To translate nominal, ordinal, interval and ratio data accurately, you must preserve what the numbers are allowed to mean. A category code such as “1 = red, 2 = blue, 3 = green” does not become a ranked scale merely because the labels are numbers. A satisfaction scale from “strongly disagree” to “strongly agree” has order, but the gap between neighbouring categories is not automatically a fixed physical distance. A Celsius temperature has equal intervals, yet 20°C is not meaningfully “twice as hot” as 10°C. Translation errors begin when familiar-looking numerals are treated as interchangeable quantities.
This guide explains how to translate nominal data, ordinal scales, interval measurements and ratio scales for research papers, surveys, dashboards, data dictionaries, educational materials, questionnaires, technical reports and statistical output. It focuses on high-intent translation questions such as whether category codes should be translated, whether ranks can be averaged, whether a zero means “none,” how Likert responses should be labelled, and how a change of wording can accidentally change the level of measurement. The aim is not to force every dataset into one philosophical taxonomy, but to protect the relationships the source actually encodes.
The practical method is to ask four questions before translating a variable: What makes two values the same or different? Is there a meaningful order? Are equal numerical differences intended to represent equal differences in the measured property? Does zero represent absence of the property so that ratios such as “twice as much” are meaningful? NIST’s discussion of measurement scales describes the familiar nominal, ordinal, interval and ratio progression, while also noting that real measurement practice can require care beyond a simple four-box classification. In translation, these questions work as a semantic safety check.
A sixty-second orientation: measurement scale controls what comparisons mean
Nominal data classify. Ordinal data classify and order. Interval data add meaningful, comparable differences. Ratio data add a meaningful zero that supports ratios. This hierarchy is useful because each step permits more kinds of statements. You can count categories on a nominal scale. You can say one ordinal category ranks above another. You can compare differences on an interval scale. On a ratio scale, statements such as “twice the mass” can be meaningful because zero represents no amount of the measured quantity.
The danger is to translate the numeral rather than the relation. If an ordinal survey code changes from 1 to 5, the target must preserve the order and labels; it must not suggest that category 4 contains twice as much of a property as category 2 unless the source defines such a quantity. If a temperature moves from 10°C to 20°C, the difference of 10 Celsius degrees is meaningful, but the ratio 2:1 is not a meaningful ratio of thermodynamic temperature on the Celsius scale. The wording around the number determines which comparisons are legitimate.
1. Nominal data are names or classes, even when they are coded with numbers
A nominal variable answers a classification question: which category does this observation belong to? Examples can include language code, product type, blood group, document status or experimental condition label. The categories may be represented by words, abbreviations or numbers. The representation does not create arithmetic meaning.
Suppose a fictional dataset uses 1 for online, 2 for classroom and 3 for hybrid instruction. The mean of those codes is not automatically a meaningful “average instruction type.” A translator should preserve the mapping rather than treating the numbers as quantities. If the target changes the order of categories for alphabetical reasons, the code-value mapping must still remain consistent with the underlying data.
This becomes important when software exports only numeric values while a data dictionary supplies the labels. Translate the dictionary and keep the codebook relationship intact. Do not translate code 2 into whichever target category appears second in a newly sorted list. Codes identify categories; list position is a presentation choice.
2. Numeric category codes are identifiers, not measurements
Numbers often look mathematical even when they function like names. Postal codes, classroom IDs, participant numbers and categorical database codes may contain digits but are nominal identifiers. Adding, averaging or ranking them does not produce a meaningful property unless the system explicitly defines such operations.
A translation system can accidentally destroy this distinction by localising digits, removing leading zeros or reformatting codes as ordinary numbers. A participant code 007 may not be equivalent to participant 7 if the identifier is fixed-width. A status code 01 may be a protected system value even if the human label beside it is translated.
Keep identifier strings separate from measured quantities in your working notes. Ask whether the value can be used in arithmetic or only for lookup and grouping. This connects to the broader eduKateSG translation principle that identifiers should be treated as protected objects rather than ordinary vocabulary.
