Data Quality Dimensions
Key Takeaways
- CBDA Domain 2 (Source Data, ~15%) expects practitioners to assess data quality against named dimensions—accuracy, validity, completeness, consistency, timeliness, and uniqueness—before analysis, not after results surprise stakeholders.
- Validity means values conform to defined formats, ranges, codes, and business rules; accuracy means values correctly represent the real-world fact—a value can be valid and still wrong.
- Classic CBDA-style traps map concrete defects to dimensions: date format mismatch across systems is consistency; blank required fields is completeness; duplicate customer keys is uniqueness; stale inventory snapshots is timeliness.
- Assess quality relative to the research question and decision risk: scorecards, profiling, and threshold agreements make go/no-go and remediation decisions transparent.
- Remediation is a business decision (fix at source, transform with rules, exclude, or proceed with documented limitations)—not an automatic cleanse-everything mandate.
Data Quality Dimensions
Quick Answer: Before analysis, CBDA practitioners assess whether sourced data is fit for the research question using quality dimensions—especially accuracy, validity, completeness, consistency, timeliness, and uniqueness. Scenario items often give a concrete defect (for example, dates stored as
MM/DD/YYYYin one system andYYYY-MM-DDin another) and ask which dimension is violated; answer with the dimension name, then plan remediation that matches decision risk.
Source Data is the second CBDA knowledge area and carries roughly 15% of exam weight. After research questions are framed (Domain 1), practitioners must obtain data that can actually answer those questions. Competency work in this area is not “download a CSV and hope”—it is deliberate evaluation of source fitness, quality risks, and preparation needs so later analysis is trustworthy. On the exam, quality questions rarely ask for vendor product names; they test whether you can name the right dimension, separate validity from accuracy, and decide what to do next when quality is imperfect.
Why Quality Assessment Comes Before Analysis
Poor data does not merely reduce model accuracy; it changes the business meaning of every chart, lift estimate, and recommendation. A complete but inconsistent customer master can invent “churn” that is really a system merge problem. Highly accurate transaction amounts with incomplete product codes can bias category insights. Timely operational feeds that fail uniqueness checks can double-count revenue.
CBDA-aligned practice therefore treats quality assessment as a gate between sourcing and analysis:
- Confirm fitness for the research question — Does this data measure the constructs the business cares about, at the right grain and population?
- Profile the data — Measure nulls, formats, ranges, duplicates, referential integrity, and freshness against expectations.
- Score and communicate — Use a simple scorecard so sponsors see tradeoffs (proceed with caveats, remediate, or change sources).
- Document limitations — Assumptions and quality debt travel with the results into interpretation and decision support.
Skipping this gate is a common failure mode in scenario stems: teams jump to modeling because “we have a lot of data,” then discover the metric definition or join key was broken only after executives see the dashboard.
Core Dimensions Tested on CBDA
Different frameworks list six, eight, or more dimensions. For CBDA study, master the following set and be ready to map any realistic defect to one primary dimension (secondary dimensions may also apply, but exam items usually have a best fit).
Accuracy
Accuracy means the stored value correctly represents the real-world entity, event, or measurement. A phone number that is dialable for the actual customer is accurate; a birth date that matches the government ID is accurate. Accuracy is about truth, not only format.
CBDA example: warehouse cycle counts show on-hand quantity 120, but a physical count finds 95. The stored figure is inaccurate even if it is a valid integer in range.
Remediation directions: source-system process controls, reconciliation to authoritative sources, sensor calibration, human data-entry training, and dual entry for critical fields.
Validity
Validity means values conform to defined formats, domains, ranges, codes, and business rules for the attribute—whether or not those values happen to be true in the real world. Validity is rule conformance.
CBDA example: a “state” field must be one of the approved two-letter codes. XX fails validity. A temperature of −400°C for ambient warehouse air fails a reasonable range rule.
Remediation: validation rules at capture, pick-lists instead of free text, schema constraints, reject/quarantine pipelines, and reference-data tables for allowed codes.
