Data Quality Dimensions
Measure trust before you build on it
Before analysis can be trusted, data must pass six quality gates: accuracy, completeness, consistency, timeliness, validity and uniqueness. Define them, measure them, then clean.
What you will learn
- Name the six quality dimensions
- Assess a dataset against each one
- Prioritize cleaning by decision impact
What is this?
Data quality measures fitness for use. The classic dimensions: accuracy (right values), completeness (no missing), consistency (same facts agree), timeliness (fresh enough), validity (fits the schema/domain), and uniqueness (no duplicates). An analyst's job starts by scoring each one on real data — then deciding what must be fixed before a decision can rely on it.
Why do you need it?
- Garbage-in resuscitation is the main driver of wrong business decisions.
- Quality problems compound: one bad join fans out into hundreds of bad rows.
- Documented quality scores communicate risk to stakeholders who will rely on the analysis.
- Catching issues at the boundary is 10x cheaper than at the dashboard.
Mental model
A building inspector's checklist
Data quality assessment
An inspector doesn't 'clean' the building — they verify foundations, wiring, plumbing, permits, occupancy and duplicates against code, then issue findings. Data quality gates do the same: measure, flag, and only then remediate.
How it works
- Accuracy: compare a sample against an authoritative source or business rule.
- Completeness: count nulls per column and empty/zero placeholders.
- Consistency: cross-check the same fact in two fields or two sources.
- Timeliness: compare event/load timestamps against the expectation.
- Validity: check values against allowed domains, formats and ranges.
- Uniqueness: count exact and key-valued duplicates.
- Prioritize fixes by the magnitude and direction of decision impact.
Six gates, one assessment pass.
In production
- A data team publishes a quality scorecard per table so executives see freshness and null rates before relying on KPIs.
- An ingestion job fails loudly when a core dimension's uniqueness drops below threshold.
- A vendor evaluation compares competing data providers on completeness and freshness scores.
Best practices
- Write the expected values down before you look — tests without expectations are decoration.
- Score each dimension, then repair only what the decision needs.
- Keep a quality dashboard with time trends so regressions are visible.
- Document 'known dirty' tables — unsolved issues cost less when they're declared.
Common mistakes
- Cleaning everything 'just in case' and losing traceability.
- Measuring nulls but not accuracy, then being surprised by wrong values.
- Fixing data in the notebook without ever restating it upstream.
- Treating freshness as a binary: 'today's data is clean'.
Debugging tips
- When a KPI disagrees with reality, run the six-gate audit — one dimension usually explains it.
- X-ray two sources of the same fact to expose consistency breaks.
- Count exact vs fuzzy dupes separately — they fail differently.
- Re-run the scorecard after a schema change; quality assumptions go stale fast.
Interview checkpoint
Hands-on lab
- Apply the six dimensions to a messy export
- Quantify each problem with pandas
- Publish a findings list, not just fixes
Score a Sales Export Across Six Gates
30 min · beginner
Quick quiz
A date stored as '2026/13/45' fails which dimension?
A table with no NULLs is automatically accurate.