To translate mean, median, mode, percentiles and quartiles accurately, preserve which statistical summary the source actually uses. A mean is not simply another word for a median. The 90th percentile is not 90% of the maximum value. A quartile boundary is not automatically one quarter of the measurement scale. These statistics describe different features of a distribution, so replacing one familiar term with another can change the evidence while leaving the numbers looking reasonable.
This guide explains how to translate average, arithmetic mean, median, mode, percentile, quartile, interquartile range and ranked statistical summaries without changing subgroup definitions, weighting, missing-data rules or direction of comparison. It is designed for educational, research, assessment, business and public-information texts where readers need a faithful summary of a dataset rather than a looser conversational paraphrase.
The reliable method is to reconstruct the statistic before rewriting the prose. Record the data group, the statistic name, the calculation rule or rank interpretation, any weighting, and any exclusions. Then translate the sentence. This makes it much harder to turn “median household value” into “average household value,” confuse a percentile rank with a percentage score, or treat a quartile as though every group contains exactly the same range of values.
A fifty-second orientation
NIST describes the mean as the sum of data values divided by the number of values and the median as the central ordered value, with the usual averaging rule for an even number of observations. NIST’s percentile guidance treats percentiles as positions in ordered data, while its box-plot guidance identifies the lower quartile with the 25th percentile, the median with the 50th percentile and the upper quartile with the 75th percentile.
Useful references include NIST’s measures of location, NIST’s percentile guidance and NIST’s box-plot explanation. These sources also reveal an important translation lesson: a statistical word identifies a procedure or position, not merely a general impression of “typical” or “high.”
Before translating a statistical summary, ask: Which dataset? Which statistic? Which weighting or subgroup? Which percentile convention, if the exact numerical boundary matters? What did the source do with missing values? Those questions define the measure more securely than the number alone.
1. “Average” can be ambiguous
In everyday language, average can mean a general typical value. In technical writing, it often means arithmetic mean, but not always. A source may use average loosely while a table specifically labels mean, median or another summary. The translation should follow the source’s technical definition rather than assume every occurrence is interchangeable.
Suppose a fictional dataset is 2, 3, 4, 5 and 16. The arithmetic mean is 6, while the median is 4. Calling both “the average” hides a meaningful difference caused by the high value 16. If the source intentionally contrasts mean and median, the target must preserve separate terms.
When the source uses average without definition, inspect surrounding formulas, tables or methodological notes. If uncertainty remains and the distinction affects interpretation, keep the general wording or raise a source query rather than silently selecting mean because it is common in your own usage.
2. The arithmetic mean depends on every included value
For the fictional values 4, 6, 8 and 10, the arithmetic mean is (4 + 6 + 8 + 10) / 4 = 7. Every included observation contributes to the sum and the count. Adding a fifth value changes both numerator and denominator.
A translation that changes the population label can therefore change the meaning of the mean without changing the reported number. “Mean score among respondents” is not necessarily the same as “mean score among everyone invited.” Non-respondents may not appear in the calculation at all.
Keep the analysis population attached to the statistic. A short heading such as “Mean” may depend on a footnote saying “valid responses only.” If the translation moves the table but drops the footnote, the statistic becomes less defined even though its number survives.
3. The median depends on order rather than the sum
For the ordered fictional values 2, 3, 9, 20 and 100, the median is 9 because it is the middle observation. The arithmetic mean is 26.8. The large value 100 pulls the mean upward but does not move the median beyond the central position.
This makes median especially informative in some skewed distributions, but the translator should not add evaluative claims such as “more accurate” or “better” unless the source does. Mean and median answer different summary questions.
Translate median with a term that preserves its ordered-position meaning. If the target phrase merely means “middle-sized on average” in a vague sense, the statistical definition may be lost. Technical vocabulary should identify the measure, not just its everyday intuition.
4. Even-sized datasets need the median rule preserved
For four ordered values 2, 4, 8 and 10, the conventional sample median described by NIST is the average of the two middle values: (4 + 8) / 2 = 6. Six need not appear as an observed value in the dataset.
A target explanation saying “the median is always the middle observed value” would therefore be incomplete. In an even-sized dataset under this common convention, the median lies between the two central observations and may be a value that no individual observation has.
Preserve the source’s actual convention if it differs. Statistical software and methodological contexts can use defined procedures. Translation should not overwrite a specified calculation rule with a simplified classroom rule unless the assignment intentionally adapts the material for beginners.
5. Mode is about frequency, not central position
The mode is commonly the most frequently occurring value or category. In the fictional list 1, 2, 2, 2, 5 and 9, the mode is 2. The median is also 2 in this case, but that coincidence does not make the definitions identical.
A dataset can have more than one mode or no unique mode. For the values 1, 1, 2, 2, 3, two values share the highest frequency. A translation that insists on “the mode” as one mandatory value may misrepresent a multimodal dataset.
Mode also applies naturally to categorical data where mean is meaningless. If shirt sizes occur as S, M, M, L, the modal category is M. Do not force every “most common category” statement into a numerical-average vocabulary simply because nearby sections discuss statistics.
6. Mean, median and mode can coincide
In a symmetrical dataset they may be equal or close, but this is not guaranteed. Their agreement is a property of the data, not a rule that lets a translator substitute terms.
For the values 1, 2, 3, 4, 5, the mean and median are both 3, while there is no unique mode because every value appears once. Even in a very simple dataset, the three summary concepts do not collapse into one.
A terminology check should therefore focus on the calculation or definition, not the reported number. If two measures happen to produce the same value, keep their labels distinct so readers understand how each was obtained.
7. Weighted mean is not the same as ordinary mean
Suppose a fictional summary combines two groups. Group A has a mean of 80 across ten observations; group B has a mean of 60 across ninety observations. The simple average of group means is 70, but the overall mean across all observations is (80 × 10 + 60 × 90) / 100 = 62.
The groups do not contribute equally because their sizes differ. A target sentence that translates “weighted mean” as “mean” may hide the weighting structure. This is especially risky in surveys, school results, economics and meta-analysis.
Preserve the weighting basis when the source states it: population weights, sampling weights, credits, exposure time or another factor. If the weighting method is unclear, do not invent one from the final number.
8. An average of averages may not equal the overall mean
The previous example shows why averaging group averages can be misleading when group sizes differ. This matters in translation because a phrase such as “average across schools” may mean each school contributes equally, while “average across students” weights schools by student count.
Those are not stylistic variants. They define different estimands. A translator should keep the unit of averaging visible: people, schools, days, machines, branches or another entity.
A useful source query is: “Is this the mean of school-level means or the mean across all students?” That question identifies the missing weighting decision precisely. The answer belongs in the translation glossary because later references to average may depend on it.
9. Percentiles describe ordered position, not percent of the maximum
The 90th percentile is a value associated with a high position in the ordered distribution. It does not mean 90% of the maximum possible value. A student can be at the 90th percentile with a raw score that is far below 90% of the test maximum.
NIST describes a percentile through the proportion of measurements falling below and above a value. Exact computational conventions can differ among software and methodologies, especially for finite samples, but the ranking idea remains central.
Translate percentile with a rank-position concept, not a percentage-score concept. This distinction is crucial in education, health, performance dashboards and any context where readers may already be familiar with ordinary percentages.
10. Percentile rank and percentile value are related but not identical phrases
A percentile value is a measurement threshold associated with a chosen rank position. A percentile rank often expresses the percentage of the reference distribution at or below a particular score under a stated convention. The direction of mapping matters.
For example, “the 75th-percentile score was 82” and “a score of 82 had a percentile rank of 75” may align in a simple presentation, but the grammatical subject changes. One sentence starts with the percentile boundary; the other starts with an observed score.
A translator should preserve which object is being described. Reversing the relationship can make a report sound as though every score labelled 75 corresponds to a value of 82, or that percentile is a raw-score unit. Keep score and rank on separate conceptual axes.
11. The median is the 50th percentile under common definitions
NIST identifies the 50th percentile with the median. This provides a useful cross-check when a source reports both. If a table’s median and P50 differ substantially, investigate whether the source uses different populations, interpolation conventions or definitions.
Do not silently force the numbers to match. The discrepancy might be legitimate because different columns use different filters or software methods. A translator’s job is to preserve the source and surface meaningful contradictions, not repair them invisibly.
When an explanatory article uses P50 as shorthand, define it at first use for a general audience. The symbol or abbreviation should remain attached to the statistical concept rather than be mistaken for a model number or score code.
12. Quartiles divide ordered data conceptually, not the measurement scale itself
The lower quartile is commonly associated with the 25th percentile, the median with the 50th, and the upper quartile with the 75th. These are positions in ordered data. They do not mean the numerical distance from minimum to maximum is divided into four equal intervals.
Consider the fictional values 1, 2, 3, 4, 100. Most data lie near the low end, so quartile boundaries can be clustered there even though the numerical range extends to 100. Equal proportions of observations do not imply equal widths on the value scale.
A target phrase such as “the first quarter of the measurement range” can therefore be misleading if the source means lower quartile. Translate quartile as a distributional rank concept, not a geometric division of the axis.
13. Different quartile algorithms can produce different boundaries in small samples
Statistical software packages can use different interpolation or ranking conventions for sample quantiles. In large datasets these differences may be small, but in small samples they can produce visibly different quartile values.
Therefore, when a source gives exact quartile numbers, preserve the stated software or method if it is part of the methodology. Do not recalculate with your preferred spreadsheet and overwrite the source because your Q1 differs slightly.
A translation can explain the concept without pretending there is only one computational convention in every context. The method is part of the data provenance when exact boundaries matter.
14. Interquartile range is a difference between quartiles
The interquartile range, or IQR, is commonly Q3 minus Q1. It measures the spread of the middle half of the ordered data under the chosen quartile convention. It is not the same as the full range from minimum to maximum.
If Q1 is 20 and Q3 is 50, the IQR is 30. The middle 50% of observations lies between the quartile boundaries according to the source method. A target that labels 30 as “the range” without qualification can suggest the full dataset runs only from 20 to 50.
Keep the modifier interquartile or its correct technical equivalent. Shortening a statistical term to a familiar everyday noun can remove the calculation rule that gives the number meaning.
15. Box plots summarise position and spread; their components should keep their roles
NIST’s box-plot guidance places the median within a box spanning the lower and upper quartiles. Variations may use whiskers and outlier rules differently. A translation should preserve what each visual element represents in the specific figure.
Do not call the box “the full range” if it represents the interquartile interval. Do not call a whisker endpoint the maximum if the plotting convention instead stops at a non-outlier boundary. The caption, method and legend define the figure.
This is where translating charts and statistical prose meet. The labels around a box plot are not separate from the data summary. A mistranslated legend can alter how readers interpret every line in the figure even when the plotted geometry is unchanged.
16. Missing data rules can change means and medians
Suppose five fictional values are 2, 4, missing, 8 and 10. If the source calculates summaries using the four observed values, the mean is 6 and the median is 6 under the usual even-sample rule. If missing is encoded incorrectly as zero, the mean becomes 4.8 and the median becomes 4.
A translation should preserve terms such as valid cases, available-case analysis, missing, not applicable and zero. These labels can control whether a value participates in the statistic.
Do not “clean up” an empty field into 0 unless the source defines zero as the intended value. The distinction between absence of data and a measured zero is both statistical and semantic.
17. Subgroups need separate statistical labels
A report may give overall median, median for group A and median for group B. The numbers can be identical or different. Their meaning depends on which observations were included in each calculation.
A target that drops “among group A” may make a subgroup statistic look like the overall statistic. This error is particularly easy in headings where space is limited and group labels repeat.
Use layout-aware QA: read each number together with its row and column headings. A statistic belongs to the intersection of those labels. Copying a value correctly is not enough if its subgroup context is lost.
18. Time averages and cross-sectional averages are different
“Average daily value for each person” can differ from “average across all person-days.” The first may give each person equal weight after computing an individual average; the second may give more weight to people with more recorded days.
This resembles the average-of-averages problem but appears frequently in longitudinal data. A translator should preserve whether the unit being averaged is a person, day, transaction, site or another repeated observation.
If the source says “mean per participant” or “mean per observation,” keep that unit visible. A small prepositional phrase can determine the weighting structure of the statistic.
19. Geometric and harmonic means are not arithmetic means
Some technical sources use geometric or harmonic means for particular kinds of data and rates. A translation that simplifies all of them to average can remove the calculation method. Even if the article does not explain the full formula, the statistic name should remain distinct.
For positive values 2 and 8, the arithmetic mean is 5, while the geometric mean is 4. The numbers differ because the operation differs. A reader who sees only “average = 4” cannot reconstruct which summary the source used.
Translate specialist mean types at concept level and preserve any formula or methodological definition. Do not let the everyday familiarity of average erase the statistical procedure.
20. Percentile comparisons require the same reference distribution
A 75th percentile in one reference group is not automatically equivalent to the 75th percentile in another group. The percentile positions may match while the underlying score thresholds differ.
Therefore a translation should preserve reference-population labels such as national sample, age group, grade level, year, region or cohort. Dropping the reference group can make a percentile appear universal when it is distribution-specific.
If the source compares percentiles across years, check whether the reference distribution itself changed. Translation should preserve the methodological caveat rather than imply direct comparability that the source did not claim.
21. A percentile is not a percentage-point change
Moving from the 60th percentile to the 70th percentile is a change of ten percentile ranks or points on that rank scale, but it does not necessarily mean a ten-percentage-point increase in the underlying measured outcome.
The raw-score difference between those percentile positions depends on the distribution. In a dense central region it may be small; in a sparse tail it may be larger. A translator should not convert percentile movement directly into percentage-score movement.
Preserve phrases such as percentile rank, percentage score and percentage points as separate terms. They occupy different mathematical layers even when all contain the word percent.
22. Worked case: skewed income-style data
Consider the fictional values 30, 32, 34, 36 and 168. The median is 34. The arithmetic mean is 60. The large final value pulls the mean far above the middle observation.
A flawed target might replace “median value was 34” with “average value was 34” because median sounds unfamiliar to general readers. That rewrite removes the very distinction the source may be using to describe skewness.
A better translation keeps median and, if needed, explains it briefly as the middle value in the ordered data. Clarity should come from explanation, not from replacing a technical statistic with a different one.
23. Worked case: weighted versus unweighted school averages
Suppose fictional School A has ten students with mean score 90, while School B has ninety students with mean score 60. The unweighted mean of the two school means is 75. The student-weighted overall mean is 63.
If the source says “average school mean,” 75 may be appropriate. If it says “average student score across both schools,” 63 is the relevant value. Translation must preserve the unit being averaged.
A target phrase such as “overall average” can be dangerously vague. Where the source is explicit, keep the explicit wording. If the source itself is ambiguous and the distinction matters, ask whether schools or students receive equal weight.
24. Worked case: percentile and raw score
A fictional report says: “A raw score of 72 corresponded to the 80th percentile in the reference cohort.” This means the score’s position is described relative to that cohort under the report’s percentile method.
A flawed translation might say “the learner scored 80% with a raw score of 72.” That invents a percentage score and discards the reference distribution. The number 80 is a percentile rank, not necessarily a percentage correct.
The repair keeps all three objects distinct: raw score 72, percentile rank 80, reference cohort. This is a good example of why a number must be translated together with its statistical label.
25. Worked case: quartiles and box-plot language
Imagine a fictional dataset with Q1 = 12, median = 18 and Q3 = 30. The interquartile range is 18. The middle half of the ordered data lies between the quartile boundaries under the source convention.
A target caption saying “the data range is 18” would be wrong if 18 is the IQR rather than the full max-minus-min range. Likewise, “the box shows all values from 12 to 30” overstates what the box represents.
Keep the terminology aligned with the visual: lower quartile, median, upper quartile and interquartile range. A figure can be visually correct and still be semantically mistranslated through its caption.
26. Practice clinic with explained answers
Practice one. Values are 2, 4, 6, 8. Mean = 5. Median = 5 under the usual even-sample rule. The two numbers match, but the calculations differ.
Practice two. Values are 1, 2, 2, 5, 20. Mean = 6. Median = 2. Mode = 2. A translation must not call all three “average 2.”
Practice three. Values are 1, 1, 2, 2, 3. There are two modes, 1 and 2, under the ordinary frequency definition. Do not force a unique mode.
Practice four. A score is at the 90th percentile. This does not mean the score equals 90% of the maximum. Preserve percentile-rank language.
Practice five. Q1 = 10 and Q3 = 25. IQR = 15. Do not label 15 as the full range unless minimum and maximum support that statement separately.
Practice six. Group A mean = 80 for ten observations; Group B mean = 60 for ninety observations. Overall observation-weighted mean = 62, not the simple mean-of-means value 70.
Practice seven. A table says “mean among valid responses.” The target should not simplify this to “mean among all respondents” if invalid or missing responses were excluded.
Practice eight. P50 and median differ in a source table. Do not silently repair the discrepancy. Check populations, methods and filters first.
Practice nine. A chart box spans Q1 to Q3. That box covers the interquartile interval, not necessarily the complete minimum-to-maximum range.
Practice ten. A source reports “geometric mean 4.” Do not translate the label as ordinary arithmetic mean. The statistic type is part of the claim.
Practice eleven. A raw score of 60 maps to the 75th percentile in one cohort and the 65th in another. Percentile meaning depends on the reference distribution.
Practice twelve. The source says median waiting time was 12 minutes. A target saying average waiting time was 12 minutes is not faithful unless the source explicitly uses average to mean median.
27. A release checklist for statistical summaries
Label every statistic explicitly: arithmetic mean, weighted mean, median, mode, percentile, quartile or IQR. Identify the analysis population. Check missing-data rules. For averages across groups, identify the weighting unit. For percentile statements, identify the reference distribution and whether the number is a raw score or rank.
Then review the target language for false synonyms. Typical, average, middle, common and central can be useful explanatory words, but they should not replace technical terms in a way that changes the calculation. Exact vocabulary is especially important when several summaries appear together.
Finally, recompute one representative example from the source where possible. A successful check does not prove every statistic is right, but it can reveal reversed labels, wrong denominators, incorrect weighting or a percentile translated as a percentage.
28. Where this article sits in the eduKate translation system
This specialist guide extends Translate | Names, Numbers, Dates and Units, the new Fractions, Ratios and Proportions guide and the Probability, Odds and Risk Ratios guide. The broad owner remains Master Art of Translation.
For terminology depth, the Vocabulary Learning Hub helps separate near-neighbours such as mean, median, percentile and percentage. How English Works supports the grammatical side: which noun a qualifier modifies, which group a comparison refers to and how scope survives sentence reordering.
The governing rule is to preserve the statistic before preserving the sentence. Once you know exactly what was calculated, on which observations and under which ranking or weighting rule, you can write naturally in the target language. Without that reconstruction, a beautifully translated “average” may still be the wrong statistic.