Completeness
Completeness means required data is present for the population and attributes needed for the analysis. Completeness can be measured at the field, record, or population level.
CBDA example: 28% of closed-won opportunities lack a “lost competitor” field that the research question needs for win/loss analysis—the dataset is incomplete for that question even if other CRM fields are full.
Remediation: improve capture processes, backfill from alternate sources, redefine the analysis population to records with required fields (with bias notes), or change the research question if the attribute cannot be obtained ethically and legally.
Consistency
Consistency means the same real-world fact is represented the same way across records, systems, time, or related attributes—and that related values do not contradict each other. Format mismatches, conflicting birth dates across systems, and metrics that do not reconcile across reports are classic consistency failures.
IIBA sample-style trap: dates stored as 03/04/2026 in the CRM and 2026-04-03 (or a different day-month interpretation) in the warehouse. The exam-ready diagnosis is primarily a consistency problem (representation and cross-system alignment), often with validity implications depending on parsing rules. When two systems show different birth dates for the same customer, consistency is the dimension to name first.
Remediation: master data management, shared reference calendars and code sets, ETL standardization rules, single source of truth for critical attributes, and cross-system reconciliation reports.
Timeliness
Timeliness means data is sufficiently up to date for the decision and available when needed. A perfect overnight batch is timely for a monthly board pack and untimely for real-time fraud blocking.
CBDA example: inventory allocation decisions use last night’s snapshot, but flash sales change stock every few minutes—allocation recommendations become operationally late even if every row is accurate for the snapshot time.
Remediation: increase refresh frequency, use event streams for critical entities, document latency SLAs, and align analysis cadence with decision cadence.
Uniqueness
Uniqueness means each real-world entity or event is represented once at the intended grain (or duplicates are controlled and explained). Duplicate primary keys, multiple “golden” customer records, and double-counted transactions break uniqueness.
CBDA example: the same order ID appears twice after a botched merge of regional databases, inflating revenue by the duplicate amount.
Remediation: deduplication rules, survivorship logic, unique constraints, identity resolution, and post-load duplicate audits.
Validity vs Accuracy: The Distinction CBDA Loves
Memorize this contrast; it appears in both straightforward definitions and scenario stems.
| Aspect | Validity | Accuracy |
|---|---|---|
| Core idea | Conforms to rules, formats, ranges, and allowed codes | Matches the real-world truth |
| Question asked | “Is this an allowed value for this field?” | “Is this the correct value for this entity/event?” |
| Example valid but inaccurate | Phone number 555-0100 matches format and length, but belongs to a different person | — |
| Example invalid (accuracy may be unknown) | Age entered as 999 when max allowed is 120 | Even if someone meant 99, the stored value fails validity |
| Primary controls | Schema, validation, reference data | Source of truth, process quality, reconciliation |
A dataset can score high on validity (clean formats, no illegal codes) and still be dangerously inaccurate. Conversely, free-text fields may hold accurate truths that fail validity until standardized. For exam answers, if the stem emphasizes business rules, formats, or code lists, choose validity. If it emphasizes truth versus reality (wrong address, wrong quantity, wrong customer linked), choose accuracy.
Mapping Defects to Dimensions (Exam Drill Table)
Use this table as a rapid lookup when a scenario describes a symptom rather than a dimension name.
| Dimension | Definition | CBDA-style example | Remediation focus |
|---|---|---|---|
| Accuracy | Values correctly reflect reality | Shipped quantity stored as 10; customer received 8 | Source controls, reconciliation, measurement fix |
| Validity | Values conform to rules/formats/ranges | Status code Z9 not in approved list | Validation rules, pick-lists, reject quarantine |
| Completeness | Required data present for use | 40% of claims missing ICD code needed for analysis | Capture process, backfill, population redefinition |
| Consistency | Same fact aligned across systems/records | Date formats or birth dates differ across CRM and warehouse | Standardization, MDM, cross-system rules |
| Timeliness | Fresh enough for the decision | Weekly churn file used for same-day retention offers | Faster refresh, latency SLA, cadence redesign |
| Uniqueness | One representation at intended grain | Duplicate customer IDs double-count retention cohort | Dedup, identity resolution, unique constraints |
When two dimensions seem plausible, pick the one the stem emphasizes. “Different birth dates in two systems” → consistency. “Field is blank for half the population” → completeness. “Value outside allowed range” → validity. “Physical count disagrees with system” → accuracy.
