. This means the statement “a parameter has a signal-to-noise ratio of 2.3” may completely shift to 1.6 with a different machine or cohort. When the threshold is fixed on such an oscillating quantity, the selected set becomes a random set: parameters with a true value of 2.5 have a 50% probability of being excluded, while those with a true value of 1.4 have a 50% probability of being included. Therefore, the first number to report should be the stability of the selected set—for instance, the proportion of the intersection relative to the union when filtering independently via two half-splits. If this number is low, all subsequent discussions on power are building towers on sand.
IV. The Cost of Downgrading Comes in Two Forms, Often Conflated
Downgrading a parameter to “comparison within only a single device” incurs losses beyond a single event. First is power loss: available information decreases, detectable differences increase; this is a statistical issue that can be remedied by adding samples or repetitions. Second is scope loss: the applicability of conclusions shrinks from “this set of Dantian parameters” to “this set of parameters on this specific device.” This is an external validity issue that cannot be compensated for by adding samples—no matter how many are added, precision only increases on that specific machine. In multi-center studies, these two accounts become entangled: with few centers and severe imbalance in enrollment, standard asymptotic approximations themselves lose control, distorting both p-values and error rates. Instances have been reported where standard approximations miss true risk factors while simultaneously generating Type II errors. If the downgrade decision does not distinguish between these two accounts, one may attempt to fill a hole that cannot be filled by adding samples.
V. Shifting the Threshold by One Unit Changes the Conclusion
The threshold of 2 lacks physiological origin; it is a convention. Conventions require sensitivity analysis; otherwise, the robustness of conclusions cannot be judged. Multi-universe analyses in behavioral science RCTs have provided stark figures: selection of preprocessing procedures alone accounts for 76.9% of the total variance in estimated effects, whereas model specification accounts for only 7.5%. Certain standardization or logarithmic transformation procedures reduce effects by over ninety percent, while others retaining raw scales amplify them by an order of magnitude. Research on cutoff selection yields the same conclusion: selecting an “optimal cutoff” that maximally separates two groups causes severe inflation of Type I errors and systematic overestimation of absolute effect sizes, even requiring contraction with bootstrapping to restore confidence interval coverage. Selection thresholds are isomorphic to cutoffs—they are also optimal splits selected from the data.
VI. Reversal: The Prior Should Not Be the False Negative Rate, but the Rule’s Self-Consistency
The original proposition assumes the rule is stable but merely conservative, thus measuring “how many false negatives are caused by conservatism.” The reversal lies in the possibility that the rule is fundamentally unstable; its problem is not how many are missed, but that different batches miss different parameters. Therefore, the validation sequence must be inverted. First, report the stability of the selected set (half-split Jaccard index, bootstrap inclusion frequency); if below a certain level, the rule is deemed irreproducible, rendering further discussion of power moot. Second, report whether conclusions flip within a dual-threshold interval (e.g., 1.5 to 3); flipping indicates conclusions are determined by convention rather than data. Third, for non-selected parameters, report “detectable difference” rather than “false negative rate,” allowing readers to judge if this resolution suffices. Fourth, explicitly distinguish whether downgrading constitutes power loss or scope loss; the former warrants a sample size plan, while the latter defines applicability boundaries. Once these four tasks are completed, the selection rule possesses a certificate of authenticity; prior to this, the samples saved merely shift uncertainty into the invisible.
Evidence and Limitations
Verified: Selection and testing not being independent causes simultaneous distortion in descriptive statistics and inference, even creating spurious effects on pure noise data (PMID 19396166); peaks filtered by thresholds carry selection bias in amplitude requiring resampling correction, and only corrected values can be used for power analysis (PMID 31866164); reporting only above-line results causes effect size inflation (PMID 34434078); optimal cutoff selection leads to Type I error inflation and overestimation of effects, requiring contraction with bootstrapping to restore coverage (PMID 15160403); selection of preprocessing procedures accounts for 76.9% of variance in effect estimates (PMID 41185384); in multi-center studies with few centers and enrollment imbalance, asymptotic approximations fail and can cause Type II errors (PMID 41569702).
Partially Verified: Trials terminated early due to lack of efficacy show effect misestimation; a systematic investigation of design and conditional power in 52 such trials did not provide a unified false negative rate estimate (PMID 31948397); measurement error correction relies on reliability coefficients, and when the tau equivalence assumption is violated, corrected means show bias and lower coverage (PMID 35604700); evidence for adult respiratory pattern measurement tools is dispersed, with no single modality reaching “clinically usable” status (PMID 42862511); aggregating multiple indicators into composite indices dilutes the predictive power of individual indicators (PMID 42645861).
To Be Verified: ① The half-split Jaccard stability of the “signal-to-noise ratio ≥ 2” threshold on Dantian parameters has never been empirically measured; ② The flip rate of conclusions within dual-threshold intervals is unknown; ③ Quantitative metrics for scope loss caused by downgrading are lacking; ④ Variance behavior of ratio-type thresholds on small denominators has not been calibrated for respiratory mechanics parameters.
Related Reading
- 「X5/AX Device Bias Maximum」Missing Subject: Ranking is an Attribute of Equipment Pairs, Not Parameters
- Cross-Device “Interchangeability Rate” Is Not an Attribute of Parameters, but a Function of «Method Pair × Value Domain»
- Head-to-Head Quantification of Dantian Parameters vs. Effect Amplitude of Practitioners-Novices
- Spacetime Frequency Topology Research Pool
L1 Network Status: This topic has not yet entered the L1 corresponding network—the
CORRESPONDENCE_NETWORK.mdnode table contains only domains related to Yijing/TCM, Physics, and Physiological Experience, lacking measurement dimensions such as “Signal-to-Noise Ratio / Selection Threshold / Selected Set Stability / Trial Endpoint Downgrading.” The edge table only extends to Section 3.4 (Heavenly Stems and Earthly Branches ↔ Five Phases), with no corresponding edges. According to the “Thin / Only Literature / No Level A Evidence” rule, this article’s evidence level is capped at C.Evidence Level: C
References: PMID 19396166, PMID 31866164, PMID 34434078, PMID 15160403, PMID 41185384, PMID 41569702, PMID 31948397, PMID 35604700, PMID 42862511, PMID 42645861
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