Data quality

Data quality is a multidimensional judgment about whether data can support a specified purpose. A widely used quality model treats accuracy, completeness, consistency, timeliness, and relevance or fitness for use as distinct but related dimensions.¹ Accuracy concerns correspondence with the thing represented; completeness concerns required values, records, or coverage; consistency concerns agreement across fields, sources, and rules; and timeliness concerns availability and currency.

These dimensions can conflict. A rapidly updated dataset may be timely but less accurate than a slower, verified release, while a highly detailed dataset may be complete for research but excessive for an operational decision. Data quality therefore requires explicit quality requirements rather than a single universal score. The intended users, acceptable error, unit of measurement, time period, and permitted missingness should be defined before assessment.

Quality management begins by translating business or research needs into measurable rules. Common measures include the proportion of records passing a validity check, the percentage of required fields populated, duplicate-record rates, reconciliation differences, error rates from sampled review, and the age of the latest update.

Controls work best across the data lifecycle: capture values as close as possible to their source, validate them at entry, document transformations, reconcile important outputs, and monitor recurring exceptions. Data governance assigns ownership and decision rights, while data lineage records where values came from and how they changed. Automated checks are useful for scale, but human review remains necessary for ambiguous classifications, changing definitions, and errors that are syntactically valid.

• Prevent: use controlled vocabularies, clear forms, reference data, and validation constraints.
• Detect: profile distributions, compare sources, test business rules, and investigate anomalies.
• Correct: retain provenance, fix source processes, and record remediation decisions.

High-quality data supports reliable analysis, safer operations, fairer decisions, and reproducible research. In electronic health records, for example, missing dates, inconsistent coding, or delayed entries can alter clinical and epidemiological conclusions; health agencies consequently use structured data-quality reviews and source verification. In finance, identity management, logistics, and public administration, quality problems can create duplicate entities, incorrect payments, failed matching, or misleading performance measures.

Improving quality is not simply a matter of removing unusual values. A rare observation may be correct, and changing it can destroy important information. Likewise, imputation can improve analytical completeness while reducing transparency if the imputed values are not identified. Quality work should preserve uncertainty, distinguish correction from estimation, and evaluate consequences for different populations. Fair access and reuse also depend on adequate metadata, persistent identifiers, and documentation of permitted interpretation, concerns emphasized by the FAIR Data Principles.²

Some of the most consequential quality failures occur in meaning and context rather than in obvious typos. A field may retain the same name while its definition, coding scheme, geographic boundary, or measurement instrument changes; this creates semantic inconsistency that ordinary format checks may miss. Data copied between systems can also acquire rounding, truncation, encoding, or time-zone errors without appearing incomplete.

Quality is often distributional: a dataset may have a high overall completion rate while missing particular regions, languages, age groups, or socioeconomic populations. Sampling bias and measurement bias therefore belong in a broader quality assessment even when individual records are internally consistent.³ Another overlooked issue is fitness under change. A model, policy, or operational workflow can make formerly suitable data unsuitable when behavior, technology, or definitions shift. Monitoring should consequently include drift, subgroup performance, version history, and the continued validity of quality rules.

The FAIR Data Principles improve findability and reuse but do not by themselves guarantee accuracy or absence of bias; those are separate quality questions.²

Accuracy

The degree to which recorded values represent the real-world entities, events, or measurements they describe.

Completeness

The extent to which required records, attributes, values, or population coverage are present.

Consistency

The degree to which data agrees with defined rules and does not conflict across fields, systems, or sources.

Timeliness

The extent to which data is current and available when it is needed.

Fitness for use

Suitability of data for a specified purpose, considering its required meaning, precision, coverage, and constraints.

Data profiling

Systematic examination of data structure, values, distributions, relationships, and rule violations.