Data quality is not just a technical check for nulls, duplicates and invalid formats. The harder question is whether the data means what people think it means, compares like with like and is reliable enough for the decision being made.

Start with definitions

A clean metric can still be misleading. Before testing the values, confirm the row meaning, population, time period and exclusions. A definition check is often the fastest way to find the real problem.

Reconcile detail back to a known total

The strongest validation checks connect detailed output to something independently credible. One population model produced 52 million Australians because the local assumptions were never reconciled to the national total. Rolling the result back up exposed the error.

Check whether the comparison is still valid

Categories, store networks, product hierarchies and geographic boundaries change. Historical data can be technically correct but no longer comparable. The ABS Census harmonisation project rebuilt older data onto consistent boundaries before any trend was interpreted.

Technical validity is not the same as decision safety

Some data is good enough for directional exploration but not for a pricing decision, financial target or customer-level intervention. The validation standard should match the consequence, and the limitations should travel with the result.

A practical review sequence

Define the row, metric, population and period. Check missing, duplicated and impossible values. Reconcile totals, test comparability, then ask whether the result makes sense in context. Finally, document what the dataset is safe to support.

The goal is not perfect data. It is honest analysis: knowing where the evidence is strong, where it is fragile and how much confidence the decision deserves.