Assessing Quality Before Analysis
Quality assessment is evidence-based work, not vibes. Typical CBDA-aligned steps:
- Define expectations from the research question — Required entities, grain (order line vs order), population filters, critical attributes, acceptable latency, and legal/privacy constraints.
- Profile — Distributions, null rates, distinct counts, min/max, pattern checks, foreign-key orphans, duplicate rates, and freshness timestamps. Profiling is the technique most often associated with finding missing values, outliers, and distribution issues before formal analysis.
- Compare profile results to expectations — Convert raw stats into dimension scores or pass/fail against agreed thresholds.
- Trace issues to sources and processes — Distinguish extract bugs from true source defects so remediation targets the right owner.
- Decide — Proceed, remediate, change source, narrow scope, or escalate—always with documentation.
Quality Scorecards
A quality scorecard is a lightweight management artifact that makes fitness-for-use visible. Typical columns: dimension, metric (for example, % non-null for required fields), threshold, current value, status (green/amber/red), owner, and remediation action. Scorecards help non-technical sponsors understand why analysis is delayed or why results carry caveats.
Best practices for CBDA scenarios:
- Tie thresholds to decision risk. Fraud blocking may require near-zero invalid account IDs; a directional market study may tolerate more missing demographic fields.
- Score at the right grain. Completeness of order headers can look excellent while line-level product codes are incomplete—and product mix analysis fails.
- Separate source quality from pipeline quality. A red consistency score after ETL may be a transformation bug, not a CRM problem.
- Re-score after remediation. Quality work is iterative; a one-time cleanse that is not monitored often decays.
Business Judgment: Not Every Defect Stops the Project
CBDA practitioners are business-oriented. Perfect data is rare. The competency is to know the risk, communicate it, and choose a proportionate response:
- Block when critical keys, legal identifiers, or outcome labels are unreliable for the decision at stake.
- Remediate when fixes are feasible within time and cost relative to value.
- Proceed with limitations when the research question can still be answered for a well-defined subset, with bias and confidence clearly stated.
- Reframe when no available source can meet minimum fitness—even excellent techniques cannot rescue unusable inputs.
Exam traps include “always delete incomplete rows,” “always impute,” or “quality is IT’s problem only.” The better answer connects quality findings to the research question, stakeholder decision, and transparent documentation.
Linking Dimensions to Later Domains
Quality work does not end when analysis starts. Dimension failures reappear when interpreting results (Domain 4), influencing decisions (Domain 5), and governing data (Domain 6). A consistency issue that was “accepted with caveats” must still appear in the story told to executives; otherwise the organization acts on a false precision. Treat dimension language as shared vocabulary across the analytics lifecycle—not a sourcing-only checklist.
Study Checklist for This Section
- Name six core dimensions and give one business example each.
- Explain validity vs accuracy with a single sentence each and one pair of examples.
- Map: date format mismatch → consistency; illegal code → validity; physical vs system count → accuracy; blank required field → completeness; late feed → timeliness; duplicate keys → uniqueness.
- Describe how a scorecard supports go/no-go decisions before analysis.
- State why quality assessment is relative to the research question, not absolute perfection.
Which dimension of data quality describes whether data values conform to defined formats, ranges, and business rules for the attribute?
An analyst finds customer birth dates stored as MM/DD/YYYY in the CRM and as YYYY-MM-DD in the enterprise warehouse, with some records resolving to different calendar days after naive parsing. Which data quality dimension is PRIMARILY violated?
A phone number field only accepts 10-digit numeric strings and passes validation, but many numbers belong to previous account holders after a flawed porting process. What is the BEST quality diagnosis?