Series ID: EDKS-ADV-VOC-ZH-0039 · Advanced English Vocabulary (Chinese Edition) · Lesson No.039 · C1 → C2 · 简体中文辅助
The average tells you where the centre is. The distribution tells you what surrounds it.
平均值告诉你中心在哪里；分布告诉你中心周围到底长什么样。
Two groups can have the same mean and tell completely different stories. One can be tightly concentrated around the centre; another can be widely spread. One can be symmetric; another strongly skewed. One can form a single cluster; another can split into two peaks. If a learner reports only the average, the sentence may be numerically correct but structurally blind.
NIST characterises a distribution through location, spread and shape: where values are centred, how much they vary, and how that variation is arranged. Cambridge also uses distribution for the way people or things are spread out and for the statistical frequency pattern of variable values. This gives Lesson 039 a broad but precise job: describe arrangement rather than centre alone.
Centre answers “where?” Spread answers “how far apart?” Shape answers “in what pattern?”
centre 问“在哪里”；spread 问“离得多远”；shape 问“以什么形状排列”。
This lesson owns the Mandarin-supported C1–C2 lexical route for distribution, spread, dispersion, variability, concentration, clustering, range, interquartile range, variance, standard deviation, skewness, symmetry, tails, outliers and multimodality. Lesson No.038 owns centre/typical values; Lesson No.036 owns change over time; Lesson No.034 owns exceptions/outliers in general; Lesson No.040 will own frequency and prevalence.
Part I — Build the distribution map | 第一部分：建立“分布地图”
1. Distribution is the broad arrangement concept | distribution = 值如何排列
In statistics, a distribution describes how values or frequencies are arranged across possible values. In geography/social science, distribution can describe how people/resources are spread across places. The shared idea is arrangement across a space—numerical, categorical or physical.
2. Distribution is not only a graph | distribution ≠ chart
A histogram, box plot or density plot can display a distribution, but the distribution is the underlying arrangement of observations. Saying “the distribution is skewed” describes the data pattern, not the visual design itself.
3. Location, spread and shape form the core triad | location–spread–shape
NIST summarises distributions with location, spread and shape. Location asks where values sit centrally; spread asks how widely they vary; shape asks whether the arrangement is symmetric, skewed, multimodal or otherwise structured. Lesson 038 owns location/centre; Lesson 039 focuses on spread and shape while coordinating the whole triad.
4. Spread is the plain-language width concept | spread
Spread describes how far apart data values are. OpenStax uses spread and variability for the degree to which observations lie close to or far from the mean. It is accessible general language and a statistical concept at the same time.
5. Variability is the broad difference/change concept | variability
Variability describes how much values differ across observations or measurements. Unlike Lesson 036’s temporal volatility, variability does not require time order. A classroom can show high variability in scores at one exam sitting.
6. Dispersion is a formal statistical spread word | dispersion
Dispersion is a formal noun for spreading/scattering and appears in statistical language for variability around a centre. It also has specialised physics/chemistry senses, so context matters. In general reports, spread may be clearer.
7. Concentration is the opposite intuition: values bunch together | concentration
Values are concentrated when many observations occupy a relatively narrow area or category. “Scores were concentrated around 70” says the distribution is tightly clustered near that level. Concentration can also have chemical and market senses; domain determines the relation.
8. Cluster means values/group members gather together | cluster
A cluster is a group of observations close together in feature/value/space. One distribution may contain several clusters. “Most scores clustered between 60 and 70” is descriptive; formal cluster analysis is a separate technical method.
9. Clustering is pattern; cluster analysis is method | avoid method inflation
Seeing visual clustering does not mean a formal clustering algorithm has identified groups. Say values appear clustered unless a defined method supports stronger technical language.
10. Range uses the extremes | range
The statistical range is the maximum minus minimum. It is simple and intuitive but uses only two extreme observations. NIST cautions that it tells little about spread near the centre. “Values ranged from 10 to 90” is descriptive; “range = 80” is the numerical spread measure.
11. “Range from X to Y” and “range of Z” differ | range grammar
“Scores ranged from 10 to 90.” “The score range was 80 points.” The first gives endpoints; the second gives width. Do not confuse interval endpoints with the calculated range.
12. Interquartile range focuses on the middle half | IQR
The interquartile range (IQR) spans the middle 50% of ordered observations between the first and third quartiles. It is less driven by extreme tails than full range. Technical learners should follow formal definitions; general learners can treat it as a spread measure focused on the central half.
13. Quartiles divide ordered data into four parts | quartiles
Quartiles are positional summaries. Q1, median/Q2 and Q3 help describe distribution beyond one centre. Exact calculation conventions can vary slightly across software/methods; preserve the source’s method if exact values matter.
14. Percentiles locate positions in a distribution | percentile
A percentile describes relative position: the 90th percentile is a threshold below which roughly 90% of observations fall under a stated convention. It is not “90% correct” or “top 90%”. Percentile language requires careful interpretation and will often accompany centre/spread summaries.
15. Variance is a technical squared-deviation measure | variance
Variance summarises squared deviations from the mean under a formal formula. Its units are squared relative to the original measurement, which makes direct interpretation less intuitive. Lesson 039 treats it as recognition/precision vocabulary, not a calculation lesson.
16. Standard deviation returns spread to original units | standard deviation
OpenStax describes standard deviation as a common measure of spread around the mean and notes that larger standard deviation means greater variation. Unlike variance, standard deviation is expressed in the same units as the data. Its usefulness depends on distribution shape and reader job.
17. Standard deviation is not “average distance” in casual wording | precision caution
People sometimes explain standard deviation loosely as the average distance from the mean. That is pedagogically intuitive but not the exact formula. In technical writing, call it a measure of spread based on squared deviations rather than inventing an inaccurate definition.
18. Small standard deviation means tighter concentration around mean | small SD
When standard deviation is small relative to the scale and context, values are more tightly concentrated around the mean. But “small” needs a reference: same units, same metric, or meaningful benchmark. A standard deviation of 10 can be tiny or huge depending on the variable.
19. Large standard deviation means wider spread, not poor quality automatically | large SD
High variability can be undesirable in manufacturing but natural in human populations. Do not translate spread into quality judgment without domain criteria. A wide distribution is a description first.
20. Shape asks how values occupy the spread | shape
Two distributions can have the same centre and standard deviation yet different shapes. Shape includes symmetry, skewness, tails, peaks and clusters. It prevents spread from becoming one-dimensional.
21. Symmetric means roughly balanced around a centre | symmetry
A symmetric distribution has broadly similar shape on both sides of a central location. This does not mean every pair is mirror-perfect unless the mathematical distribution is exactly symmetric. In empirical description, roughly/approximately symmetric is often safer.
22. Asymmetric is broad; skewed is statistical-directional | asymmetric vs skewed
Asymmetric simply means not symmetric. Skewed in statistics usually describes a distribution with a longer/more pronounced tail on one side. Use the more technical term only when that shape relation is intended.
23. Right-skewed means long tail toward higher values | right skew
A right-skewed distribution contains a longer upper/right tail. Mean is often pulled above median, though the relationship is not a universal shortcut for every dataset. Income and waiting-time examples commonly illustrate this shape.
24. Left-skewed means long tail toward lower values | left skew
A left-skewed distribution has a longer lower/left tail. Scores on an easy test can show this pattern when most students score high and a smaller group extends downward.
25. Tail is the thin/extreme end of a distribution | tail
Tail describes the low-frequency extreme region of a distribution. Upper tail contains high values; lower tail low values. Tail language is important because centre and middle spread can look healthy while tail outcomes remain severe.
26. Long-tailed means extremes extend farther | long tail
A long-tailed distribution has observations extending far from the centre. “Long tail” also has business/cultural uses, so statistical context matters. Technical heavy-tail definitions are more specific and should not be used casually.
27. Heavy-tailed is technical | heavy tails
Heavy-tailed has formal probability/statistical meanings regarding tail behaviour. General learners should recognise the term but avoid using it merely to mean “there are some outliers.”
28. Peak / mode describes high-density regions | peaks
A distribution can have one prominent peak (unimodal) or several (multimodal). A peak in a distribution is not the same as a time-series peak from Lesson 036. Here the horizontal axis represents values/categories, not time.
29. Unimodal means one prominent mode/peak | unimodal
Unimodal distributions have one main peak. The term is technical recognition vocabulary. A unimodal distribution can still be skewed or widely dispersed.
30. Bimodal / multimodal reveal multiple clusters | multiple modes
Two or more peaks can indicate mixed subpopulations, processes or categories. They do not automatically prove distinct causal groups, but they warn that one central value may be unrepresentative.
31. Gap is a low-density region between values/clusters | gap
A visible gap can separate clusters. Saying “there is a gap between 50 and 70” is descriptive. Inferring that two natural classes exist requires more evidence.
32. Outlier is an unusually distant observation | outlier
Outliers sit far from the main body under a rule or judgment. Lesson 034 owns exception terminology broadly; here the outlier’s job is distributional: it affects spread, range, mean and tail interpretation. Do not delete it simply because it is inconvenient.
33. Extreme value is descriptive; outlier can be analytical | extreme vs outlier
The maximum or minimum is an extreme value by position. Whether it is an outlier depends on a criterion or context. Using outlier can imply abnormality; extreme observation is sometimes more neutral.
34. Concentrated and dispersed form a useful contrast | concentration vs dispersion
“Scores were tightly concentrated around 70.” “Scores were widely dispersed across the full range.” These expressions give readers an intuitive picture without requiring formulas. Add exact measures when technical precision is needed.
35. Part I checkpoint: describe location, spread and shape separately | 第一部分检查点
For any distribution, ask three questions: Where is the centre? How widely are values spread? What shape does the arrangement have? Then choose mean/median, range/IQR/standard deviation, and symmetry/skew/cluster/tail language. One statistic cannot answer all three.
Part II — Choose a spread measure for the job | 第二部分：根据任务选择离散程度指标
36. Full range is easy but extreme-sensitive | 全距简单，但受极端值影响
Range uses only maximum and minimum. If values are 49,50,51,52, range is 3. Add one erroneous value 500 and range jumps to 451. The statistic is useful for total observed span, but it can be dominated by one extreme or data error.
37. Range answers “how far from lowest to highest?” | range question
Range does not tell us whether values fill the interval evenly, cluster in the centre or sit at two ends. Datasets 0,50,100 and 0,1,99,100 share range 100 but different internal arrangements. Use range as an endpoint measure, not a full distribution portrait.
38. IQR answers “how wide is the middle half?” | IQR question
The interquartile range focuses on the middle 50% and is therefore much less affected by the most extreme values. It is useful with median and skewed data because both are based on ordered positions rather than squared distances.
39. IQR can hide tail problems | middle-half limitation
If the middle 50% remains stable while the top 1% becomes extreme, IQR may barely move. This is not failure; it is what the measure is designed to do. Tail-sensitive decisions need tail metrics in addition.
40. Standard deviation answers overall distance from mean under a formal method | SD job
Standard deviation summarises variation around the mean, giving larger influence to observations farther away because the underlying variance uses squared deviations. It is especially interpretable with roughly symmetric distributions and formal statistical models that use mean/variance structure.
41. Standard deviation is scale-dependent | units matter
A standard deviation of 10 marks means something different from 10 dollars or 10 milliseconds. Comparing standard deviations across variables with very different scales can be meaningless without normalisation or domain context.
42. Variance squares the unit | variance units
If the original variable is measured in metres, variance is in square metres. This mathematical property makes variance useful in modelling but less intuitive for general readers. Standard deviation restores the original unit by taking the square root.
43. Mean absolute deviation is another distance-based spread concept | absolute deviation
Absolute-deviation measures use absolute rather than squared distances. Different formulas exist, so specify the centre and method. Recognition matters because “average deviation” can be ambiguous between several measures.
44. Median absolute deviation is robust spread vocabulary | MAD
Median absolute deviation is based on median distances and is resistant to extreme tails. It is specialist vocabulary. General learners should recognise it as a robust spread measure, not confuse MAD with mean absolute deviation.
45. Coefficient of variation normalises spread by mean | CV
Some fields compare relative variability using the coefficient of variation, typically standard deviation relative to mean under appropriate conditions. It is technical and can behave poorly when means are near zero or when scales do not support ratio interpretation. Do not use “relative standard deviation” casually without domain rules.
46. Percentile spread focuses on positions | percentile intervals
“10th to 90th percentile” describes a central 80% interval by rank. This can communicate practical spread without assuming symmetry. Percentile intervals are especially intuitive for latency, income and performance distributions.
47. p95 is not “95% of average” | percentile misconception
The 95th percentile is a position threshold, not 95% of the mean or maximum. If p95 latency is 500 ms, roughly 95% of observations are at or below that value under the definition. Percentile grammar should remain positional.
48. Deciles and quintiles are rank partitions | decile / quintile
Deciles divide ordered data into ten portions; quintiles into five. These terms are common in income and performance reporting. “Top quintile” means the highest fifth under the chosen measure, not “top 5%.”
49. Distribution can be narrow or wide | qualitative spread
“A narrow distribution” suggests values are tightly concentrated; “a wide/broad distribution” suggests greater spread. This is useful plain language when exact spread measures are unnecessary or unavailable.
50. Tight / tightly clustered are intuitive | tight spread
“Scores were tightly clustered around the median.” This is strong descriptive language if the graph supports it. It does not specify a formal clustering algorithm.
51. Widely dispersed is stronger than varied | dispersion adjective
“Responses were widely dispersed” foregrounds broad spread. “Responses varied” is weaker and only says they were not identical. Modifier choice should match visible/quantitative spread.
52. Heterogeneous describes diversity, not only numerical spread | heterogeneity
A heterogeneous group can differ across categories, backgrounds or mechanisms, not merely have high standard deviation. “Heterogeneous population” is broader than “highly variable scores.” Do not use the words as automatic synonyms.
53. Homogeneous means comparatively similar | homogeneous
A homogeneous group is similar in a relevant characteristic. Statistical homogeneity can have formal meanings. In general prose, specify the dimension: homogeneous in age, skill, product type or response pattern.
54. Concentration can be local or systemic | concentrated distribution
A distribution may be concentrated around a centre, in one region, among one customer segment or in a few firms. The noun concentration can describe spatial/statistical arrangement or market dominance. Name the dimension to avoid ambiguity.
55. Market concentration is not statistical spread alone | market concentration
In economics/business, market concentration concerns how market shares are distributed across firms. High concentration means a few firms hold large shares. This is conceptually related to unequal distribution but uses specialised measures such as HHI in some contexts.
56. Geographic concentration describes spatial clustering | geography
“Population is concentrated along the coast.” This is distribution language across physical space, not numerical variance. The same English noun family travels across domains because the underlying idea is arrangement and clustering.
57. Dispersion can be geographic | geographic dispersion
“Customers are geographically dispersed” means spread across many locations. It is a natural professional use of dispersed and does not require statistical calculation.
58. Dense and sparse are arrangement adjectives | dense / sparse
Dense clusters contain many observations in a small region; sparse regions contain few. In data science, density can be formal. In general description, “densely clustered” and “sparsely populated” are intuitive.
59. Uniform distribution is a formal concept | uniform
A uniform distribution assigns equal density/probability across a defined range in a formal model. Empirical data can be “roughly evenly spread” without being statistically uniform. Avoid upgrading visual evenness into a formal distribution claim.
60. Normal distribution is a specific statistical family | normal distribution
Normal distribution is not “a distribution that looks normal/ordinary.” It is a formal statistical distribution with defined shape/properties. General learners should recognise the term but avoid using normal simply because a histogram looks bell-shaped.
61. Bell-shaped is descriptive, not proof of normality | bell-shaped
A histogram may look bell-shaped while deviating from a formal normal distribution in tails or other features. “Approximately bell-shaped” is safer than “normally distributed” without analysis.
62. Skewness is a formal measure of asymmetry | skewness
Skewness quantifies asymmetry under statistical formulas. Skewed can be used qualitatively in data description. Do not attach a formal skewness value unless it is calculated.
63. Kurtosis is specialist tail/shape vocabulary | kurtosis
Kurtosis has formal definitions related to the fourth moment and tail/peak characteristics depending interpretation. It is specialist recognition vocabulary. Avoid reducing it to “peakedness” as a universal plain definition.
64. Outlier-sensitive spread vs robust spread | robustness
Range and standard deviation can be strongly affected by extremes. IQR and median absolute deviation are more robust. The correct choice depends on whether extremes are noise or meaningful parts of the decision problem.
65. Robust does not mean more accurate universally | robust ≠ superior
A robust statistic deliberately reduces influence of extreme observations. If extremes are genuine and consequential, reducing their influence may hide risk. “Robust” describes behaviour under unusual values/assumptions, not a universal quality ranking.
66. Scale and spread must be interpreted together | relative spread
A standard deviation of 5 around mean 10 is large relative to the centre; SD 5 around mean 1,000 may be tiny. Sometimes relative measures are useful. Do not judge “high variability” from the spread number alone.
67. Spread can shrink while centre moves | independent dimensions
A class can improve its mean from 60 to 75 while scores become more tightly clustered. Centre rises, spread falls. Another class can keep mean 75 while spread widens. Always separate location from variability.
68. Spread can widen without anyone getting worse | divergence
If top performers improve faster while lower performers remain stable, variance increases even though nobody declines. “Performance became more variable” does not mean average performance worsened.
69. Spread can narrow through convergence | convergence
Values may move closer together over time. Convergence can describe narrowing differences across groups/values, though formal meanings vary by field. It is the opposite direction of divergence, not necessarily movement toward a particular centre.
70. Divergence means moving farther apart | divergence
Group outcomes may diverge even if the overall mean stays constant. This is important in education, inequality and regional analysis. Do not use “variation increased” without checking whether spread actually widened.
71. Polarisation is stronger than spread | polarisation
Polarisation usually implies movement toward opposing ends or groups, not merely wider variability. Ratings splitting into 1s and 5s may be polarised; a uniform spread from 1 to 5 is varied but not necessarily polarised.
72. Segmentation describes meaningful subgroups | segmentation
Business/education teams may segment a distribution into groups. Segmentation can be a practical classification, not a naturally occurring cluster. Avoid implying segments were discovered statistically unless a method supports that claim.
73. Tail risk is not the same as wide spread | tail risk
A distribution can have modest central spread but rare catastrophic tail values. Risk decisions may focus on tails rather than standard deviation or IQR. “Low average variability” does not eliminate extreme-event risk.
74. Ceiling and floor effects compress observed spread | measurement bounds
If many scores hit a maximum, the observed upper tail is truncated and spread may appear smaller than underlying ability variation. Lesson 038 touched this for centres; here the effect is on distribution shape and variability.
75. Truncation changes distribution | truncated distribution
Removing observations below/above a threshold changes centre, spread and shape. A truncated sample should not be described as though it were the full population distribution.
76. Censoring retains partial information differently | censored data
Censoring records that a value lies beyond a limit without observing its exact magnitude. It differs from truncation, where such observations may be absent. Technical interpretation requires domain methods; vocabulary should preserve the distinction.
77. Sampling can make spread look different | sampling variability
Different samples from the same population can show different ranges, standard deviations and tails. Sample spread is an estimate/observation, not automatically the exact population spread.
78. Measurement error can inflate observed spread | error variance
If instruments or scoring are noisy, observed variability includes measurement noise as well as real differences. “Participants varied widely” may overstate real heterogeneity if measurement error is substantial.
79. Reliability and variability are not opposites | reliability
A reliable instrument can measure a genuinely variable population consistently. High spread among people does not imply poor measurement reliability. Likewise low spread does not prove reliability.
80. Part II checkpoint: spread measure = question + sensitivity | 第二部分检查点
Use range for full observed span, IQR for middle-half spread, standard deviation for mean-based overall variation, percentiles for positional tails, and robust measures when extreme values should have less influence. Always state the scale and remember that no spread measure shows the whole shape.
Part III — Read distribution shape across real domains | 第三部分：在真实领域中读懂分布形状
81. Exam scores: same mean, different spread | 考试成绩
Two classes can both have mean 70. Class A scores mostly 65–75; Class B scores 30–100. “The classes performed similarly” is too broad. Their centres match, but Class B is much more variable. A stronger report separates location and spread: “Both classes had a mean of 70, but scores were far more dispersed in Class B.”
82. Exam scores: narrow spread can mean consistency, not excellence | consistency vs level
A class tightly clustered around 50 is consistent but not necessarily high-performing. A class widely spread around 80 can have a higher centre and greater variability. “Consistent” describes spread; “strong” evaluates level. Keep the dimensions separate.
83. Easy tests can create left-skew and ceiling effects | exam shape
If most students score near the maximum and a smaller group extends downward, the distribution may be left-skewed and compressed at the top. Mean and standard deviation can become harder to interpret because the test ceiling limits observed spread. This is a measurement pattern, not proof that all students have similar ability.
84. Hard tests can create right-skew and floor effects | exam floor
If most scores cluster near the bottom with a long upper tail, the distribution may be right-skewed. A floor effect can compress lower differences. “Most students failed” does not by itself describe the shape; inspect where scores cluster and how far the upper tail extends.
85. Income distributions are often right-skewed | income
A small number of very high incomes can extend the upper tail and pull the mean above the median. “Income is widely distributed” is vague; a better description might be “Income is strongly right-skewed, with a long upper tail and mean well above median.” That tells the reader centre and shape together.
86. Inequality is not identical to variance | inequality vocabulary
Income inequality can be measured with specialised metrics such as Gini coefficients and percentile shares. Variance or standard deviation captures spread but is not synonymous with inequality in every analytical sense. Use the domain’s conventional metric rather than calling any wide distribution “high inequality.”
87. Wealth can be even more tail-heavy than income | tail emphasis
Wealth distributions may contain extremely long upper tails. Median wealth can describe the midpoint household while total wealth concentration depends heavily on the upper tail. A centre-only summary misses who holds the mass of the distribution.
88. Waiting times often have a long right tail | operations
Most customers may wait briefly while a few wait much longer. Median tells the middle experience; p90/p95 reveal the tail; mean reflects total waiting burden. “Average wait is six minutes” cannot tell whether almost everyone waited six minutes or most waited one minute while a few waited an hour.
89. Latency distributions require tail vocabulary | computing latency
Web services often report median, p95 and p99 latency because user experience in the upper tail matters. A low median and high p99 means most requests are fast while a small fraction are very slow. “Performance is stable” would be misleading if tail latency is severe.
90. Error rates can be concentrated by source | operational clustering
If 80% of errors come from 5% of components, errors are concentrated rather than evenly distributed. This is not merely a high average error rate. Concentration language identifies where problems cluster across units.
91. Manufacturing: same mean, different tolerance performance | quality
Two factories can have mean diameter exactly on target. Factory A has tiny standard deviation; Factory B has wide spread causing many parts outside tolerance. “Both factories are on target” describes centre only. Quality requires spread relative to specifications.
92. Process capability is specialist vocabulary | capability
Manufacturing uses formal process-capability measures comparing process spread with specification limits. General learners should recognise that standard deviation alone does not tell whether a process meets tolerance. Technical capability indices belong to quality engineering, not generic spread vocabulary.
93. Geographic distribution can be clustered, dispersed or concentrated | geography
Population can be concentrated along a coastline, clustered around cities or dispersed across rural areas. These terms describe spatial arrangement rather than statistical standard deviation. The common conceptual job is still distribution shape across a space.
94. Spatial density and spatial concentration differ | density vs concentration
Density measures amount per area under a defined unit; concentration describes how unevenly that amount is distributed. A country can have low average population density but high concentration in a few cities. Lesson No.037 owns density as a ratio-like measure; Lesson No.039 owns arrangement.
95. Market concentration asks whether share is dominated by a few firms | market structure
A market with four equal firms and one with one dominant firm plus many small firms can have similar firm counts but different concentration. Market share distribution, not merely number of firms, determines concentration. Technical competition analysis uses specialised measures.
96. Customer revenue concentration creates dependency risk | customer concentration
If one customer supplies 40% of revenue, revenue is highly concentrated even if thousands of customers exist. This distributional fact matters for risk because losing one large customer has outsized impact. Count of customers alone would hide concentration.
97. Portfolio concentration is not return variability | finance distribution
A portfolio can be concentrated in a few assets yet show low recent volatility, or diversified across many assets but volatile due to common shocks. Concentration describes allocation; volatility describes movement over time. Lesson 036 owns volatility; Lesson 039 owns allocation spread.
98. Survey responses can cluster at the centre or extremes | survey shape
Two surveys can both average 3 on a five-point scale. One has most responses at 3; another splits between 1 and 5. The first is centrally concentrated; the second polarised/bimodal. Average alone erases opinion structure.
99. Acquiescence or response style can shape distributions | survey caution
Survey distributions may reflect response tendencies, question wording or scale design as well as underlying attitudes. Describing the shape is observational; explaining why the shape exists requires evidence.
100. Categorical distribution is category frequencies/shares | categorical data
A categorical distribution describes how observations are allocated across categories: 40% red, 35% blue, 25% green. Mean and standard deviation may be meaningless if categories lack numeric order. Distribution remains useful without numerical centre.
101. Frequency distribution links Lesson 039 and 040 | frequency distribution
A frequency distribution records how often each value/category occurs. Lesson 039 owns the overall arrangement; Lesson 040 will own the lexical distinction among frequency, prevalence, incidence and recurrence. The same table can support both jobs.
102. Relative-frequency distribution uses proportions | relative frequency
Instead of raw counts, categories can be expressed as shares/percentages. Lesson 037 supplies the denominator logic. The distribution shape should be similar if all counts share the same denominator, but the numerical scale changes.
103. Histogram shows numeric distribution in bins | histogram
A histogram groups numerical values into intervals and displays frequencies/densities. Bin width can change apparent peaks and shape. Saying “the data are bimodal” from a histogram alone may depend on bin choices; confirm with appropriate analysis.
104. Bar chart is not histogram | bar vs histogram
Bar charts typically display categorical values with separate bars; histograms display numeric intervals with meaningful adjacency/order. Treating a category bar chart as a continuous distribution can lead to wrong shape language.
105. Box plot compresses centre, quartiles and extremes | box plot
A box plot shows median, quartile spread and whiskers/outliers under a convention. It is useful for comparing distributions compactly, but it can hide multimodality because the interior shape is compressed.
106. Density plot smooths distribution shape | density plot
Kernel density plots smooth observations into a continuous-looking curve. The smoothing parameter affects apparent peaks. Use them as estimated shape displays, not literal raw-frequency counts.
107. Violin plot combines density and summary | violin plot
Violin plots visualise distribution density and often include median/quartile markers. They can reveal multimodality hidden by box plots, but unfamiliar audiences may need explanation.
108. Scatter is not distribution by itself | scatter
A scatter plot displays pairs of variables and relationship structure. Each variable also has a marginal distribution, but “scatter” describes two-dimensional spread/relationship. Do not call a scatter plot a histogram-like distribution display.
109. Distribution over time vs time trend | two axes
A time series describes how a value changes over time. A distribution across all time points describes how often values occupy levels. The same data can have both a trend and a distribution, but the questions differ.
110. Day-of-week distribution is categorical, not trend | temporal categories
Counts by Monday, Tuesday, etc. form a distribution across weekdays. A “Monday peak” is a category-frequency maximum, not necessarily a trend because weekday categories repeat cyclically rather than form one directional timeline.
111. Log transformation can reduce visible right skew | transformation
Taking logs can compress high values and change distribution shape. A distribution may look more symmetric on log scale. Do not compare raw and transformed standard deviations/means as though they share the same units.
112. Standardisation changes scale, not rank structure | z-scores
Standardising values to z-scores shifts/rescales a variable around mean/standard deviation under a common convention. The relative ordering remains but units change. Exact statistical interpretation belongs to formal methods.
113. Min–max normalisation changes range by construction | rescaling
Rescaling values into 0–1 changes range and standard deviation numerically without changing basic ordering/shape linearly. Spread measures are scale-dependent, so transformed values must be labelled.
114. Combining groups can create multimodality | mixture distribution
If two populations with different centres are pooled, the combined distribution may show multiple peaks or broad spread. What looks like high individual variability may partly be subgroup mixture. Segment before explaining.
115. Simpson-like aggregation can change distribution story | aggregation
Overall distributions can differ from subgroup distributions because group sizes and centres differ. The technical phenomenon can be complex; the language repair is simple: inspect subgroup structure before treating the aggregate shape as one homogeneous population.
116. Outlier can form its own meaningful subgroup | not noise automatically
Several “outliers” may cluster together and reveal a second process or segment. Before deleting extremes, ask whether they are measurement errors, rare genuine events or evidence that the population is mixed.
117. Anomaly describes unexpected pattern, not a spread measure | anomaly
An anomalous observation or region deviates from expected pattern. It can be an outlier, structural break or unusual cluster. The noun is interpretive; distribution measures only show arrangement.
118. Long tail can dominate totals | contribution concentration
A small number of high values may account for a large share of total output, revenue or cost. This creates both distributional skew and contribution concentration. Report centre plus tail contribution when the total burden is driven by extremes.
119. Pareto-like language is descriptive only with evidence | 80/20 caution
People casually call any concentrated distribution “80/20” or “Pareto”. Do not use the label unless the observed relationship approximately supports it or the formal Pareto model is actually intended. “A small share accounts for most outcomes” may be safer.
120. Even distribution can mean balanced allocation, not statistical uniformity | even
“Resources are evenly distributed across schools” may mean similar allocations, not values drawn from a mathematical uniform distribution. Use evenly spread/balanced for general communication; reserve uniform distribution for statistical probability contexts.
121. Uneven distribution can describe imbalance without formal metric | uneven
“Benefits were unevenly distributed” is a strong general phrase when some groups received much more than others. It does not specify variance, Gini coefficient or another inequality metric.
122. Distributional shift means arrangement changed | distribution shift
A distributional shift can involve centre, spread, shape or category shares changing. Machine learning uses distribution shift technically. In general analysis, specify what shifted: mean rose, variance widened, categories rebalanced, tail thickened.
123. Concept drift is not any distribution shift | ML caution
Machine learning distinguishes various forms of data/concept shift. Do not use concept drift as a generic synonym for “the distribution changed.” Technical labels should follow the field’s formal definition.
124. Distribution overlap describes similarity of ranges/densities | overlap
Two groups can have different means but highly overlapping distributions. “Group A is higher” may be true on average while many individuals in B exceed individuals in A. Distribution overlap prevents group-average differences from becoming deterministic individual claims.
125. Separation describes how distinct groups are | separation
Well-separated distributions have little overlap; poorly separated groups overlap strongly. Classification performance often depends on separation, not merely difference in group means.
126. Within-group and between-group variability are different | levels of variation
Groups can differ strongly in their means while each group is internally consistent, or groups can share means while each is internally variable. “Variation between schools” and “variation within schools” answer different questions.
127. Pooled spread can exceed subgroup spread | mixture effect
If two tight clusters sit far apart, the combined distribution has large spread even though each subgroup is homogeneous. This is why heterogeneity can arise from between-group separation rather than within-group noise.
128. Spread can be desirable or undesirable depending job | value neutrality
Wide product variety may be desirable; wide manufacturing dimension spread may be poor quality. Broad student achievement spread may signal differentiated needs. Distribution terms describe first; evaluation comes second.
129. “More consistent” usually means lower variability, not higher mean | consistency
A system can become more consistent while its average deteriorates. “More consistent performance” should refer to narrower spread/reduced variation, not automatically better level.
130. Part III checkpoint: distribution language protects against average-only thinking | 第三部分检查点
Across domains, ask what is hidden behind the centre: tails, clusters, subgroup mixtures, spatial concentration, category imbalance, overlap or inconsistent spread. Never call two groups similar merely because their averages match.
Part IV — Mandarin-to-English distribution control | 第四部分：中文母语学习者的分布与离散词汇转换
Mandarin words such as 分布、离散、分散、集中、波动、聚集 and 尾部 overlap across statistics, geography, finance and ordinary description. English separates time movement from cross-sectional arrangement more sharply. The key question is whether values are changing over time or arranged differently across observations at one period.
131. 分布 = distribution | 分布的核心是“如何排列”
“成绩分布” → score distribution. “人口分布” → population distribution. “收入分布” → income distribution. The common idea is arrangement across values, categories or space. Do not translate every 分布 as spread; distribution is the broader owner.
132. 离散程度 = dispersion / spread / variability | 离散程度
In technical statistics, dispersion is formal. In general explanation, spread or variability is often clearer. “成绩离散程度较高” → Scores were widely spread / showed greater variability. Choose register and whether an exact spread measure is available.
133. 分散 = dispersed / spread out | 分散强调不集中
“数据点较分散” → The observations are widely dispersed. “人口分散在多个地区” → The population is dispersed across several regions. The word travels between statistical and spatial domains naturally.
134. 集中 = concentrated | 集中强调聚在有限区域
“分数集中在 70 附近” → Scores were concentrated around 70. “收入集中在少数人手中” → Income/wealth is concentrated among a small group. Same word family, different domain interpretation.
135. 聚集 = cluster / clustered | 聚集
“数据在两个区域聚集” → Values cluster in two regions / form two clusters. Do not automatically write “cluster analysis shows…” unless a formal clustering method was used.
136. 波动 is temporal in Lesson 036; variability is broader here | 波动 vs 离散
“价格波动很大” over time may be high volatility / large fluctuations. “不同学生成绩差异很大” at one exam sitting is high variability / wide spread, not volatility. The time axis decides the vocabulary.
137. 方差 = variance | 方差
Use variance for the formal squared-deviation statistic. It should not be translated as general “difference.” “组内方差” → within-group variance; “方差增大” → variance increased.
138. 标准差 = standard deviation | 标准差
“标准差较小” → The standard deviation was small or scores were tightly clustered for a plain audience. Do not abbreviate as SD without defining it for unfamiliar readers.
139. 极差 = range | 极差不是“极端差异”
Statistical 极差 means maximum minus minimum, conventionally range. “极差为 80” → The range was 80. If you mean the interval itself, say values ranged from X to Y.
140. 四分位距 = interquartile range (IQR) | 四分位距
Use the formal term directly. “四分位距较小” → The IQR was narrow/small, indicating that the middle 50% was tightly clustered. Do not translate it as “quartile distance.”
141. 四分位数 = quartile | quartile
“第一四分位数” → first quartile (Q1); “第三四分位数” → third quartile (Q3). Exact computational convention can vary; retain the source values/method.
142. 百分位数 = percentile | percentile
“第 90 百分位” → 90th percentile. Do not write “90 percentile” or interpret it as 90% score. A percentile is a positional threshold within a distribution.
143. 分位数 = quantile | quantile is technical
Quantile is the broader technical family containing quartiles, percentiles and related cut points. General readers may not need the term unless discussing methodology.
144. 偏态 / 偏斜 = skew / skewness / skewed | 偏态
“分布右偏” → The distribution is right-skewed. “偏度为…” → skewness was… if formally calculated. Use adjective for descriptive shape and noun skewness for the property/measure.
145. 右偏 / 正偏 = right-skewed / positively skewed | right skew
Both terms occur in statistics. Right-skewed is often more intuitive because it names the long-tail direction. Positive skew refers to the conventional sign of skewness. Do not confuse “positive” with desirable.
146. 左偏 / 负偏 = left-skewed / negatively skewed | left skew
Again, negative describes mathematical direction, not poor performance. An easy test can produce negative/left skew because many scores cluster near the high end.
147. 对称分布 = symmetric distribution | 对称
If the data are only visually approximate, say roughly/approximately symmetric. “Perfectly symmetric” is a stronger mathematical statement.
148. 均匀分布 = uniform distribution, but only when formal | 均匀分布
For a probability model, uniform distribution can be exact. If empirical resources are simply spread evenly across regions, use evenly distributed. Chinese 均匀 does not automatically imply the formal probability family.
149. 正态分布 = normal distribution | 正态不是“普通”
Use the technical term normal distribution. Do not explain it as “a normal/ordinary distribution.” Normally distributed has statistical meaning.
150. 钟形 = bell-shaped | 钟形不等于证明正态
“大致呈钟形” → approximately bell-shaped. This is a visual description and is safer than declaring normality without formal assessment.
151. 尾部 = tail | 尾部
“上尾” → upper tail; “下尾” → lower tail. “尾部风险” → tail risk in finance/risk contexts. Distribution tails refer to extreme regions, not simply end points.
152. 长尾 = long tail / long-tailed | 长尾
In statistical description, long-tailed can describe extended extremes. In business, “the long tail” can describe many low-volume niche items. Check domain before carrying the term across contexts.
153. 厚尾 / 重尾 = heavy-tailed, with technical caution | 重尾
Use heavy-tailed only when the field’s formal concept is intended. Do not translate every distribution with visible extremes as heavy-tailed.
154. 峰 = peak / mode depending sentence | distribution peak
“分布有两个峰” → The distribution has two peaks / is bimodal if modes are genuinely distinct. “最高频值” → mode. Peak is the visual/high-density region; mode is the most frequent value/category.
155. 双峰 = bimodal | 双峰
“呈双峰分布” → The distribution is bimodal. Do not infer two causal populations automatically; bimodality is a shape description first.
156. 多峰 = multimodal | 多峰
Use multimodal when several modes/peaks are present. It is technical but useful for recognising mixture structures.
157. 离群值 = outlier | 离群值
Outlier is the conventional statistical term. Do not translate it as “exception value.” Whether a point qualifies as an outlier can depend on the method/context.
158. 异常值 = anomalous value / outlier / abnormal reading | 异常值
Chinese 异常值 may refer to an outlier, measurement anomaly or impossible/invalid value. English should distinguish statistical distance from data-quality abnormality. Not every outlier is an error.
159. 极端值 = extreme value | 极端值
Extreme value is neutral and can refer to high/low observations by position. It is safer than outlier when no statistical rule labels the observation unusual.
160. 高度集中 = highly concentrated | 高度集中
“收入高度集中于顶部群体” → Income is highly concentrated among the top group. If a specific concentration metric exists, report it rather than relying only on adjective strength.
161. 高度分散 = widely dispersed / highly variable | 高度分散
Use widely dispersed for spatial/statistical arrangement or highly variable for broad numeric differences. High volatility belongs to time movement and is not the default translation.
162. 差异大 can be high variability / wide spread, not “large difference” only | 差异大
Across many observations, values vary widely / show substantial variability is more natural than repeatedly saying “there is a big difference.” Difference often compares specific items; variability describes a distribution.
163. 一致性高 can mean low variability / high consistency | 一致性
“评分一致性高” may mean raters agree (reliability/agreement) or scores vary little. These are not identical. Clarify whether consistency concerns distribution spread or measurement agreement.
164. 收敛 = convergence | 收敛
“地区差距逐渐收敛” → Regional outcomes are converging. In mathematics, convergence has formal meanings. In general comparative data, it means values/groups move closer together.
165. 发散 = divergence | 发散
“两组结果开始发散” → The two groups began to diverge. This can describe separation over time; the cross-sectional consequence is wider between-group spread.
166. 两极化 = polarisation | 两极化
Use polarisation when values/opinions increasingly concentrate toward opposing ends or camps. Wider spread alone is not enough.
167. 重叠 = overlap | 分布重叠
“两组分布高度重叠” → The two distributions overlap substantially. Different means can coexist with substantial overlap; this matters when translating group-level differences into individual predictions.
168. 分离 = separation | group separation
“两组分离明显” → The groups are well separated. In clustering/classification, separation can become technical; in general prose, name the variable and amount of overlap if possible.
169. 分布变宽 / 变窄 = distribution widened / narrowed | spread change
“分布变宽” → The distribution became more dispersed / widened. “分布变窄” → Values became more tightly concentrated / the spread narrowed. Do not say “the distribution grew” unless a different quantity is intended.
170. 组内差异 / 组间差异 = within-group / between-group variation | levels
These terms should stay separate. Large overall variability may come from differences between groups even when each group is internally homogeneous.
171. 集中度 = concentration / concentration measure | 集中度
Market concentration, geographic concentration and response concentration use different metrics. Do not translate 集中度 as “concentration rate” unless that is the domain’s conventional term.
172. 密度 = density, but separate from spread | 密度
Density can mean amount per area or statistical probability density. A high-density region contains many observations relative to a value interval/space. It is not the same as low standard deviation, though the concepts can relate.
173. 频数分布 = frequency distribution | bridge to Lesson 040
“频数分布” → frequency distribution. Here frequency supplies counts per value/category; distribution describes the arrangement. Lesson 040 will deepen how often/occurrence vocabulary.
174. 相对频数分布 = relative-frequency distribution | proportion-based distribution
Use relative-frequency distribution when category/value counts are converted to proportions/percentages. Lesson 037’s denominator logic applies.
175. Part IV checkpoint: translate the shape, not only the statistic | 第四部分检查点
Mandarin-to-English distribution control requires three layers: exact measure (标准差/IQR/极差), qualitative arrangement (集中/分散/偏态/双峰), and domain (statistical, spatial, market, technical). The word should tell the reader what kind of spread or shape is being described.
Part V — Distribution failure laboratory | 第五部分：分布失误实验室
The fastest way to see why distribution vocabulary matters is to hold one summary constant while changing the shape around it. Every case below is designed to expose a specific failure: average-only thinking, spread-only thinking, tail blindness, subgroup mixing, visual overinterpretation or outlier misuse.
176. Case 1 — same mean, narrow vs wide distribution | 同均值，不同离散
Class A: 68,69,70,71,72. Class B: 30,50,70,90,110. Both mean 70. “Both classes performed the same” ignores variability. Better: “Both classes had the same mean, but Class B showed much greater spread.”
177. Case 2 — same median, radically different tails | 同中位数，不同尾部
Two datasets share median 50. One ranges 45–55; another ranges -1,000 to 5,000. Median alone suggests the same middle, but risk and resource burden differ sharply. Tail-sensitive decisions need more than the midpoint.
178. Case 3 — same standard deviation, different shape | 同标准差，不同形状
A symmetric unimodal distribution and a bimodal mixture can be constructed with similar standard deviations. “Variability is the same” may be true numerically, but clustering differs. Spread is not shape.
179. Case 4 — same range, different internal arrangement | 同极差，不同内部结构
0,50,100 and 0,1,2,98,99,100 share range 100. One has a central observation; the other clusters near extremes. Range sees only endpoints. It cannot tell whether the interval is filled, empty in the middle or clustered.
180. Case 5 — IQR stable while upper tail explodes | IQR 不变，尾部扩大
The middle 50% stays within 45–55 while the largest few values increase from 70 to 10,000. IQR is unchanged, but tail risk grows dramatically. Robust middle-spread measures intentionally ignore some extreme movement.
181. Case 6 — standard deviation rises because one outlier appears | 单个离群值扩大标准差
A tightly clustered dataset receives one extreme value. Standard deviation rises sharply. Before declaring the population “more variable,” ask whether the point is a genuine observation, measurement error or new subgroup.
182. Case 7 — deleting an outlier hides a real event | 删掉离群值掩盖真实事件
A server has one ten-second latency event among thousands of 100-ms requests. Removing it makes the distribution look clean but erases a severe user experience. Outlier handling should follow the analytical question, not aesthetic preference.
183. Case 8 — “outlier” is actually a data-entry error | 真错误 vs 真极端值
A height of 18 metres appears in a human dataset. Here validation suggests a unit/data-entry error. Distribution analysis should not treat invalid data as a meaningful tail. The first question is data quality, not robust statistics.
184. Case 9 — bimodality reveals two groups | 双峰揭示混合群体
Scores cluster near 40 and 90. Overall mean 65 is unrepresentative. Separating novice and advanced learners may explain the two peaks. But bimodality alone does not prove the causal identity of the groups; it signals a mixture worth investigating.
185. Case 10 — visual bimodality created by bin width | 直方图分箱制造双峰
A histogram looks bimodal with narrow bins but unimodal with wider bins. The display choice changes visual peaks. Before naming the underlying distribution bimodal, inspect raw data or robust density/method evidence.
186. Case 11 — bell-shaped is not proof of normality | 钟形不等于正态
A histogram looks roughly bell-shaped, but tails are much heavier than a normal model. “Normally distributed” overstates what a visual impression establishes. “Approximately symmetric and bell-shaped” is descriptive and safer.
187. Case 12 — normal-looking centre, dangerous tail | 中间正常，尾部危险
Most financial losses are modest, but rare extreme losses dominate risk. Mean and standard deviation around ordinary periods can look stable. Tail behaviour matters because the decision concerns catastrophic extremes, not typical days.
188. Case 13 — same mean and SD, different tail thickness | 同均值同标准差，不同尾厚
Distributions can match first two moments but differ in how probability mass sits in tails. This is why mean + standard deviation cannot fully describe shape. Technical tail measures may be needed for specialised risk work.
189. Case 14 — polarisation vs uniform spread | 两极化 vs 均匀分散
Survey A is split between 1 and 5. Survey B is evenly spread across 1–5. Both can have similar mean and broad spread, but only A is strongly polarised. Polarisation implies concentration at opposing ends, not merely high variability.
190. Case 15 — narrow distribution around bad target | 一致但很差
A factory produces every part at 9.5 mm when target is 10.0 ±0.1. Spread is tiny; quality is consistently wrong. “Low variability” is not “good performance.” Centre and specification matter.
191. Case 16 — wide distribution around excellent mean | 平均优秀但差异很大
A class mean is 85, but scores range from 40 to 100. “High-performing class” may be fair in centre terms but hides a struggling tail. Educational decisions need subgroup/spread information.
192. Case 17 — concentration vs density | 集中 vs 密度
Country A has high population density everywhere; Country B has low national density but nearly everyone concentrated in two cities. Density and concentration answer different spatial questions.
193. Case 18 — market count vs concentration | 企业多不代表竞争分散
A market contains 100 firms, but one holds 80% share. Firm count is large while market concentration is high. Distribution of shares, not count alone, determines concentration.
194. Case 19 — customer count vs revenue concentration | 客户多不代表收入分散
A business has 5,000 customers but 60% of revenue comes from one client. “Highly diversified customer base” based only on count is misleading. Revenue distribution is concentrated.
195. Case 20 — two group means differ but distributions overlap heavily | 均值不同但大量重叠
Group A mean 70; Group B mean 75. Both have SD 20. Large overlap means many individuals in A exceed many in B. “B is better than A” may describe average, not every member. Group-level statements should not become deterministic individual claims.
196. Case 21 — same overall spread, different within/between structure | 总体离散相同，结构不同
Dataset A is one wide cloud. Dataset B is two tight clusters far apart. Overall variance can be similar. One reflects continuous individual diversity; the other subgroup separation. Shape and grouping change interpretation.
197. Case 22 — pooling creates false heterogeneity | 合并组制造“高差异”
Each school has tightly clustered scores, but school means differ. Pooled national spread is large. “Students vary widely within schools” is false; variation is mainly between schools. Level of analysis matters.
198. Case 23 — standardising hides original units | 标准化后失去原单位
Z-scores make variables comparable on a standardised scale but no longer express dollars, seconds or marks. “SD = 1” after standardisation is partly true by construction under the method, not evidence that raw distributions had equal spread.
199. Case 24 — log transform changes apparent spread | 对数变换改变形状
Raw incomes are highly right-skewed; log incomes look more symmetric. Neither graph is “the real distribution” without context. They represent the variable on different scales. State the transformation.
200. Case 25 — p95 improves while median worsens | 中位数与尾部方向相反
After an update, median latency rises from 100 to 120 ms, but p95 falls from 2 s to 500 ms. Typical requests are slightly slower, tail experience dramatically better. “Performance improved” needs dimension-specific wording.
201. Case 26 — median improves while p99 collapses | 中间更好，极尾更差
Median latency falls, but rare requests become much slower. Centre-only dashboards can miss tail deterioration. Report the distribution feature relevant to users/risk.
202. Case 27 — range widens because one valid rare event | 全距扩大但主体不变
A rare but genuine 100-point observation extends range; IQR and median stay stable. “The whole population became more variable” may overstate change. Better: “The observed range widened because of one extreme case, while central spread remained stable.”
203. Case 28 — IQR narrows because central group homogenises | 中间收窄
Middle 50% converges while tails remain unchanged. “Distribution narrowed” may be too broad; “IQR narrowed” is exact. Name the spread measure when only one part changes.
204. Case 29 — standard deviation falls after truncation | 截断制造低离散
A programme excludes applicants below a threshold. The admitted group naturally has narrower score range/SD than all applicants. “Participants are more homogeneous” is descriptively true but partly caused by selection design.
205. Case 30 — ceiling compresses high performers | 天花板压缩分布
An easy test produces many 100s. Observed SD shrinks, but ability differences above the test ceiling are invisible. “High scorers are homogeneous” is unsupported.
206. Case 31 — missing low values changes skew | 缺失改变偏态
If low-performing participants disproportionately fail to respond, the observed distribution shifts upward and may look more symmetric. Distribution shape describes observed data, not automatically the full target population.
207. Case 32 — measurement noise inflates variance | 测量噪声扩大离散
Repeated noisy measurements produce wider observed spread than true values. “Individuals differ greatly” may actually reflect unreliable measurement. Variability source needs diagnosis.
208. Case 33 — identical standard deviation on different scales | 跨尺度误比
SD 10 on a 0–100 score and SD 10 on a 0–10,000 income scale are not comparable as relative variability. Units and central scale matter.
209. Case 34 — high CV near zero becomes unstable | 相对变异陷阱
Coefficient of variation divides by mean; when mean approaches zero, the ratio can become huge or unstable. Technical relative-spread measures have assumptions and should not be used mechanically.
210. Case 35 — percentiles depend on population/reference | 百分位依赖参照群体
A score at 90th percentile nationally may be 60th percentile within an elite cohort. Percentile is a relative position in a specified distribution, not an intrinsic property of the score.
211. Case 36 — top 10% is not 90% correct | percentile confusion
Being at the 90th percentile means relative rank, not necessarily scoring 90/100. A difficult test can have 90th-percentile score 70; an easy test may have it 98.
212. Case 37 — quartile boundary differs by method | software conventions
Different statistical packages can compute sample quartiles/percentiles using slightly different interpolation conventions. If exact cut points matter, report software/method rather than assuming every Q1 is identical.
213. Case 38 — box plot hides bimodality | compressed visual
Two groups with a bimodal mixture may produce a box plot that looks ordinary. A violin/histogram reveals the two peaks. Visualisation choice can conceal shape even while centre/IQR are correctly shown.
214. Case 39 — density plot invents smoothness | smoothing artefact
Kernel density estimates draw a smooth curve through finite observations. Small bandwidth creates many peaks; large bandwidth erases them. Treat the density as an estimate, not raw reality.
215. Case 40 — population distribution changes because categories redefined | classification shift
If a business changes customer-segment definitions, category shares/distribution shift even if behaviour is unchanged. Distributional change may reflect measurement/classification rather than real-world change.
216. Case 41 — geographic dispersion vs travel burden | spatial meaning
Customers can be geographically dispersed but still near transport hubs, producing lower service burden than a less dispersed population in inaccessible locations. Spatial dispersion describes arrangement, not operational difficulty automatically.
217. Case 42 — product variety vs demand concentration | assortment
A store offers 10,000 products, but 80% of sales come from 50. Product catalogue is diverse; demand distribution is concentrated. Variety of options and concentration of outcomes are different.
218. Case 43 — high variance does not mean random | structured variation
Scores vary widely because two curriculum tracks have different centres. Variance is high, but variation is structured rather than random. “High variability” should not imply unpredictability without evidence.
219. Case 44 — low variance does not mean unbiased | systematic error
A faulty sensor reports 10.0 repeatedly when true value is 12.0. Readings have low spread but are systematically wrong. Precision/consistency is not accuracy.
220. Case 45 — two groups with equal SD but different means | location vs spread
Group A mean 50 SD 10; Group B mean 80 SD 10. They have similar spread but different centre. “The distributions are identical” is false; only variability is similar.
221. Case 46 — two groups with equal mean but different SD | centre vs spread
Both mean 70; SDs 5 and 25. “Same average” does not mean same consistency or overlap with thresholds. Report spread.
222. Case 47 — different means but complete overlap is impossible only in degenerate cases | avoid deterministic language
In real noisy data, group distributions often overlap even when means differ meaningfully. Avoid sentences such as “Group A individuals are higher” from group means alone. Say “Group A had a higher mean” unless classification evidence supports individual prediction.
223. Case 48 — concentration can increase while variance falls | concentration/spread alignment
If observations move toward the centre, variance may fall and concentration increase. But market concentration can increase through share dominance in a way that does not map directly to ordinary variance. Domain definition still matters.
224. Case 49 — “more diverse” is not always wider numerical spread | diversity
A team can become more diverse in categorical background while salaries become more tightly clustered. Diversity and statistical spread may concern different dimensions. Name the variable.
225. Part V checkpoint: distribution failures are usually dimension failures | 第五部分检查点
The recurring mistake is to let one dimension stand for all others: mean for spread, SD for shape, range for concentration, volatility for cross-sectional variability, outlier for error, density for concentration, subgroup difference for individual certainty. Keep location, spread, shape, tail, grouping and domain separate.
Part VI — FENCE, drills and mastery | 第六部分：FENCE、练习与真正掌握
226. Build the distribution FENCE | 分布 FENCE
| Fence | Question | Language job |
|---|---|---|
| F0 Variable | What is distributed? | score, income, location, category |
| F1 Location | Where is the centre? | mean, median, mode |
| F2 Spread | How wide? | range, IQR, SD, variability |
| F3 Shape | Symmetric or skewed? | right/left skew, bell-shaped |
| F4 Peaks | One cluster or several? | unimodal, bimodal, clusters |
| F5 Tails | What happens at extremes? | tails, outliers, percentiles |
| F6 Measure | Which spread statistic suits the job? | range/IQR/SD/MAD/etc. |
| F7 Domain | Statistical, spatial, market or technical? | preserve domain meaning |
| F8 Transfer | Can I paraphrase without flattening the shape? | meaning-preserving output |
227. Build a distribution-control card | 分布控制卡
| Field | Example |
|---|---|
| Variable | monthly household income |
| Centre | median = 5,000 |
| Spread | IQR = 2,000 |
| Shape | right-skewed |
| Tail | small high-income upper tail |
| Clusters | none established |
| Domain | income distribution |
| Companion | mean + percentile shares |
228. Practice A — same centre, different spread | 练习 A
- 68,69,70,71,72
- 30,50,70,90,110
- 60,65,70,75,80
Describe which groups share a centre and how their spread differs. Use tightly clustered, moderately spread, widely dispersed before reaching for a formal statistic.
229. Practice B — same spread, different shape | 练习 B
Create two small datasets with similar range or SD but one unimodal and one bimodal. Explain why the spread measure alone does not describe clustering.
230. Practice C — choose range, IQR or SD | 练习 C
- You need the full observed minimum-to-maximum span.
- You need a robust spread of the middle half in skewed data.
- You need a mean-based spread measure in a roughly symmetric technical model.
- You need an easily explained interval for general readers showing the central 80%.
231. Practice D — percentile interpretation | 练习 D
Explain the 90th percentile of exam score, latency and income without saying “90% of the maximum” or “90% correct.” Then identify the reference population for each.
232. Practice E — skew direction | 练习 E
- Most values low, few very high.
- Most values high, few very low.
- Balanced tails around centre.
- Two separate peaks.
Choose right-skewed, left-skewed, roughly symmetric, bimodal and explain the tail/peak, not only the label.
233. Practice F — outlier or extreme? | 练习 F
- Largest valid income in a sample.
- Human height recorded as 18 metres.
- One genuine server request lasting 30 seconds.
- A point flagged by a formal IQR rule.
Decide whether extreme value, data error, outlier or another label is justified. Do not use outlier as a synonym for “I dislike this point.”
234. Practice G — Mandarin repair | 练习 G
- 两组平均值相同，但第二组离散程度更高。
- 分数集中在 70 附近。
- 收入分布明显右偏，并有长上尾。
- 中间 50% 的四分位距较窄。
- 两组分布高度重叠。
- 市场份额集中在少数公司手中。
- 这个离群值是真实事件，不应该自动删除。
- 短期价格波动属于 Lesson 036 的 volatility，不是这里的横截面 dispersion。
235. Practice H — centre + spread + shape | 练习 H
For one real dataset, write a three-sentence description: sentence 1 centre, sentence 2 spread, sentence 3 shape/tails. Avoid causal interpretation. This creates a complete distribution description without overloading one sentence.
236. Practice I — overlap and group averages | 练习 I
Two groups have means 70 and 75 with broad overlap. Write one safe group-level sentence and one unsafe deterministic individual-level sentence; explain why the latter exceeds the distribution evidence.
237. Practice J — visualisation audit | 练习 J
- Histogram with suspicious bin-dependent peaks.
- Box plot that may hide bimodality.
- Density plot whose smoothing changes peak count.
- Truncated axis making spread look large.
Write the caveat each display needs before you make a strong shape claim.
238. A 15-minute distribution session | 15 分钟训练
| Time | Task |
|---|---|
| 0–3 min | Choose one real distribution. |
| 3–6 min | Name centre + spread. |
| 6–9 min | Describe symmetry/skew and peaks. |
| 9–12 min | Inspect tails/outliers. |
| 12–15 min | Write one plain-English summary. |
239. A 30-minute spread laboratory | 30 分钟分布实验室
| Time | Task |
|---|---|
| 0–5 min | Create same-mean/different-spread datasets. |
| 5–10 min | Add one outlier; compare range/IQR/SD. |
| 10–15 min | Create a bimodal mixture. |
| 15–20 min | Compare subgroup vs pooled spread. |
| 20–25 min | Translate Mandarin distribution terms. |
| 25–30 min | Explain FENCE orally. |
240. Seven-day distribution challenge | 七天训练
| Day | Focus |
|---|---|
| 1 | distribution / spread / variability / dispersion |
| 2 | range / IQR / SD / percentiles |
| 3 | symmetry / skew / tails |
| 4 | clusters / modes / overlap |
| 5 | outliers / extremes / robustness |
| 6 | Mandarin transfer |
| 7 | full FENCE report |
241. Thirty-day distribution-control challenge | 30 天系统升级
- Collect 20 same-centre/different-spread examples.
- Compare range, IQR and SD in 15 datasets.
- Identify 15 skewed distributions and tail direction.
- Find 10 multimodal examples where one mean is misleading.
- Audit 10 outlier decisions for evidence.
- Describe 10 geographic/market concentrations.
- Translate 20 Mandarin spread/shape terms.
- Pair 20 centre statements with spread/shape companions.
- Explain 10 distributions orally under one minute.
242. Twelve-week progression | 12 周路线
| Weeks | Focus | Evidence of mastery |
|---|---|---|
| 1–2 | spread vocabulary | range/IQR/SD jobs distinct |
| 3–4 | shape / skew / tails | accurate shape description |
| 5–6 | clusters / multimodality / overlap | no average-only thinking |
| 7–8 | outliers / robustness | evidence-based handling |
| 9–10 | cross-domain concentration | statistical vs spatial/market control |
| 11–12 | centre-spread-shape synthesis | independent distribution reporting |
243. First weak-link diagnostic | 第一个弱点诊断
| Symptom | Weak link | Repair |
|---|---|---|
| I compare groups only by mean. | spread blindness | centre + spread pairs |
| I call all variability volatility. | time/cross-section confusion | 036 vs 039 contrast |
| I call bell-shaped normal. | technical overclaim | descriptive vs formal distribution |
| I delete every outlier. | data-quality reasoning | error vs genuine extreme audit |
| I think low SD means good. | evaluation/description confusion | target-centre examples |
| I think one SD describes tails fully. | shape/tail blindness | percentile/tail practice |
| I translate 波动 as variability everywhere. | bilingual domain control | time vs cross-section drill |
| I infer natural groups from visual clusters. | method inflation | descriptive vs clustering-method language |
244. How a strong tutor teaches distribution vocabulary | 高水平导师怎样教？
A weak lesson defines standard deviation. A strong lesson holds the mean constant and changes spread, then holds spread constant and changes shape, then inserts one extreme, then mixes two tight groups. The learner watches different measures respond to different structures and learns that no single statistic owns the whole distribution.
245. The learner should become their own distribution editor | 自我编辑
Before writing “the groups are similar,” ask whether only means match. Before writing “more variable,” state the spread measure or evidence. Before writing “normal,” ask whether you mean ordinary or normally distributed. Before removing outliers, ask what they are. Before calling a distribution bimodal, check whether the graphing method creates the peaks.
246. Reading harvest: extract distribution bundles | 阅读提取
Collect phrases such as tightly clustered around, widely dispersed, strongly right-skewed, long upper tail, interquartile range, 95th percentile, substantial overlap, geographically concentrated. Store the whole bundle, because modifiers/prepositions carry the shape.
247. Listening harvest: hear shape contrasts | 听力提取
Presenters often contrast “same average, wider spread,” “median stable, tail worsened,” “overall variability driven by between-group differences.” Capture the contrast structure, not only the statistic names.
248. Writing activation: centre → spread → shape → caveat | 写作激活
Use a four-sentence architecture: central location; spread measure; shape/clusters/tails; methodological caveat. This produces a rigorous distribution paragraph without hiding everything inside an “average”.
249. Speaking activation: 60-second distribution explanation | 口语激活
Explain one chart using FENCE in one minute. A listener should be able to tell where values centre, how wide the spread is, whether the distribution is skewed/multimodal, and whether any tail/outlier matters.
250. Paraphrase test: preserve shape | 改述测试
Original: “Scores were strongly right-skewed with a long upper tail.” Unsafe: “Scores varied widely.” The paraphrase deletes direction and tail shape. Safe: “Most scores lay toward the lower end, with a smaller number of much higher scores extending the upper tail.”
251. Summary test: keep the feature that changes interpretation | 摘要测试
If mean and median differ because of skew, keep skew/centre relation. If p99 drives user pain, keep the tail. If two clusters make mean unrepresentative, keep multimodality. Summary compression should remove detail, not the reason the average is misleading.
252. Canonical ownership boundary | 本课所有权边界
Lesson 039 owns distribution arrangement: spread, dispersion, variability, concentration, range, IQR, variance, standard deviation, skew, symmetry, tails, outliers, clustering, multimodality and overlap. Lesson 038 owns centre/typical values; Lesson 036 owns change/volatility over time; Lesson 037 owns quantitative denominators; Lesson 040 owns frequency/prevalence/incidence.
253. Recommended reference floor | 推荐参考资源
- NIST/SEMATECH — Location, Spread and Shape — distribution triad.
- OpenStax — Measures of Spread — range, variance, standard deviation and variability.
- NIST/SEMATECH — Measures of Scale — spread measures and robustness.
- Cambridge Dictionary — distribution — statistical and spatial senses.
- Cambridge Dictionary — dispersion — spread/scattering senses.
- Cambridge Dictionary — spread — general spread language.
254. SEO language map | 本课关键词范围
This lesson naturally serves distribution vocabulary English, spread vs variability, dispersion English, range IQR standard deviation, skewness English, right skew left skew, tails outliers English, bimodal multimodal distribution, concentration vs dispersion, percentile vocabulary, same mean different distribution, standard deviation meaning, C1 C2 statistics vocabulary, 英语分布词汇, 离散程度英语, 方差标准差英语, 极差四分位距英语, 偏态英语, 离群值英语, 双峰分布英语 and 中文母语高级英语.
255. Your assignment | 本课作业
Choose one authentic distribution and write a 350-word report using the full FENCE. Include at least one centre, one spread measure, one shape description and one tail/cluster observation. Then translate the report into Mandarin and back, checking that temporal volatility has not replaced cross-sectional variability.
256. Why this owner earns longform depth | 为什么这一课需要长文深度？
The reader job is not memorising SD and IQR. It is learning why the same average can hide consistency, inequality, polarisation, multimodality, tails, geographic concentration or subgroup structure; why robust measures deliberately ignore extremes; why outliers are not automatically errors; and why time volatility differs from cross-sectional spread. Those are separate reasoning failures and require a full distribution owner.
257. Final rules | 最后的规则
The average tells you where the centre is. The distribution tells you what surrounds it.
平均值告诉你中心在哪里；分布告诉你中心周围到底长什么样。
Never call two groups similar just because their averages match.
不要因为两个群体的平均值一样，就说它们相似。
Series: EDKS-ADV-VOC-ZH · Lesson 039.
