To translate missing data, N/A, null, blank fields and zero values accurately, you must preserve the reason a value is absent. Zero means a measured or computed quantity equal to zero. Blank may mean nothing was entered, but it may also be a formatting choice. N/A may mean not applicable, not available, not assessed or something else defined by the source. Null is often a database state meaning no value is stored. These are not interchangeable labels, and a fluent translation that collapses them can silently change the dataset.
This guide explains how to translate missing values, unknown data, not applicable responses, nulls, blanks, zeros, suppressed values, withheld data, below-detection-limit results and special codes in spreadsheets, surveys, databases, research tables, dashboards, APIs and technical reports. It solves a common translation search intent that sits between language and data engineering: how do you localise what is not there without inventing a value, erasing the reason for absence, or turning a true zero into missingness?
The safest workflow is to separate three layers before translating: the semantic state, the machine representation and the human-readable label. The semantic state may be unknown, not applicable, not collected, withheld, structurally impossible or genuinely zero. The machine representation may be NULL, NaN, an empty field, −999, 99, a dot, a dash or a status code. The human-facing label may be “Not available” or another phrase. Translate the label only after you know which state it represents. OECD reporting guidance, for example, explicitly distinguishes missing data from zero values and uses status codes to preserve why data are absent.
A sixty-second orientation: absence has types
Data systems often need to distinguish at least four broad situations: a value is known and equals zero; a value should exist but is unknown or unavailable; a value does not apply to this observation; or a value is deliberately withheld or suppressed. Many real systems add further states such as not collected, not yet reported, below detection limit, invalid, refused, skipped or pending.
The translator’s first job is therefore classification, not wording. If a source uses “—” in several ways, inspect the legend. If a survey codebook defines 97 as “refused,” 98 as “don’t know” and 99 as “not applicable,” those numeric values are not ordinary measurements. If a database field is null, do not replace it with zero because zero may be a valid observed value.
1. Zero is a value; missing is a state
A true zero tells you something quantitative. A count of zero events means the observation process recorded none under the stated conditions. A missing count tells you that the system does not currently provide a valid number. These states can lead to very different averages, totals, rates and decisions.
Suppose three fictional sites report 4, 0 and missing incidents. The mean of the two observed values is 2. If you incorrectly translate missing as 0 and calculate across all three, the mean becomes about 1.33. The change is not linguistic; it is a new analysis created by mistranslation.
Keep zero visible when it is real. Do not replace it with a dash simply because a table looks cleaner without zeros. Conversely, do not fill blank or null values with zero unless the source definition or data-processing specification explicitly authorises that imputation.
2. “N/A” is ambiguous until the source defines it
N/A can stand for not applicable, not available, not answered, not assessed or another project-specific phrase. In some organisations the intended meaning is obvious from a style guide; in others the abbreviation has been reused loosely over years. Do not translate N/A from habit.
A target-language equivalent for “not applicable” may be wrong when the source really means “data unavailable.” Those states differ. Not applicable can mean the variable has no meaningful value for this observation. Unavailable can mean a value exists in principle but cannot currently be supplied.
Ask an answerable source question when necessary: “Does N/A in this table mean not applicable or not available?” If the project contains both meanings, replace the ambiguous source abbreviation with distinct target labels only when the source owner approves the semantic split. Translation should clarify an established distinction, not invent one.
3. Null is usually a data-layer concept, not a word to localise blindly
In databases and programming environments, NULL or null often means that no value is present for a field. It is not necessarily the same as an empty string, zero, false or the word “unknown.” The exact semantics depend on the system, schema and application logic.
If NULL is a protected database token, do not translate it inside machine-readable data. Translate the human explanation around it. A documentation sentence can say that NULL indicates no stored value, while the literal token remains NULL in code examples or data extracts.
A common localisation failure occurs when a target interface displays the word “null” to ordinary users because the application exposes an internal state. The correct fix may be a user-friendly label such as “Not provided,” but that is an interface-design decision requiring knowledge of the underlying state. The translator should not guess that every null means the same thing.
4. Blank, empty string and null can be different
A spreadsheet cell can appear blank because it contains nothing, because a formula returns an empty string, because text is hidden, or because a display rule suppresses a value. A database can distinguish NULL from a zero-length string. A CSV file can distinguish some empty-field patterns depending on its schema and parser.