3. Nominal categories can be hierarchical without becoming ordinal
Categories can have structure without having a single rank order. A document may classify a device as electrical, then battery-powered, then rechargeable. That hierarchy does not mean rechargeable is “greater than” electrical. It expresses nested membership. NIST notes that nominal properties can themselves be hierarchical.
Translation should therefore preserve parent-child relationships in taxonomies. A target-language editor who flattens a hierarchy into one alphabetical list can make later filters or summaries ambiguous. Likewise, adding numbered bullets for readability should not imply that category 1 has lower status than category 2.
Where a classification tree matters, translate the headings and node labels while preserving the tree structure. If the source hierarchy has official codes, keep them. Human-readable wording can change; the category identity and parentage should not.
4. Ordinal data add order, but not necessarily equal spacing
Ordinal scales allow statements such as higher, lower, earlier, later, better ranked or more severe according to the scale definition. The crucial limitation is that the gap between adjacent levels need not be equal. Category 4 can rank above category 3 without being “one unit more” in a physical sense.
Consider a fictional four-level difficulty scale: easy, moderate, difficult, very difficult. Translating the terms requires preserving order. If a target word chosen for moderate sounds stronger than the word chosen for difficult, the scale becomes internally inconsistent. Vocabulary precision is therefore part of measurement preservation.
Before release, read the translated labels as a sequence. Ask whether a native reader would order them in the same way as the source. Direct dictionary equivalents can fail this test because words carry different intensity, politeness and connotation across languages.
5. Likert response categories are a translation problem before they are a statistical problem
Survey scales such as strongly disagree, disagree, neither agree nor disagree, agree and strongly agree are ordinal response categories at the item level. The translation must preserve direction, balance and intensity. A weak translation of the middle category can push respondents toward one side; a target word for strongly may sound extreme compared with the source.
Do not translate only the endpoint words. Read the entire response set together. If “agree” becomes a phrase closer to “somewhat agree” while “strongly agree” stays very strong, the intervals in perceived intensity may change even if the order remains. The instrument’s measurement properties can be affected by wording.
Where a validated target-language instrument exists, use its established wording rather than inventing a fresh translation. If you are creating a new translation, document the response mapping and any cognitive-testing or back-translation process required by the project. A survey scale is not ordinary decorative copy.
6. “Neutral,” “neither,” “not sure” and “not applicable” are not synonyms
A middle response in an agreement scale may mean neither agree nor disagree. “Not sure” may mean the respondent cannot decide. “Not applicable” may mean the question does not apply. “Prefer not to answer” expresses another status entirely. Collapsing these into one vague neutral category can change both measurement and missing-data interpretation.
If the source distinguishes substantive midpoint from nonresponse options, keep them distinct in the target. They may be analysed differently. A midpoint can be part of the ordinal scale, while not applicable may be excluded from the scored variable and stored as a special code.
This is where measurement-scale translation connects directly to missing-data translation. The visible survey form and the dataset must agree about what each response means. A translator should therefore check the codebook, not only the questionnaire text, when the project includes both.
7. Ranks preserve order but not the size of the gaps
First, second and third place are ordinal positions. The difference between first and second does not tell you how much better the first performance was. The winning time might be milliseconds ahead or minutes ahead. Rank compresses magnitude into order.
A translation should keep rank language distinct from score language. “Rank improved from tenth to fifth” is not the same as “score improved by fifty percent.” Likewise, a five-position change does not reveal a fixed change in the underlying ability or performance measure.
If a table includes both score and rank, preserve both headings. Do not convert one into the other to reduce columns. They answer different questions: score indicates a measured or calculated quantity under its scale; rank indicates relative position within a comparison set.
8. Interval scales make differences meaningful
An interval scale supports meaningful differences because equal numerical intervals represent equal differences in the measured property under the scale definition. Calendar years are a useful conceptual example: the difference between 2000 and 2010 is ten years, and the difference between 2010 and 2020 is also ten years. But year 2000 is not “twice” year 1000 in a meaningful temporal-amount sense.
