Marketing data quality: authorising or blocking a decision.
Key distinctions
Three conditions before calculating.
Quality relative to the use
Data can be enough to guide an investigation and remain insufficient to reallocate a budget. The threshold must be attached to the cost of an error and to the reversibility of the decision.
Non-compensatory controls
Excellent freshness does not repair identifiers that do not join. Completeness, validity, consistency, freshness and traceability must include blocking gates.
Explainable anomalies
Mechanically correcting an extreme value can delete a real event. Every treatment states the rule, the volume affected, the effect on the results and the ability to return to the raw data.
Method
Building a data-based authorisation to analyse
- 01Name the decision
Define the expected output, the granularity, the horizon and the cost of an error; this step turns a generic checklist into verifiable requirements.
- 02Test the dimensions
Measure completeness, validity, consistency, uniqueness, freshness, matching and financial reconciliation on the perimeter genuinely analysed.
- 03Qualify the gaps
Distinguish random error, systematic bias, definition break and structural gap; document the segments and periods most affected.
- 04Assign a status
Green authorises the decision, amber requires a sensitivity analysis, red limits the use to exploration; the correction plan has an owner and a date.
The five quality gates before analysis
Each control has an acceptance threshold. Here the matching rate blocks attribution and optimisation despite a favourable overall average.
A quality average of 90% that must block optimisation
The dataset shows 96% completeness, 91% validity, 85% freshness, 78% exposure-to-sale matching and 99% spend reconciliation. The simple average reaches 89.8%.
The matching threshold to attribute and optimise was set at 90%. The gap of 12 points concerns mainly the mobile channel, which represents 34% of sales. The average therefore cannot offset the defect.
Descriptive analysis remains authorised. Attribution and reallocation are deferred until the mobile identifiers are corrected, or framed by a sensitivity analysis that excludes that channel.
Acceptance conditions
What must be true to act.
- 01
Thresholds are defined by use before the scores are observed.
- 02
No blocking defect is offset by an average.
- 03
Every correction keeps the raw data and a trace of the transformation.
- 04
The report states the segments, periods and decisions made fragile.
Limits
What this analysis does not prove.
- Technically compliant data can remain causally insufficient.
- Automatic controls detect changes of business definition poorly.
- A high overall rate can hide poor quality concentrated on a decisive segment.
References
Works cited.
- ISO/IEC 25012:2008, Data Quality Model.
- Batini, C. et al. (2009), “Methodologies for Data Quality Assessment and Improvement”, ACM Computing Surveys.
- Redman, T. C. (2013), Data Driven: Profiting from Your Most Important Business Asset.