Data Quality Assessments: A Theoretically Structured Overview of Approaches and Methods
Ralph Foorthuis
stat.OT, cs.DB, stat.ME
2026-08-23
Heineken's Foorthuis maps 15 data-quality methods onto formal/informal logic and norm/data drivers; LLMs sit in the exploratory quadrant because generation is not deterministic.
Everyone agrees that bad data wrecks reporting, analytics, and models. The hard part is how to assess it. Existing reviews either stop at dimension lists (completeness, consistency, accuracy) or drop into SQL-level checks such as string-length tests and foreign-key analysis. DAMA's body of knowledge speaks to stakeholders, but it still does not give a map you can use to pick a method.
Ralph Foorthuis, in Heineken's digital and technology group, supplies a mid-level typology: what kind of logic does the assessment use, and what drives it. The intended readers include auditors and strategy owners, not only engineers who write validation rules.
Two axes.
The cross product is four quadrants:
LLMs sit in quadrant four for a blunt reason. The training algorithm is formal, but sampled generation with a temperature setting is not fully predictable, and results move with the prompt, so the method does not count as a deterministic rule. If an agent later calls a strict validator, that slice belongs back in specification-based validation or data-based comparison.
An appendix restates the author's earlier anomaly typology: five data dimensions, three super-types, nine basic types, and more than sixty subtypes.
This is a classification paper. There is no new accuracy number and no experiment that tests the typology against alternatives. The deliverable is coverage:
| Quadrant | Methods | Typical tools |
| Specification-based | 4 | schema, domain rules, distributional tests, supervised detectors |
| Data-based comparison | 3 | cross-system values, digests, reality checks |
| Perception-based | 4 | interviews, desk review, observation, questionnaires |
| Exploratory analysis | 4 | unsupervised anomalies, manual EDA, profiling, LLMs |
Table 1 also tags each method by structured vs unstructured data, whether it can explore, and whether it takes input from the external world. Most methods handle both structured and unstructured data. Checksums handle unstructured data only by ignoring the data model.
Specification-based checks are somewhat tautological: if an attribute may only be A or B, a C is wrong by construction. The conclusion is nearly always true and almost never insightful. Exploratory analysis is the opposite: you can start immediately, and the findings are less certain. The Kenyan hospital mixed-methods study is cited as a side result: mixing methods made the assessment more robust because one method could correct another.
For people who train models, the map is a warning against treating data quality as "a few more Great Expectations rules." Representativeness, class imbalance, noise, and drift sit in the distributional-test slot of specification-based validation. Using an LLM to flag missing text or impute values sits in the exploratory quadrant, with human oversight still required.
The Heineken framing is practical: methods are defined by the work to be done, not by the name of the output, which makes it easier to pick tools for an audit. The contribution is a mid-layer map, not a new algorithm and not a new benchmark.
The author states that the typology has not been evaluated empirically, and that future work should test how methods combine into a full audit. The LLM slot is also left slightly open: later techniques may earn a place in another quadrant.
The 15 methods come from the author's reading of the literature. There is no search protocol, no inclusion rule, and no count of papers per class. The claim that no such theoretically structured overview exists yet is an awareness claim, not a systematic review. DAMA's seven common dimensions are adopted as-is; overlap among dimensions is noted, not resolved. Issues that are specific to training data (label noise, synthetic contamination, eval-set leakage) appear only as a brief mention of representativeness, imbalance, noise, and drift inside distributional tests. They do not get their own class.