同一笔预算投给例还是投给人:n-k 平面上的等精度线是什么形状,决定「加人」与「加例」在哪个价位交叉

Written by

in

Language

Evidence level: C

判读研究的预算只有两条腿——请人的钱和找例的钱;等精度线在「例数×人数」平面上长什么样,决定了钱该往哪条腿上放,而这个配比没有一个放之四海皆准的答案,它随判读者一致性 ρ 和人力单价同时变。

一、上篇留下的对偶问题

上一篇(判读者边际递减)算清了「多请一个判读者还值不值」,但那是在例数固定的前提下问的。真实场景里预算是死的:多请一个判读者的钱,等于少找几个例的钱。这两笔投入可以互相兑换吗?兑换价是多少?这就是预算分配的对偶侧。概化理论早有现成的数学工具——在预算约束下用拉格朗日乘数法解最优样本量,2024 年有系统工作把它推广到多侧面设计(PMID 39418281)。问题是:这套解析解放到真实判读场景里,和全网格暴力仿真的实测值到底差多少?差距决定了解析公式能不能直接当设计初值用。

二、先验货:解析式全网格过关

用与上篇同款的可加性模型(真值加判读者噪声),把 ρ=0.4 和 0.6、例数 n=30 到 240、人数 k=2 到 12 共 40 个格子全部仿真,实测半宽与解析方差公式逐格比对。结果:40 格的实测与解析比值全部落在 0.97 到 1.05 之间,最大偏差 5%。这比预想的「偏差小于 10% 才可当初值」门槛还要好——结论是:在这个模型族里,解析折算公式不是近似品,可以直接当设计初值,全网格仿真只需要做终检。

三、等精度线的形状:低 k 端陡、高 k 端平

固定 n=60 看等精度线怎么走:人数从 2 加到 4,半宽降 36%(ρ=0.4);4 到 6 再降 9 个点;到 12 人时总共只降了 53%。ρ=0.6 时更平,加满到 12 人只降 42%。这条线的形状是典型的双曲线式:低人数端陡峭(前几个判读者最值钱),高人数端几乎水平(再多请人在精度上白花钱)。这就是「加人」和「加例」交叉点的几何来源——人便宜时先堆人吃陡段,人贵过某个价位后倒过来全投例。

这也解释了为什么只看一张固定 n 的信度表会做错预算:表上 k=12 比 k=2 的半宽只好看一半,容易得出「判读者没什么用」的结论;但把省下的例数预算换算回来就知道,陡段那 4 个人(2 到 6 人)贡献了全部降幅的八成多,后面 6 个人只是在平段上摊薄。等精度线不是一条直线,沿它走,每一步的「汇率」都在变。

四、最优配比表:随人力单价和 ρ 双变

固定总预算 120 单位(找一例花 1 单位),解四个价位档的最优配比:人力单价等于例价时,最优是约 20 个判读者、100 例;人力贵 3 倍,降到约 10 人、89 例;贵 10 倍,约 5 人、70 例;贵 30 倍,只剩 2 到 3 人、48 例。更反直觉的是 ρ 的方向:判读者越一致(ρ=0.6),各价位的最优人数反而越少——一致性 0.4 时平价档要 19.6 人,0.6 时只要 15.7 人。道理和上篇一脉相承:越靠谱的团队越不用堆人,因为他们的边际递减来得更早。

拿人力贵 3 倍这一档做个具体演算:如果守着老习惯只请 2 个判读者、把例堆到 114 例,半宽约 0.093;挪预算到 10 人 90 例,半宽压到 0.051,等于不花一分钱把精度提高了将近一半。但继续往人这边挪就亏了——20 人 60 例时半宽回升到 0.057。最优配比是内点不是端点,两头走极端都要付精度税。

五、配比表要不要按 ρ 分档

分档判据的答案是分价位:平价档(人力不贵)两个 ρ 的最优人数差 4 人,超过 2 人的分档门槛,必须按 ρ 分档;人力贵 3 倍时差 1.9 人,压线;贵 10 倍以上差距不足 1 人,单表通用即可。另外,因为解析式全网格过关,预想中「解析解在低例数高人数角系统性乐观」的翻车区并没有出现——守则可以放心把折算公式当初值。落到判读线的条款上就是三句话:报方案先报人力单价对例价的比值;比值 10 以上直接用 4 到 5 人的通用小配比;比值低时必须先估 ρ 再定人数,别抄别人的配比表。