Celsius temperature provides another familiar example. A change from 10°C to 20°C is a 10-degree increase, and 20°C to 30°C is another 10-degree increase. However, the zero point is conventional rather than an absence of thermodynamic temperature, so ratio statements based directly on Celsius values are misleading.
Translation should therefore preserve whether the source describes a difference, level or ratio. Words such as twice, half, double and percentage increase can be invalid on an interval scale even when subtraction is meaningful. Keep the source’s comparison type attached to the correct scale.
9. Temperature translation shows why zero matters
Imagine a fictional report saying one chamber operated at 10°C and another at 20°C. It is correct to say the second was 10 Celsius degrees warmer under the stated readings. It is not correct to say it had twice the thermodynamic temperature. If the temperatures are converted to kelvins, the ratio is completely different because 10°C and 20°C correspond to approximately 283.15 K and 293.15 K.
A translation that introduces “twice as hot” because the displayed number doubled therefore changes the measurement logic. Conversely, converting Celsius to Fahrenheit requires an affine transformation with both multiplication and an offset. The numerical zero moves, but equal temperature intervals are preserved up to the scale factor.
This example teaches a general rule: never infer ratio meaning from the visual appearance of two numbers. Ask what zero means on that scale and whether the source property supports meaningful multiplicative comparison.
10. Ratio scales add a meaningful zero
A ratio scale supports meaningful ratios because zero represents absence of the measured quantity under the scale. Mass, elapsed duration, absolute temperature in kelvins and many count variables can behave this way in ordinary measurement contexts. Twenty kilograms is twice ten kilograms; twenty seconds is twice ten seconds.
This does not mean every variable containing zero is a ratio scale. A score of zero on a test may mean no points earned, not no ability. An account balance of zero can be a reference point between positive and negative amounts rather than absence of every relevant financial property. The meaning of zero belongs to the construct and scale definition.
Translate zero together with the variable definition. If the source says “zero indicates no events observed,” keep that. If zero is an arbitrary coded category, do not reinterpret it as none. A single digit can represent absence, reference, code or missingness depending on context.
11. “Twice as much” is a measurement claim, not merely a stylistic phrase
Ratio language appears frequently in summaries: twice as long, half the mass, three times the count, 20% more. Those phrases assume that multiplicative comparison is meaningful for the quantity. A translator should not introduce them merely to make prose more vivid.
For a ratio-scale quantity, the arithmetic still needs the correct baseline. A length rising from 10 m to 20 m doubles. A value “200% of baseline” equals twice baseline, while “200% more than baseline” equals three times baseline. Measurement level permits the ratio; language determines which ratio is claimed.
The existing eduKateSG guide to percentages and percentage points develops that baseline distinction in detail. Here, the additional question is whether multiplicative comparison is meaningful for the scale at all. Both conditions must be satisfied.
12. Codes can change language without changing category identity
A multilingual database may store code EN for English, FR for French and JA for Japanese while the user interface displays translated language names. The labels can be localised; the protected code values may remain unchanged. Similar structures occur with disease classifications, product categories, school subjects and administrative regions.
Do not translate an underlying code merely because its mnemonic resembles an English abbreviation. It may be part of an external standard, API or reference table. Human-readable text and machine identifiers belong to different layers.
When codes are local rather than standardised, document whether the target system expects the original code or a newly mapped code. A translation team should never infer that requirement from appearance alone. Changing labels is linguistic; remapping identifiers is a data migration task.
13. Reversing an ordinal scale can silently reverse the meaning of a result
Some questionnaires code 1 as strongly agree and 5 as strongly disagree. Others use the opposite direction. If the target form reorders the visible options but the scoring code remains unchanged, higher numerical values can switch from positive to negative meaning.
This error may survive ordinary proofreading because every label is translated correctly in isolation. The problem is the relationship between label order and stored code. A translation quality check must compare the form, codebook and scoring specification together.
Use an anchor item in your review notes: “1 = strongly disagree; 5 = strongly agree,” for example. Verify the first, middle and last positions after localisation. If the source uses reverse-scored items, preserve the item wording and scoring rule separately rather than manually flipping values during translation.
