An independent study reference written by Dr Phuc V. Nguyen. It is not official subject material — for assessment requirements always follow your subject outline and vUWS.

Data Foundationsintermediate

Data quality dimensions

Q: A postcode field passes every validation rule, and every address involved was confirmed with the customer in the past seven days, yet parcels are being misdelivered. Which dimension is most likely failing?

Accuracy, because well-formed postcodes do not match the real addresses. A four-digit number can be perfectly well formed and still be the wrong postcode for the address it sits beside. Validity is a format test and cannot detect that, so accuracy has to be checked against an authoritative address reference. Stale records would produce the same symptom, and the stated capture window is what rules that out, since without it timeliness would explain the misdeliveries just as well.

Q: A team reaches 100 per cent completeness on the industry field after making it mandatory at data entry. Sales analysis by industry then gets worse. Why is that plausible?

Staff selected the first available option to move on, so the field is now complete and largely wrong. Forcing a value converts a visible gap into an invisible error. A null at least announced that the value was unknown, whereas a default chosen under time pressure looks like evidence and is treated as such by everything downstream.

Quality becomes manageable once it is broken into dimensions that can each be measured. The common set is completeness (are required values present), accuracy (do values match an authoritative reference), consistency (do related values agree with each other and across systems), timeliness (is the data current enough for the decision), uniqueness (is each real-world entity represented once) and validity (does the value obey its format or code list). Each becomes a percentage computed by a stated rule on a named field, so quality moves from an opinion to a monitored number, and a failing threshold points at a specific defect rather than a general complaint.

Why it matters

Saying the data is bad is like saying a car is bad. Useless. Is the fuel gauge wrong, are two tyres flat, or has the registration expired? Each has a different fix and a different urgency. The dimensions are the separate gauges. Once you can say that completeness on mobile numbers is 92 per cent and falling, you have something a person can act on straight away.

Before you read on — recall

A postcode field passes every validation rule, and every address involved was confirmed with the customer in the past seven days, yet parcels are being misdelivered. Which dimension is most likely failing?

Formulas

Completeness

C = \dfrac{n_{\text{populated}}}{n_{\text{required}}} \times 100\%

Measured on a named field, across the rows where that field is genuinely required. Of 50,000 customer records, 4,100 have no mobile number, so completeness on that field is 45,900 divided by 50,000, about 91.8 per cent. Completeness across a whole table means nothing, because the fields are not equally required.

Uniqueness

U = \left(1 - \dfrac{n_{\text{duplicate}}}{n_{\text{total}}}\right) \times 100\%

If 620 of 50,000 customer rows are extra copies of an entity already present, uniqueness is about 98.8 per cent. The figure depends entirely on the matching rule used to decide two rows are the same person, so the rule belongs in the published definition of the measure.

Timeliness as currency against a tolerance

A = t_{\text{now}} - t_{\text{updated}} \;\le\; \tau

Age

A

is the gap between now and the last update, and

\tau

is the tolerance the decision allows. A price list refreshed nightly has an age of up to 24 hours, which is fine for a periodic margin review and unacceptable for a live pricing engine. Timeliness only means something against a stated tolerance.

Worked examples

Scenario

A monthly data quality report shows an overall quality score of 94 per cent. The executive sponsor asks what to do about the missing 6 per cent.

Solution

A blended score cannot answer that, because it hides which dimension failed and on which field. Splitting it shows validity at 99.6 per cent, completeness on mobile number at 91.8 per cent, and uniqueness on customer at 98.8 per cent. Those imply three different responses: a change at the point of capture, a matching and merge exercise, and no action at all for validity. Scores are useful for tracking a trend, and the action always comes from the dimension and field underneath.

Scenario

A customer record shows a date of birth of 1 January 1900. The validation rule passes it.

Solution

The value is valid, because it is a well-formed date inside the accepted range, and it is not accurate, because nobody in the file was born then. It is almost certainly a default keyed to get past a mandatory field. Validity checks format and code lists and is cheap to automate. Accuracy needs comparison against something authoritative, which is expensive, so a practical compromise is to monitor suspicious concentrations. One date holding four per cent of all records is a signal no format rule will ever raise.

Common mistakes

✗Accuracy and validity are the same thing. Validity asks whether a value obeys its format or code list, accuracy asks whether it matches reality, and a well-formed wrong value passes validity easily.
✗A single overall quality score is enough to manage by. Blended scores hide which dimension and which field failed, and the action always comes from that detail.
✗Completeness means no nulls anywhere. A null is only a defect where the field is required for the use in question, and forcing values into optional fields manufactures data that is worse than a null.
✗Timeliness is about how fast the pipeline runs. It is the age of the data relative to the decision, so a slow pipeline that delivers a monthly figure on the second of the month is perfectly timely.

Revision bullets

•Completeness, accuracy, consistency, timeliness, uniqueness, validity
•Each dimension is a percentage computed by a stated rule on a named field
•Validity is cheap to automate, accuracy needs an authoritative reference
•Uniqueness depends on the matching rule, so publish the rule with the number
•Timeliness is age against a decision tolerance, not pipeline speed
•Blended scores hide the field and dimension that actually need action

Quick check

A team reaches 100 per cent completeness on the industry field after making it mandatory at data entry. Sales analysis by industry then gets worse. Why is that plausible?

Connected topics

Metadata Fitness for Use Why Data Goes Bad Data Cleansing KPI Design

More in Data Foundations

What Big Data Means Forms of Data Data Warehouse ETL and ELT Star Schema Measures & Dimensions Cloud for Analytics Sampling Bias

Sources

Wang & Strong (1996)
Wang, R. Y., & Strong, D. M. "Beyond Accuracy: What Data Quality Means to Data Consumers." Journal of Management Information Systems, 12(4), 5-33, 1996.
Derives and groups data quality dimensions empirically rather than asserting them.
ISO/IEC 25012:2008
International Organization for Standardization. ISO/IEC 25012:2008, Software engineering: Software product Quality Requirements and Evaluation (SQuaRE), Data quality model.
A standardised data quality model separating inherent characteristics from those that depend on the system holding the data.
DAMA-DMBOK (2017)
DAMA International. DAMA-DMBOK: Data Management Body of Knowledge, 2nd ed. Technics Publications, 2017.
Treats dimensions as things to be defined, measured and monitored against thresholds rather than described.

How to cite this page

Dr. Phil's Quant Lab. (2026). Data quality dimensions. Derivatives Atlas. https://phucnguyenvan.com/concept/ba-data-quality-dimensions