证据与局限

已验证: 概化理论在预算约束下拉格朗日最优样本量方法成熟(PMID 39418281);有效受试者数与判读者数互换公式有直接文献支撑(PMID 16143966);按精度反解双方向样本量方法成熟(PMID 24338600)。本文 40 格仿真验证解析公式偏差不超过 5%,属模型内确证。

部分验证: 最优配比随人力单价剧变(20 人到 2 人)与随 ρ 反向移动(一致性越高人数越少),在可加性模型加平均配对相关估计量下成立,但只覆盖完全交叉设计。

待验证: 轮换、缺席、专业分工导致的非均衡设计下最优配比怎么移(承接上篇待验证);单次 ICC 估计自带不确定度会把最优配比往哪个方向推(PMID 34477455 已证不确定度会放大所需样本量,但没算配比侧);概化系数置信区间双约束设计下配比是否仍由精度单目标决定(PMID 35754612)。

相关阅读

证据等级: C
参考文献: PMID 39418281, PMID 16143966, PMID 24338600, PMID 35754612, PMID 34477455

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。N200 研究内核边表(EDGE_TABLE_science-yixue.md,63 行)grep「预算/配比/等精度/判读者/评分者」0 命中;对应网络中仅脾容积「测量误差预算」(E-004/KB-488)等生理测量条目,与本文的「判读设计预算分配」不是同一条边。故本文按最弱档 C 处理,全部数字为模型内仿真结论,非已验证机制。


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。

Same Budget for Examples or People: What Shape Do Iso-cost Lines Take on the n-k Plane, and at What Price Do \”Adding People\” and \”Adding Examples\” Intersect?

Evidence level: C

Allocating a Fixed Budget to Cases or Persons: What Shape Do Equal-Precision Lines Take on the n-k Plane, and Where Do “Adding Cases” and “Adding Persons” Intersect?

In Yijing Studies, the precision of judgment rests on two legs—funds for recruiting participants and funds for collecting cases. The shape of equal-precision lines on the “number of cases × number of persons” plane determines where to place each dollar, and this allocation ratio has no universal answer; it varies with the consistency of the judge (ρ) and the unit cost of labor.

I. Dual Problems Left by the Upper Part

The preceding section (on diminishing marginal returns for readers) calculated whether “adding one more reader is still worth it,” but this was under the constraint of a fixed number of cases. In real-world scenarios, the budget is fixed: the cost of hiring one additional reader equals the opportunity cost of not finding a few more cases. Can these two expenditures be converted into one another? What is the conversion rate? This constitutes the dual side of budget allocation. Generalized theory already possesses ready-made mathematical tools—using the Lagrange multiplier method to solve for optimal sample size under budget constraints. In 2024, systematic work extended this to multi-aspect design (PMID 39418281). The question remains: when this analytical solution is applied to real-world reading scenarios, how does it differ from the empirically measured values of full-grid brute-force simulation? This gap determines whether the analytical formula can be directly used as the initial design value.

II. Prior Load: Analytical Full-Grid Pass

Using the same additive model as the previous section (true values plus reader noise), we simulated all 40 cells across ρ=0.4 and 0.6, sample sizes n=30 to 240, and participant counts k=2 to 12. Measured half-widths were compared against the analytical variance formula cell-by-cell. Results: all 40 simulated-to-analytical ratios fell between 0.97 and 1.05, with a maximum deviation of 5%. This exceeds the expected threshold of “deviation under 10% to serve as initial values.” Conclusion: within this model family, the analytical conversion formulas are not approximations but can serve directly as design initial values; full-grid simulation is required only for final verification.

III. Shape of the Equal-Precision Line: Steep at Low k, Flat at High k

Fixed n=60 to trace the equal-precision line: increasing raters from 2 to 4 reduces half-width by 36% (ρ=0.4); from 4 to 6 reduces it by another 9 points; reaching 12 raters yields a total reduction of only 53%. At ρ=0.6 the line is flatter, and filling to 12 raters reduces half-width by just 42%. This curve exhibits a characteristic hyperbolic shape: the low-rater end is steep (the initial few raters contribute most to precision), while the high-rater end is nearly flat (adding more raters beyond this point yields diminishing returns in precision). This geometry arises from the intersection of “adding raters” and “adding cases”—when raters are inexpensive, one initially accumulates them along the steep segment; once rater cost exceeds a certain threshold, resources shift entirely to cases.