Translation should therefore avoid statements such as “blank means missing” unless the data specification confirms it. In one workflow, blank may mean “not reported.” In another, blank may be an accepted empty comment. In a third, every numeric blank may be invalid input.
When documenting a system, use precise nouns: null value, empty string, empty field, visually blank cell, omitted element. These phrases may feel technical, but they protect the reader from assuming that several distinct machine states are one thing.
5. Missing-value sentinels are codes, not measurements
Legacy datasets frequently use numeric sentinels such as −9, −99, 9999 or 99 to represent missing states. The code is chosen because it lies outside the expected valid range or follows a project convention. It should not be analysed as an ordinary observation.
NIST’s Dataplot documentation illustrates configurable missing-value codes and shows why a software system may assign a specific value when an empty field is encountered. The important translation lesson is that the number’s meaning comes from the codebook. A value of −99 may mean missing in one file and a perfectly valid temperature in another.
Never translate a sentinel into a natural-language number phrase. “−99” should not become “minus ninety-nine” in a reader-facing explanation unless you are describing the code itself. If a display is intended for users, render the defined semantic label while preserving the underlying code in the data layer.
6. “Unknown” and “not collected” are different
Unknown means the value is not known to the dataset or respondent under the relevant process. Not collected means the process did not attempt to obtain it, or the field was outside the collection design. The distinction matters when evaluating data quality and coverage.
If a target table translates both as “missing,” readers lose information about whether the gap reflects ignorance or design. A field not collected this year may be available next year if the form changes. A genuinely unknown attribute may remain unknowable even though the question was asked.
Preserve operational vocabulary in metadata and footnotes. Ordinary readers may not need every code, but analysts and auditors often do. A translated dataset should support the same interpretation of coverage as the original.
7. “Not applicable” means the question has no valid value here
Not applicable is structural. A variable may not make sense for a particular observation. For example, “number of previous pregnancies” may be not applicable under a study’s defined population categories, while “number unknown” would mean the question applies but the value is unavailable.
Do not turn not applicable into zero unless zero is a meaningful value under the variable definition. Zero previous events and a question that does not apply are logically different. They may belong to different denominators in later analysis.
Survey interfaces often need a distinct N/A option for this reason. If the target wording for “not applicable” is too similar to “none,” respondents may select it incorrectly. Translation quality affects data quality before analysis even begins.
8. “Not available” describes access, not logical applicability
Not available can mean the value exists in principle but is inaccessible, delayed, not released, or missing from the current source. This is different from not applicable. The field belongs to the observation, but the dataset does not currently provide it.
Consider a dashboard that shows current-year population for most regions but “Not available” for one region because publication is delayed. Translating that as “Not applicable” would imply that the region has no meaningful population value, which is absurd. The target must preserve the access or reporting status.
Temporal phrases can help: not yet available, unavailable at time of publication, pending release. Use them only if supported by the source. Do not add a future promise merely because unavailable data often arrive later.
9. “Not reported” may reflect process rather than absence
A source can mark a value as not reported even when an underlying value exists. The reporting entity may have omitted it, a publication may have excluded it, or a rule may not require disclosure. The translator should not equate not reported with zero or unknown.
In research tables, “NR” can mean not reported. In clinical summaries, it may mean another domain-specific term. Never expand a short code without checking the legend. The same two letters can carry different meanings across disciplines.
When space allows, a full target phrase is often safer than reusing an ambiguous abbreviation. But if the publication has a standard bilingual legend, preserve that convention so readers can compare tables consistently.
10. Refused and don’t know are response outcomes, not interchangeable missingness
Survey respondents may refuse to answer, say they do not know, skip a question accidentally, or be routed around the question by design. Each creates an absent substantive value, but the mechanisms differ.
A refusal can contain information about privacy or willingness. Don’t know can indicate genuine uncertainty. A skipped item may be accidental or technical. A routing skip can be structurally appropriate. Collapsing all four into “blank” destroys metadata about response behaviour.
Translate the response options consistently between questionnaire and dataset. If “Prefer not to answer” is shown to participants but stored as code 97 labelled “Refused,” document the mapping. The human phrase and technical code can differ while representing the same response state.
11. Skip patterns create structural missingness
Questionnaires often route respondents based on earlier answers. A person who answers “No” to “Do you own a vehicle?” may skip questions about vehicle model. Those later blanks are not ordinary nonresponse; the questions were not asked because they did not apply under the survey logic.
If the target questionnaire changes the wording of the routing question, ensure the skip condition still maps to the correct response. A mistranslated yes/no direction can send respondents into the wrong branch and generate invalid data rather than merely bad prose.
Survey translation therefore includes logic testing. Review the visible question, response code, branching rule and downstream missing-value label. Structural absence is part of instrument behaviour.
12. Suppressed or withheld data may exist but be intentionally hidden
Public statistics and confidential reports sometimes suppress values to protect privacy, comply with disclosure rules or avoid unreliable estimates. A suppressed cell is not the same as a missing measurement. The value may exist internally but is deliberately not shown.
Translate suppressed, withheld, confidential and restricted according to the source rule. These words carry different implications. “Suppressed for confidentiality” tells the reader why the value is absent. “Not available” would erase that governance information.
Never attempt to reconstruct a deliberately suppressed value from totals unless the publication explicitly allows derived disclosure. Translation should preserve confidentiality protections, not undermine them through arithmetic or explanatory prose.
13. Below detection limit is not zero
Laboratory and environmental data may report values as below a detection or quantification limit. This does not necessarily mean the true concentration is zero. It means the measurement procedure cannot reliably quantify the value below a stated threshold under the defined method.
A source might use “<0.01 mg/L,” “BDL,” “ND” or another code. The exact meaning of ND can vary: not detected, no data, not determined or another project-specific phrase. Check the method and legend before translating the abbreviation.
Do not replace <0.01 with 0.00 for visual consistency. That changes a censored or thresholded measurement into an exact zero. Likewise, do not invent a midpoint value unless the analytical method explicitly specifies an imputation rule.
14. Trace, negligible and zero can describe different concepts
Some reports use words such as trace, negligible, minimal or below reporting threshold. These may be qualitative descriptions rather than exact numerical zeros. Translate them with their evidential strength intact.
“Negligible” can mean small enough to be ignored for a specified purpose, not physically nonexistent. “Trace” can mean detected at a very low amount. A source that carefully distinguishes trace from zero should not become a target table with both shown as 0.
Where the project has quantitative thresholds for these terms, preserve the threshold definitions. If not, avoid inventing one. Qualitative magnitude words often depend on domain and purpose.
15. NaN is a computational value, not simply “missing”
NaN commonly means “Not a Number” in floating-point computing and data-analysis systems. It can arise from missing data representations or from undefined numerical operations depending on the software and workflow. Treat it as a technical token unless the schema defines a human-facing semantic meaning.
A displayed NaN might be better localised as “Not available” for a general user only if the application team confirms that missingness is the intended state. In developer documentation or machine data, preserve NaN exactly.
Do not translate the letters into a target-language acronym inside code. A programming environment expects the original token or its own language-defined syntax. Translate explanations, not protected keywords.
16. Infinite and undefined are not missing values
Computations can produce infinity, negative infinity or undefined results. These states may look like data problems, but they are not automatically missing observations. Division by zero, logarithms outside a domain or overflow can produce special numerical states.
If a report distinguishes “undefined” from “not measured,” preserve the distinction. An undefined ratio can arise even when all source values are known. Calling it missing suggests absent input rather than a mathematical limitation.
This matters in dashboards where a rate change from zero baseline may be shown as undefined or not meaningful. Do not translate that as 0% change. The dedicated eduKateSG article on percentage changes explains why zero baselines require special treatment.
17. “No data” can mean dataset absence, not individual missingness
A map may display “No data” for an entire region because the dataset does not cover that geography. This is different from a single respondent having a missing value. The scope of absence matters.
Translate dataset-level phrases such as data unavailable for this period, not covered, outside survey scope or no observations collected according to the source. Avoid a generic blank that leaves users wondering whether the site failed to load.
Good localisation can improve transparency by preserving the reason in the interface. But do not add a reason the source does not provide. “No data” should not become “No data because reporting is delayed” without evidence.
18. Dashes, dots and special symbols need legends
Printed statistical tables often use “—”, “..”, “.”, “:” or other compact symbols to represent absent or special values. The same symbol can mean different things across publications. A dash may mean zero, not applicable, negligible, no occurrence or data not available depending on house style.
Never translate the symbol by intuition. Read the table notes. If the source legend says “— = not applicable” and “0 = zero,” preserve that distinction in the target legend. If space permits, expanding the target label may be clearer than retaining an unfamiliar symbol, but the semantic mapping must remain exact.
When tables are exported to plain text, some special symbols may be lost. Keep an accessible textual legend so the meaning survives formatting changes.
19. Missingness mechanisms are analytical concepts, not ordinary labels
Statistical methodology may describe data as missing completely at random, missing at random or missing not at random under formal definitions. These phrases are technical terms. “At random” in this context does not mean careless or arbitrary in everyday language.
Translate the established methodological terminology used in the target discipline. If no stable equivalent exists, retain the English acronym MCAR, MAR or MNAR with a careful explanation rather than inventing a colloquial phrase that changes the conditional assumptions.
NIST’s missing-data materials discuss methods for working with incomplete information; broader statistical literature defines these mechanisms in model-based terms. A translator should preserve the technical category and avoid reducing it to “randomly missing” unless that is the established target-language term.
20. Imputed values are not observed values
Imputation fills or models missing values under a specified procedure. The resulting number may be useful for analysis, but it was not directly observed in the same way as an original measurement. A translated table should preserve any marker distinguishing observed and imputed values.
Do not translate imputed as measured, estimated as observed, or completed as verified. Those words carry different evidential status. If multiple imputation is used, the analysis may combine results across several imputed datasets rather than create one definitive replacement value.
A footnote such as “missing values were imputed using…” is part of the methodology, not optional clutter. Translate the method name and preserve parameters or model descriptions relevant to reproducibility.
21. Deletion rules need their scope preserved
Complete-case analysis, listwise deletion and pairwise deletion describe different ways of handling missing observations. Translating all of them as “missing cases removed” erases the scope of removal.
Listwise deletion may exclude an entire observation from a multivariable analysis if required variables are missing. Pairwise procedures may use available pairs for particular calculations. Complete-case definitions can depend on the set of variables in the analysis.
Preserve the method name and, where the source explains it, which variables or analyses define completeness. Sample size can differ across tables because of missingness rules; translation should not imply that every analysis used the same n.
22. Counts of missing values are data too
A quality report may show n missing, percent missing or response rate. Those summaries describe the pattern of absence and can be analytically important. Do not hide them in translation because they seem like administrative details.
If 20 of 100 records are missing a variable, the observed denominator for analyses of that variable may be 80. A table percentage calculated among observed records differs from a percentage calculated over all records. Preserve denominator labels such as valid percent, total percent or response rate.
The dedicated eduKateSG article on percentages and denominators is useful here: missingness changes which population a reported percentage refers to. Translation must keep that denominator visible.
23. A worked survey example
Consider a fictional survey variable “Household vehicle count.” Valid values are 0–8. Code 97 means respondent refused, 98 means don’t know, and 99 means not applicable because the question was skipped for a defined housing category. The questionnaire displays “Prefer not to answer,” “Don’t know” and “Not applicable” rather than the numeric codes.
A faulty translation might convert 97, 98 and 99 into ordinary counts because they are numeric, producing impossible households with 97 vehicles. Another might collapse all three to blank and remove the reason for absence. A third might translate zero as “No answer” because zero visually resembles an empty quantity.
The repair is to keep 0 as a valid substantive value meaning zero vehicles, preserve the three special response states, and maintain the code-to-label mapping. In analysis, the special codes should be excluded or handled according to the source specification, not averaged with valid counts.
24. A worked laboratory example
Use this fictional table: Sample A = 0.00 mg/L, Sample B = <0.01 mg/L, Sample C = ND, Sample D = blank. The method note defines 0.00 as a measured rounded result, <0.01 as below reporting threshold, ND as not determined because the run failed quality control, and blank as no sample submitted.
These four cells must not be translated into the same “zero” or “no data” label. Sample A contains a quantitative result. Sample B contains a censored threshold statement. Sample C has a failed determination. Sample D lacks a submitted specimen.
A good target table preserves both the numerical form and the method note. If ND has an established target abbreviation, use it only if the legend clearly defines it. Otherwise, a full phrase may be safer. Never infer a concentration for C or D.
25. A worked dashboard example
Imagine a public dashboard with four states: 12 incidents, 0 incidents, data pending, and suppressed for confidentiality. The visual design uses numbers for the first two and icons for the latter two. A translation should preserve the distinction even if the target layout has less space.
A tempting but harmful simplification is to show “—” for both pending and suppressed. Readers can no longer tell whether the value is delayed or intentionally hidden. If those statuses affect interpretation or trust, the target interface needs distinct labels or a legend.
Zero should remain 0. It communicates that reporting occurred and no incidents were recorded under the dashboard definition. Replacing 0 with “No data” would turn evidence of absence into absence of evidence.
26. A worked database/API example
Suppose an API returns {"middle_name": null, "nickname": "", "dependants": 0}. The three fields have different states. No middle-name value is stored. Nickname is present as an empty string under this hypothetical schema. Dependants is a valid numeric zero.
A localisation layer that converts all three to the empty string destroys information. An analytics layer that converts them all to zero creates false numeric data. The correct behaviour depends on the schema, but the translation principle is stable: do not merge distinct machine states merely because the user interface wants one visual placeholder.
Developer documentation should preserve literal tokens in code and explain them in natural language outside the code block. Human-facing interfaces can use friendly labels, but those labels must map back to the correct internal state.
27. Practice clinic: choose the correct missingness meaning
Practice one. A count is 0. Should it become blank? No. Zero is a valid value unless the source defines otherwise.
Practice two. A table shows N/A but the legend says “not available.” Should you translate it as “not applicable”? No. Follow the legend.
Practice three. A database stores NULL. Should the token be translated inside SQL or JSON? No. Preserve machine syntax and translate the explanation.
Practice four. A survey code 99 means refused. Should it be averaged with 1–5 responses? No. It is a special code, not a scale value.
Practice five. A lab reports <0.1. Should it become zero? No. Preserve the less-than threshold and method meaning.
Practice six. A value is suppressed for confidentiality. Should you calculate it from visible totals and publish it? No. Preserve the suppression rule.
Practice seven. A field says “not collected.” Is that the same as “unknown”? No. One describes process design; the other describes knowledge state.
Practice eight. A duration ratio is undefined because the denominator is zero. Is that missing data? Not necessarily. Inputs can be known while the mathematical result is undefined.
28. Frequently asked translation questions
Is blank the same as null? Not necessarily. A blank can be visual or textual; null is often a distinct data state. Check the schema.
Does N/A always mean not applicable? No. It is ambiguous across organisations. Use the source legend or style guide.
Can missing values be coded with numbers? Yes. Legacy systems often use sentinel codes. Those numbers should not be interpreted as ordinary measurements.
Is below detection limit the same as zero? No. It means the method cannot quantify the value below a stated threshold under the defined procedure.
Should every missing category be displayed to general users? Not always, but any simplified display must preserve the distinctions required for correct interpretation and downstream use. Simplification is a product decision, not an automatic translation step.
What is the strongest release check? For every absent-looking cell, ask: “If I replaced this with zero, would the meaning change? If I replaced it with not applicable, would the meaning change?” If yes, preserve the specific state.
29. Connect this specialist guide to the broader eduKateSG translation system
This article is a specialist branch of Master Art of Translation. It connects to Translate | Names, Numbers, Dates and Units, Translate | Nominal, Ordinal, Interval and Ratio Data, and Translate | Percentages, Percentage Points and Basis Points. For terminology precision, use the protected Vocabulary Learning Hub; for negation, quantifiers, scope and reference, use How English Works.
High-quality missing-data translation protects the difference between “none,” “we do not know,” “this does not apply,” “we did not collect it,” “we cannot show it,” and “the calculation has no defined value.” Those distinctions may occupy only one cell, one symbol or two letters, yet they can change totals, denominators, privacy interpretation and analytical conclusions. Translate absence with the same care you would translate a measured number. What is missing can be data too.