Data quality
Missing values in business datasets
A blank cell is a question, not a diagnosis. Before filling gaps in a business dataset, establish what each kind of missing value means and how it entered the records.
Last checked
Do not turn unknown into zero or delete incomplete rows simply to make an export look tidy. Preserve the meaning first.
Separate absence from a measured value
A blank may mean that a question was not asked, an instrument failed, a field did not apply or information was deliberately withheld. A zero may be a genuine measurement. Treating all these cases as the same value can change what a collection appears to say.
The W3C tabular data model distinguishes a cell's string representation from its interpreted value and allows metadata to identify null values. It provides a technical example of why a visible code and its meaning are different. It does not tell you the meaning of a code in your own system.
Ask the people who collected and maintained the records before changing them. If nobody can establish the meaning, record it as unresolved. An honest unknown is more useful than a confident but unsupported definition.
Make a small map of the missing-value codes
Choose an important field and list the representations found in the authorised system: blank cells, text labels, sentinel numbers or absent fields. For each, record the known meaning, where that explanation came from and whether the convention changed over time.
An invented example is a fictional inspection result recorded as not checked. That should not be relabelled as passed simply because no defect was entered. A different record marked not applicable may be complete for its purpose even though it has no measurement.
Use the data dictionary guide to keep these meanings alongside the field definition. Separate the recorded code from any later interpretation. Do not publish real identifiers or unusual combinations of values to illustrate the problem.
Check where the gaps occur
A single overall completeness figure can hide important differences. Our suggested internal review asks whether gaps cluster around particular periods, locations, product types, collection methods or system changes. Use only groupings that can be assessed appropriately within your organisation.
Record the denominator whenever you count. Missing measurements among records where a measurement was expected are different from missing values across every row. Keep the scope explicit so a later reader can reproduce the calculation without guessing which records were excluded.
This exercise should produce a description, not an unsupported explanation of cause. If older records contain more gaps, possible reasons still need investigation. Do not claim that a change in software or staff caused the pattern merely because the dates appear to coincide.
Label filling, filtering and inferred values
If values have already been filled, estimated or copied forward, document that processing separately from original observations. Ask whether the original value and the processing rule remain identifiable. A field that looks complete after processing is not necessarily a fully observed field.
Before any new processing, ask what the proposed task requires and who can approve the change. Removing incomplete rows may also remove a meaningful part of the collection. This guide is not an instruction to alter operational records, reconstruct personal information or run a new model over them.
An internal processing note can state the rule, the affected fields, the version, the reason and the validation performed. Keep unsupported assumptions visible. Use the dataset versioning guide to connect a transformed copy to the correct explanation.
Prepare a useful initial description
A safe outline can explain that the collection contains known missing-value categories, that some meanings remain uncertain and that coverage varies by period. Where an aggregate is appropriate to disclose, state how it was calculated. Do not add an impressive completeness percentage without a reproducible check.
Keep technical usefulness separate from permission to use or share the records. Missing names do not establish that a collection is anonymous, and a complete field does not establish that it may be licensed. Refer rights questions to the relevant advisers.
Fairlode accepts descriptions only at this stage. Explain the issue in ordinary language or with an invented example; no spreadsheet, record extract or screenshot is needed. A documented limitation helps an assessment, but it does not promise a suitable buyer or a sale.
General information, not legal advice, a valuation or a buyer offer. Illustrative examples are not customer case studies.
Sources and further reading
Sources checked 16 November 2026. Historical research is background, not evidence of current buyer demand.