Data quality is not an abstract score.
Data quality is not an abstract score. It determines which decisions remain defensible with available data. A complete series badly reconciled across spend, exposure and sales can produce a precise model on the wrong scope. Controls must therefore be tied to intended use: describing, comparing, attributing, forecasting and optimising do not require the same evidence.
Decision and method.
Allow, limit or defer analysis against decision-specific quality thresholds; never let an average compensate for a blocking defect. State the expected output, granularity, horizon and cost of error. Measure completeness, validity, consistency, uniqueness, freshness, matching and financial reconciliation on the actual analysed scope. Separate random error, systematic bias, definition break and structural absence, recording affected segments and periods. Green allows the decision, amber requires sensitivity analysis and red limits use to exploration; the remediation plan has an owner and date. Preserve raw data and transformation trace for every correction.
Worked example.
The dataset has 96% completeness, 91% validity, 85% freshness, 78% exposure-to-sales matching and 99% spend reconciliation: simple average 89.8%. Matching threshold for attribution and optimisation is 90%. The 12-point gap primarily affects mobile, representing 34% of sales. The average cannot compensate. Descriptive analysis is allowed; attribution and reallocation are deferred until mobile identifiers are fixed, or bounded by a sensitivity excluding that channel.
Checks and limits.
Define thresholds by use before observing scores; do not compensate a blocking defect; preserve raw data and transformation trace; report segments, periods and decisions made fragile. Technically compliant data may remain causally insufficient. Automated checks poorly detect business-definition changes, and a high global rate can conceal poor quality in a decisive segment.
Resources and sources.
Download the control register. Related: define necessary data, control MMM data, build the analytical dataset. Sources: ISO/IEC 25012:2008; Batini et al. (2009); Redman (2013).