This also explains why relying solely on a fixed-n reliability table can lead to budget misallocation: the table shows k=12 offers only half the half-width reduction of k=2, tempting the conclusion that “raters are useless”; yet converting the saved case budget reveals that those first 4 raters (2 to 6) account for over 80% of the total reduction, while the subsequent 6 raters merely dilute precision along the flat segment. The equal-precision line is not a straight line; traversing it, the “exchange rate” of each step continuously changes.

IV. Optimal Allocation Table: Co-varying with Labor Unit Price and ρ

Fixed total budget of 120 units (based on one example costing 1 unit), the optimal allocation across four price tiers is calculated as follows: when labor unit price equals example price, the optimum is approximately 20 judges and 100 examples; when labor is 3 times more expensive, it drops to approximately 10 people and 89 examples; when 10 times more expensive, about 5 people and 70 examples; when 30 times more expensive, only 2 to 3 people and 48 examples. Counterintuitively, the direction of ρ matters: the more consistent the judges (ρ=0.6), the fewer people required at each price tier—consistency at 0.4 requires 19.6 people for the standard tier, while consistency at 0.6 requires only 15.7. This aligns with the previous section: more reliable teams require fewer personnel because their marginal diminishing returns occur earlier.

Taking the 3-times-more-expensive labor tier as a concrete calculation: adhering to old habits with only 2 judges and piling up 114 examples yields a half-width of approximately 0.093; shifting the budget to 10 people and 90 examples compresses the half-width to 0.051, effectively improving precision by nearly half without spending additional money. However, continuing to shift toward more people becomes counterproductive—20 people and 60 examples results in the half-width rising to 0.057. The optimal allocation lies at an interior point, not at the extremes; pushing to either extreme incurs a precision tax.

V. Should the Allocation Table Be Stratified by ρ?

The criterion for stratification is the cost tier:
– 平价档 (low-cost tier): When labor costs are not high, the optimal difference in personnel count between two ρ values is 4 people. This exceeds the stratification threshold of 2 people; therefore, stratification by ρ is required.
– 人力贵 3 倍 (labor costs 3 times higher): The difference is 1.9 people, which falls exactly on the threshold.
– 人力贵 10 倍以上 (labor costs more than 10 times higher): The difference is less than 1 person; a single unified table suffices.

Furthermore, because the analytical solution covers the full grid, the anticipated failure zone—where “the analytical solution would be systematically optimistic for low sample sizes and high personnel counts”—did not appear. Consequently, the rules allow the conversion formula to be used directly as the initial value.

The clauses for the interpretation lines can be summarized in three sentences:
1. When reporting a proposal, first report the ratio of labor unit price to sample price.
2. If this ratio exceeds 10, directly apply the general small allocation of 4 to 5 people.
3. If the ratio is low, estimate ρ first before determining the number of people; do not copy others’ allocation tables.

Evidence and Limitations

Verified: The generalized theory yields a mature Lagrange-optimal sample size method under budget constraints (PMID 39418281); direct literature supports the interchange formula between effective subject count and rater count (PMID 16143966); and the precision-based inversion of bidirectional sample sizes is established (PMID 24338600). This paper’s 40-grid simulation confirms the analytical formula’s deviation remains within 5%, validating it internally.

Partially Verified: The optimal ratio shifts dramatically with personnel unit cost (from 20 to 2) and moves inversely with ρ (higher consistency requires fewer people), holding true under additive models with average paired correlation estimators, but only covering fully crossed designs.

Pending Verification: How the optimal ratio shifts under unbalanced designs caused by rotation, absenteeism, or specialization (continuing from the previous paper’s pending status); in which direction the inherent uncertainty of single ICC estimates pushes the optimal ratio (PMID 34477455 has proven uncertainty amplifies required sample size but did not address the ratio side); and whether the ratio under dual-constrained confidence interval designs remains determined solely by precision (PMID 35754612).

Related Reading

Evidence level: C
References: PMID 39418281, PMID 16143966, PMID 24338600, PMID 35754612, PMID 34477455

Evidence level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration

L1 network status: This topic has not yet entered the corresponding L1 network. The N200 research core edge table (EDGE_TABLE_science-yixue.md, 63 lines) yields zero hits for “budget/ratio/precision/rater/scorer”; the corresponding network contains only physiological measurement items such as “spleen volume measurement error budget” (E-004/KB-488), which is not the same edge as this article’s “reading design budget allocation.” Therefore, this article is treated at the weakest level C; all numbers are simulation results within the model, not verified mechanisms.


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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *