Why Star Ratings Can Use Broader Data Than Rankings | 星级评级为什么能使用比排名更广的数据

📚 Why Star Ratings Can Use Broader Data Than Rankings | 星级评级为什么能使用比排名更广的数据

A ranking and a star rating look like two ways of doing the same job: telling you which school, university or hospital is “better”. In practice they are mathematically different objects. A ranking demands a strict order, so every piece of evidence must ultimately be squeezed into a single comparable number. A star rating only demands a label, so evidence can stay multi-dimensional, incomplete, uncertain and even non-comparable. That single structural difference explains why star systems can absorb far more data than league tables without collapsing into nonsense.

排名和星级看起来都在做同一件事：告诉你哪所学校、哪所大学或哪家医院”更好”。但在数学上，它们是两类不同的对象。排名要求一个严格的顺序，因此所有证据最终都必须被压缩成一个可比较的数字；星级只要求一个标签，因此证据可以保持多维度、不完整、带有不确定性，甚至彼此不可比。正是这一结构性差异，解释了为什么星级体系能吸纳比排行榜多得多的数据，而不会自我崩塌。

1. Rankings Force Everything onto One Axis | 排名把所有信息压缩到一根轴上

A ranking is the output of a sorting operation. You build a composite score for each institution, then sort. Formally, for institution i and indicator j:

Sᵢ = Σⱼ wⱼ · xᵢⱼ → Rankᵢ = position of Sᵢ in the sorted list

Every indicator xᵢⱼ must be on a scale that can be multiplied by a weight wⱼ and added. Percentages, z-scores, ratios – all must be reduced to the same unit. Anything that resists that reduction gets dropped.

排名是一次排序操作的输出。先为每个机构计算综合得分，再排序。对机构 i 和指标 j：

Sᵢ = Σⱼ wⱼ · xᵢⱼ → Rankᵢ = Sᵢ 在排序列表中的位置

每个指标 xᵢⱼ 都必须落在一个能乘以权重 wⱼ 再相加的量表上。百分比、z 分数、比率——全都要被换算成同一单位。任何抗拒这种换算的信息，都会被剔除。

Qualitative evidence such as inspection narratives cannot be added to a numeric score, so it is normally excluded.

诸如督导报告叙述这样的质性证据无法与数值得分相加，因此通常被排除在外。
Indicators measured on different scales need standardisation, which destroys their original meaning.

不同量表的指标需要标准化，而标准化会破坏其原本含义。
Breadth of data therefore becomes a liability, not an asset: more indicators mean more arbitrary conversion decisions.

因此数据越广反而成为负担而非优势：指标越多，武断的换算决定就越多。

2. Star Ratings Are Labels, Not Orderings | 星级评级是标签，不是顺序

A star rating is a classification function. It maps a profile of evidence to one of a small number of categories:

f(x₁, x₂, …, xₙ) → {★, ★★, ★★★, ★★★★}

Because the output set has only four or five elements, the function is allowed to be non-linear, conditional and threshold-based. It does not have to produce a total order.

星级评级是一个分类函数。它把一组证据映射到少数几个类别之一：

f(x₁, x₂, …, xₙ) → {★, ★★, ★★★, ★★★★}

由于输出集合只有四五个元素，这个函数可以是非线性的、带条件的、基于阈值的。它不必产生一个全序。

Ranking: strict total order, no ties allowed in principle, differences between adjacent positions are meaningful.

排名：严格全序，原则上不允许并列，相邻名次之间的差异被视作有意义。
Star rating: ordinal categories with many ties, differences within a star band are explicitly declared meaningless.

星级：有序类别，允许大量并列，同一星级内部的差异被明确宣布为无意义。
Ties are the key. They create “room” for extra data that never has to be ranked against anything.

并列是关键。并列腾出了”空间”，让额外数据无需与任何东西比较排序。

3. Why Ties Make Breadth Affordable | 为什么并列让”广数据”变得可负担

Suppose two institutions differ on a new indicator. In a ranking, one must move above the other, which shifts every position below them. In a star system, both can simply remain ★★★. The new data is used to check that the star is still justified, but it does not force a distinction that the evidence cannot support.

假设两个机构在一个新指标上有差异。在排名中，其中一个必须排到另一个前面，所有后续名次随之移动；在星级体系中，两者都可以只是保留 ★★★。新数据被用来检验这个星级是否仍然成立，但不会强迫做出证据无法支撑的区分。

Rank difference = Δposition (always non-zero if scores differ) | Star difference = 0 or ±1 (usually 0)

This is why star systems can safely include indicators such as student wellbeing surveys, destination data, inclusion metrics or safeguarding compliance, which are valuable but noisy, hard to scale, or only available for some institutions.

这就是为什么星级体系可以放心纳入诸如学生幸福感调查、升学去向数据、包容性指标或安全保障合规等指标——它们很有价值，但噪声大、难以量化，或只对部分机构可得。

4. Threshold Logic vs Relative Logic | 阈值逻辑与相对逻辑

Rankings are inherently relative: your position depends on everyone else. Star ratings are usually absolute: a star is awarded if a standard is met.

排名本质上是相对的：你的位置取决于其他人。星级评级通常是绝对的：只要达到标准就授予星级。

Feature	Ranking	Star Rating
Comparison type	Relative (norm-referenced)	Absolute (criterion-referenced)
Stability over time	Low: one new entrant shifts everyone	High: standards can stay fixed
Data requirement	Complete, comparable, single-scale	Partial, multi-source, mixed-scale
Tolerance for noise	Very low	Manageable with safeguards

Because a criterion-referenced judgement does not depend on the distribution, institutions far apart in raw performance can hold the same star. That frees the rating body to collect indicators that are informative about quality but would be unfair as competitive ranking inputs.

由于标准参照的判断不依赖于分布，原始表现相差很远的机构也可以持有同一星级。这就解放了评级机构，使其能够收集那些对质量有信息量、但作为竞争性排名输入会不公平的指标。

5. Non-Comparable Indicators Can Coexist | 不可比指标可以共存

In a ranking, apples and oranges must be converted into the same currency. In a star system, they can be treated as separate conditions joined by logical operators rather than by addition:

★★★ ⟺ (outcome ≥ θ₁) AND (progress ≥ θ₂) AND (compliance = pass)

A conditional rule never requires an exchange rate between outcomes and compliance. That is the mathematical freedom rankings do not have.

在排名中，苹果和橘子必须被兑换成同一种货币；在星级体系中，它们可以作为彼此独立的条件，用逻辑运算符而非加法连接：

★★★ ⟺ (成果指标 ≥ θ₁) 且 (进步指标 ≥ θ₂) 且 (合规 = 通过)

条件式规则从不需要在成果与合规之间设定汇率。这正是排名所不具备的数学自由。

Binary indicators (pass/fail, present/absent) work naturally as gate conditions.

二值指标（通过/未通过、有/无）可以自然充当门槛条件。
Ordinal survey scales can be mapped to bands without pretending the gaps between points are equal.

有序量表调查可以直接映射到分档，无需假装各点之间的距离是相等的。
Narrative evidence can act as a modifier that changes a star only in defined circumstances.

叙述性证据可以作为修正项，只在明确规定的情形下改变星级。

6. Handling Missing Data and Small Cohorts | 处理缺失数据与小样本

Rankings punish missing data brutally. An institution with no data on one indicator either receives a zero, a national average, or is excluded entirely – all three are distortions.

排名对缺失数据的惩罚极为严酷。某个指标没有数据的机构，要么被记零分，要么被赋予全国平均值，要么被整体排除——三种做法都是扭曲。

A star system can do something more honest: withhold the star, publish an “insufficient evidence” label, or award a provisional rating with a note. Because stars are categories rather than positions, “unrated” is a legitimate output.

星级体系可以采取更诚实的做法：暂不授予星级、发布”证据不足”标签，或标注为临时评级。由于星级是类别而非位置，”未评级”本身就是一个合法输出。

Output set = {★, ★★, ★★★, ★★★★, Insufficient evidence}

7. Confidence Intervals and Star Bands | 置信区间与星级分档

Statistical uncertainty is handled differently by the two systems. A ranking hides uncertainty: two institutions separated by 0.3 points look ranked as if the difference were real. A star rating can build uncertainty into the rule.

两个体系处理统计不确定性的方式不同。排名会隐藏不确定性：相差 0.3 分的两个机构被排出名次，仿佛差异真实存在。星级体系可以把不确定性写进规则本身。

Suppose an institution’s progress score is +0.15 with a 95% confidence interval of [-0.10, +0.40], and a star threshold requires progress ≥ 0.00. Because the interval crosses the threshold, the honest decision is to treat the evidence as inconclusive for that criterion, not to rank the institution below one scoring +0.20.

假设某机构的进步分数为 +0.15，95% 置信区间为 [-0.10, +0.40]，而某星级要求进步分数 ≥ 0.00。由于该区间跨越阈值，诚实的做法是判定该项证据尚无定论，而不是把这所机构排到得分 +0.20 的机构之后。

Point estimate: the single value used in a ranking.

点估计：排名所使用的单一数值。
Interval estimate: the range a star system can act on, producing “provisional” or “confirmed” labels.

区间估计：星级体系可据此行动的区间，从而产生”临时”或”确认”标签。
Overlapping intervals mean the difference is not statistically distinguishable at the chosen confidence level.

区间重叠意味着在所选置信水平下，差异在统计上无法区分。

8. Compensation vs Non-Compensation | 补偿性与非补偿性

A weighted sum is a compensatory rule: a very high score on one indicator can offset a very low score on another. Rankings are almost always compensatory, because addition is commutative and forgiving.

加权求和是一种补偿性规则：某一指标的极高分可以抵消另一项指标的极低分。排名几乎总是补偿性的，因为加法可交换且宽容。

Star systems can choose non-compensatory logic, which is why they can absorb more indicators safely:

Compensatory: Sᵢ = Σⱼ wⱼ xᵢⱼ | Non-compensatory: Star = min over j of g(xᵢⱼ)

The minimum function means no amount of excellence in one area rescues a failure in another. Safeguarding, financial probity or minimum teaching hours are naturally expressed this way – and none of them fit comfortably into a ranking’s arithmetic.

星级体系可以选择非补偿性逻辑，这正是它能安全吸纳更多指标的原因：

补偿性：Sᵢ = Σⱼ wⱼ xᵢⱼ | 非补偿性：星级 = 对各 j 取 g(xᵢⱼ) 的最小值

取最小值的含义是：某一领域的卓越无论多高，都无法挽救另一领域的失败。安全保障、财务诚信或最低教学时数天然适合这样表达——而它们都不容易塞进排名的算术里。

9. Where Star Ratings Fail | 星级评级的局限

Broader data does not mean better decisions automatically. Star systems have their own failure modes, and a strong answer must acknowledge them.

数据更广并不自动意味着决策更好。星级体系有其自身的失效模式，一个优秀的回答必须承认这一点。

Ceiling and floor effects: if 70% of institutions hold ★★★, the rating stops discriminating among them.

天花板与地板效应：如果 70% 的机构都是 ★★★，该评级在它们之间就失去了区分能力。
Arbitrary thresholds: moving a cut-off from 0.00 to 0.05 can reclassify hundreds of institutions overnight.

任意阈值：把分界线从 0.00 移到 0.05，一夜之间可能让数百所机构改变类别。
Gaming: any published criterion invites optimisation towards the criterion rather than the goal (Goodhart’s law).

博弈：任何公开的标准都会诱发”以达标为目标”而非以目标本身为目标的优化（古德哈特定律）。
Loss of fine resolution: star ratings deliberately discard information that a rank order would preserve.

分辨率损失：星级评级有意舍弃了排名所保留的细粒度信息。

In short, stars trade resolution for breadth. Rankings trade breadth for resolution.

总之，星级用分辨率换取广度；排名用广度换取分辨率。

10. Worked Example: Two Schools, Two Systems | 实例演算：两所学校，两套体系

Consider two schools with the following profiles. The indicator weights in a hypothetical ranking are 50% attainment, 40% progress, 10% destinations.

设想两所学校的数据如下。假设某排名体系的权重为：成绩 50%、进步 40%、升学去向 10%。

Indicator	School A	School B
Attainment score	+0.45	+0.10
Progress score	-0.05	+0.40
Destinations (%)	88	84
Safeguarding	Pass	Pass
Wellbeing survey	Lower quartile	Top quartile

A composite ranking based only on the first three indicators would place School A above School B, and it would have to discard the wellbeing column entirely because there is no defensible weight for it. A star system with thresholds (attainment ≥ 0.00, progress ≥ 0.00, compliance = pass, wellbeing not in bottom quartile) would award ★★ to both, while registering that School B looks stronger on progress and wellbeing and School A stronger on attainment.

若只依据前三个指标做综合排名，学校 A 会排在 B 之前，而幸福感这一列必须被完全丢弃，因为无法为其设定有说服力的权重。若改用阈值式星级（成绩 ≥ 0.00、进步 ≥ 0.00、合规 = 通过、幸福感不在最低四分位），两校都得到 ★★，同时记录下 B 在进步与幸福感上更强、A 在成绩上更强。

Ranking: A > B (wellbeing discarded) | Stars: A = B = ★★ (wellbeing retained)

11. Reading Star Ratings Critically | 批判性阅读星级评级

Exam questions and real-world reports both reward the same habits. Ask what the rating could include that a ranking could not, and what it chose to hide.

考试题目与现实报告所奖励的习惯是一致的。要问：这个评级能包含哪些排名无法包含的东西？它又隐瞒了什么？

Check the output set: how many categories are there, and what proportion of institutions sits in each?

检查输出集合：有几个类别？各机构落在各类别的比例是多少？
Check the rule type: compensatory (weighted sum) or non-compensatory (minimum, gate conditions)?

检查规则类型：补偿性（加权求和）还是非补偿性（取最小值、门槛条件）？
Check the thresholds: are they evidence-based, negotiated, or chosen for a target distribution?

检查阈值：是有证据依据的、协商产生的，还是为达到某种目标分布而选定的？
Check uncertainty treatment: are confidence intervals published, and can a star be provisional?

检查不确定性处理：是否公布置信区间？星级是否可以是临时性的？
Check missing data policy: is “insufficient evidence” available, or is missing data silently imputed?

检查缺失数据政策：是否提供”证据不足”选项？还是缺失数据被悄悄插补？

These five questions turn a vague discussion of “which measure is fairer” into a precise statistical argument, which is exactly what high-scoring answers do.

这五个问题能把”哪种度量更公平”的模糊讨论，转变为精确的统计论证——这正是高分答案所做的。

12. Key Takeaways | 核心要点总结

The central insight is structural, not moral. Rankings need a single comparable number because they produce a strict order; star ratings need only a category, so they can accept multi-dimensional, mixed-scale, incomplete and uncertain evidence.

核心洞见是结构性的，而非道德性的。排名必须得到一个可比较的单一数字，因为它要产生严格顺序；星级只需要一个类别，因此可以接受多维、混合量表、不完整且带有不确定性的证据。

Strict total order (ranking) vs ordinal categories with ties (stars).

严格全序（排名）对比允许并列的有序类别（星级）。
Relative thresholds (ranking) vs absolute thresholds (stars).

相对阈值（排名）对比绝对阈值（星级）。
Compensatory addition (ranking) vs non-compensatory minimum and gate conditions (stars).

补偿性加法（排名）对比非补偿性取最小值与门槛条件（星级）。
Implicit uncertainty (ranking) vs published confidence intervals and provisional stars (stars).

隐含不确定性（排名）对比公布置信区间与临时星级（星级）。
The cost is resolution: stars deliberately give up fine-grained distinctions in exchange for breadth.

代价是分辨率：星级有意放弃细粒度区分，以换取广度。

Remember the one-line answer for a timed question: a ranking is a sorting problem that requires full commensurability; a star system is a classification problem that requires only thresholds and logical conditions – and thresholds can be applied separately to each dimension.

记住限时答题的一句话答案：排名是一个要求全面可公度的排序问题；星级是一个只需阈值与逻辑条件的分类问题——而阈值可以对每个维度分别施加。

Published by TutorHao | Mathematics & Statistics Revision Series | aleveler.com

更多咨询请联系16621398022（同微信）

Why Star Ratings Can Use Broader Data Than Rankings | 星级评级为什么能使用比排名更广的数据