14. Composite scores can behave differently from individual ordinal items
Researchers often combine multiple ordinal questionnaire items into a summed or averaged score. The statistical treatment of such composite scores depends on the instrument, model and analytic convention. A translator should not decide that issue from the presence of Likert-style response options alone.
The source may treat a validated multi-item scale as approximately continuous and report means, standard deviations and regression coefficients. Preserve the source’s stated analytic treatment. Do not “correct” the manuscript by replacing every mean with a median unless the task includes methodological review and the authors approve a change.
At the same time, avoid implying that one individual response category has equal physical spacing merely because the composite score is analysed quantitatively. Keep item-level labels, scoring rules and scale-level summaries conceptually separate.
15. Statistical summaries must fit the variable’s meaning
Frequency counts and proportions make sense for nominal categories. Medians and percentiles can be meaningful for ordinal data because order is preserved. Means and standard deviations are commonly used when equal-interval interpretation is defensible. Ratio-scale data additionally support meaningful coefficient-like ratios and multiplicative summaries in suitable contexts.
Translation should preserve the statistic the source actually reports, but measurement-scale knowledge helps detect suspicious combinations. If a report says “average country code = 2.7,” the number may be a coding artefact rather than a meaningful summary. Do not automatically replace it, but flag it when your task includes technical review.
The key distinction is between translating a report faithfully and validating the analysis. A translator can recognise that a statement appears conceptually odd without silently rewriting the methodology. Use subject knowledge to ask better questions, not to hide source problems.
16. A histogram, bar chart and ordered category chart imply different structures
Visualisation can reveal the intended measurement level. A bar chart with separated bars often represents categories. A histogram uses adjacent bins to represent intervals on a quantitative axis. An ordered category chart may show ordinal levels. Translation should preserve axis labels and ordering so the graph continues to communicate the same structure.
Alphabetically sorting ordinal labels can destroy the intended sequence. Sorting “low, medium, high” as “high, low, medium” may be convenient for a localisation library but misleading for the reader. The data visualisation should use semantic order rather than target-language alphabet unless alphabetical order is itself the source logic.
Similarly, a nominal chart may legitimately be sorted by frequency. That does not make the category scale ordinal. Distinguish presentation order from measurement order. One can change dynamically; the other belongs to the variable definition.
17. Dates and calendar years illustrate interval-style reasoning
Calendar years support meaningful differences within a defined calendar. The interval from 2000 to 2010 is ten years. But ratios of year numbers are not meaningful in the same way. Saying year 2000 is twice year 1000 confuses the label of a position on a calendar scale with a duration.
This matters in translation when years are reformatted across eras or calendar systems. Converting between calendar labels is not simply multiplying a quantity. A change of era may move the numerical zero and origin while preserving chronological intervals according to specified conversion rules.
Keep calendar point and duration separate. “Ten years later” is an interval; “the year 2010” is a position. The same unit word year can serve both functions, so surrounding grammar matters.
18. Scores may look quantitative while representing constructed scales
Educational, psychological and performance scores are often constructed from items, transformations or models. A score of zero may not mean no ability, and a score of 80 may not represent twice the construct of a score of 40. The numerical scale needs its own interpretation.
Translation should preserve score, points, percentile, rank and proficiency level as distinct terms. A percentile is not a percentage score; a rank is not an interval score; a proficiency category is not automatically the same type of quantity as the underlying scaled score.
This is particularly important in exam and assessment material. A student at the 80th percentile is not necessarily someone who answered 80% of items correctly. The target should preserve the statistical interpretation supplied by the assessment system.
19. Percentiles are ordinal positions, not equal-distance units
Percentiles locate a value relative to a distribution. The difference between the 50th and 60th percentile does not necessarily represent the same raw-score difference as the gap between the 80th and 90th percentile. Percentile rank therefore behaves differently from a raw measurement scale.
A translation should not say a person at the 80th percentile has “twice the ability” of someone at the 40th percentile. The numbers refer to relative positions, not ratios of the underlying construct. Preserve words such as percentile, percentile rank and percent correctly.
The dedicated eduKateSG article on means, medians, percentiles and quartiles covers these distribution summaries in detail. Here, the key lesson is that numerical range alone does not determine measurement level.
20. Binary variables are nominal even when coded 0 and 1
A binary variable may be coded 0/1 for convenience: no/yes, absent/present, control/treatment or unsuccessful/successful. The codes support useful mathematical operations in models, but their substantive meaning remains categorical. Zero is often a reference category rather than a universal absence of the construct.
Suppose treatment=1 and control=0. Translating 0 as “none” would be wrong; the control group exists. Similarly, female=1 and male=0 in a particular legacy dataset would not make male an absence of sex. The code identifies categories within a data model.
When interpreting regression coefficients for binary variables, translate the category contrast explicitly if the source does. “Treatment versus control” is clearer than “one versus zero.” The numeric coding belongs to the software layer; the categories belong to the substantive meaning.
21. Measurement units do not automatically reveal measurement level
A variable with a unit often behaves quantitatively, but the unit alone does not settle every interpretive question. Temperature in Celsius has a unit yet lacks a true zero for ratio statements. A coded severity stage may have numbers but no physical unit. A count can have a meaningful zero but may be transformed or censored in analysis.
Use the variable definition, not typography, to determine meaning. Superscripts, percent signs and units help, but they are evidence rather than complete interpretation. This is the same discipline used throughout technical translation: symbols must be read in context.
A glossary can record both preferred term and scale meaning. For example: “severity grade — ordered category, larger number indicates greater severity; distances between grades not assumed equal.” Such notes prevent later translators from treating the field as ordinary arithmetic.
22. Transformations can change representation without changing the underlying ordering
A monotonic transformation can preserve order while changing distances. For example, taking a logarithm of positive ratio-scale data preserves which observations are larger but changes numerical differences and ratios on the transformed scale. The transformed variable must be labelled clearly.
A translation should distinguish “income” from “log income,” “concentration” from “log concentration,” and “raw score” from “standardised score.” The source may discuss the transformed variable because the model operates on that scale. Dropping log or standardised is not a stylistic simplification; it changes the quantity being analysed.
This leads naturally to the separate eduKateSG guide on logarithmic scales. Measurement-scale translation and transformation translation meet at the same principle: keep the representation contract visible.
23. A worked survey translation
Consider this fictional source instrument: “How confident are you that you can complete the task independently? 1 Not at all confident; 2 Slightly confident; 3 Moderately confident; 4 Very confident; 5 Extremely confident; 9 Not applicable.” The substantive response scale is ordinal from 1 to 5. Code 9 is not the highest confidence level; it is a special response outside the scale.
A poor translation might place “not applicable” after extremely confident without visual separation and later treat 9 as a valid high score. Another might translate moderately with a word stronger than very, reversing the intensity. A reliable translation preserves order and marks code 9 according to the dataset’s missing or special-value convention.
The review should compare the displayed questionnaire, codebook and scoring instructions. If a composite confidence score excludes code 9, the target data-processing description must say so. No amount of fluent wording can compensate for a broken mapping between response and stored value.
24. A worked temperature translation
Use this fictional sentence: “The chamber temperature increased from 10°C to 20°C, an increase of 10 Celsius degrees.” That is a meaningful interval statement. If the source then says “the absolute temperature increased from 283.15 K to 293.15 K,” the kelvin representation uses a ratio scale with absolute zero.
A faulty target might say “the temperature doubled from 10°C to 20°C.” That introduces a ratio interpretation unsupported by the Celsius scale. Another might write “increased by 100%” using the displayed Celsius values as the baseline. This too depends on a ratio operation that the scale does not support in the intended physical sense.
The correct translation preserves the interval comparison. If a ratio of absolute temperatures is relevant, the source should use or define the appropriate absolute scale. Do not manufacture that reinterpretation as an explanatory flourish.
25. A worked coding translation
Imagine a data dictionary with “Region: 1 North, 2 East, 3 South, 4 West.” The numbers identify nominal categories. Suppose a target language naturally lists directions in another order. You may reorder a human-readable legend only if the code stays attached to the correct region; otherwise code 1 could become East simply because East appears first alphabetically.
A second field says “Service level: 1 Basic, 2 Standard, 3 Premium.” This variable may have an intended order because the service tiers are ranked. Although both variables use 1,2,3-type codes, their measurement relationships differ. One is nominal direction category; the other is ordinal service tier.
The translator cannot determine measurement level from code format. Read definitions, labels and business rules. A good glossary entry can record “Region — nominal” and “Service level — ordered tier” so later analysts and translators do not infer the wrong comparisons.
26. Practice clinic: identify what the number is allowed to mean
Practice one. Product colour is coded 1 red, 2 blue, 3 green. Can you say green is three times red? No. The numbers are nominal codes. Translate the labels and preserve code identity.
Practice two. Satisfaction is coded 1 dissatisfied, 2 neutral, 3 satisfied. Is satisfied exactly one unit above neutral in the respondent’s psychological state? The order is defined; equal distance is not automatically established. Preserve the ordinal wording.
Practice three. Temperatures are 10°C and 20°C. Can you translate the second as twice the first? No. Celsius supports equal differences but not that direct ratio interpretation.
Practice four. Durations are 10 s and 20 s. Can the second be twice the first? Yes, under the ordinary duration scale. Preserve unit and baseline.
Practice five. A binary field uses 0 control, 1 treatment. Does zero mean no participant? No. It is a category code and reference level.
Practice six. A survey uses 9 for not applicable after a 1–5 response scale. Should 9 be treated as highest agreement? No. Preserve special-value status outside the substantive scale.
Practice seven. A chart sorts low, medium, high alphabetically and displays high, low, medium. Is the underlying ordinal scale changed? The data definition may not be, but the visual translation is misleading. Restore semantic order.
Practice eight. A test score is zero. Does that prove zero ability? Not necessarily. Translate the score according to the assessment definition rather than treating the numerical zero as absence of the latent trait.
27. Frequently asked translation questions
Should nominal codes be translated? Human-readable labels often should be; machine or standard codes often should not. Preserve the mapping between code and category.
Can ordinal categories have numbers? Yes. Numbers can represent ordered positions, but the arithmetic distance between codes may not represent equal distance in the underlying construct.
Is every survey scale ordinal? Individual ordered response categories often are, but composite scores and validated instruments may be analysed under additional assumptions. Preserve the source’s stated model rather than imposing a universal rule.
Why is Celsius interval rather than ratio? Equal differences are meaningful, but the zero point is not an absolute absence of thermodynamic temperature, so direct ratios of Celsius values are not meaningful in the usual physical interpretation.
Does zero always mean none? No. Zero can be an absolute zero, a reference point, a category code, a score value or a missing-value sentinel depending on the variable.
What is the strongest release check? For every variable, complete this sentence: “These values can validly be used to express ___, but not automatically ___.” If the target wording permits a comparison the source scale does not, revise it.
28. Connect this specialist guide to the eduKateSG translation architecture
This article is a specialist branch of Master Art of Translation, not a competing broad translation hub. It connects naturally to Translate | Mean, Median, Mode, Percentiles and Quartiles, Translate | Standard Deviation, Variance and Z-Scores, and Translate | Correlation, Regression Coefficients and R-Squared. For precise word relationships, use the protected Vocabulary Learning Hub; for comparison, negation, modifiers and reference, use How English Works.
A reliable measurement-scale translation does not ask only whether every label has a dictionary equivalent. It asks whether the target reader is allowed to make the same comparisons as the source reader—and no stronger ones. Preserve category identity, ordinal direction, equal-interval meaning, true-zero interpretation, special codes and any documented transformations. When those relationships survive, the translated dataset remains the same dataset in another language rather than a new measurement system created accidentally by wording.