Data
Domain 3: Data
The Data domain tests whether you can specify, acquire, assess, prepare, and govern data for analytics. CAP questions focus on data quality dimensions, sampling, bias, missing values, transformation, privacy, and documentation.
Core rules and concepts
- Start from data requirements tied to the analytic question: entities, grain, time window, labels, features, and refresh cadence.
- Profile data before modeling: row counts, keys, duplicates, ranges, distributions, missingness, and match rates across sources.
- Evaluate quality on standard dimensions: accuracy, completeness, consistency, timeliness, validity, and uniqueness. Quantify defects and their impact.
- Treat missing data deliberately: understand the mechanism (MCAR, MAR, MNAR), then choose deletion, imputation, or missing-indicator approaches and validate the effect.
- Handle outliers by investigating cause, not automatic removal; outliers may be the signal in fraud or failure prediction.
- Manage sampling and bias: define the target population, avoid convenience samples, and check coverage against the deployment population.
- Protect data: apply minimization, access control, de-identification, retention limits, and legal or privacy review. Document lineage, transformations, and a data dictionary.
Worked example
Two source systems are merged for a customer view. Profiling shows 12% of records fail to match on customer ID, 3% are duplicate accounts, and signup dates are inconsistent across systems. The analyst quantifies match rate by channel and finds that one channel is systematically unmatched, which would bias the model. The team adds a probabilistic match rule, deduplicates using the most recent verified record, and documents the rule. They also add a missing-match indicator rather than dropping records. Before modeling, they compare the merged population to the deployment population and confirm coverage. This prevents a silent sampling bias that would make the model look accurate offline but fail in production.
Traps that cost marks
- Accepting source data without profiling or quantifying quality defects.
- Dropping missing or unmatched records without checking whether missingness is informative.
- Removing outliers automatically and erasing the rare events the model must detect.
- Ignoring sampling bias between training data and deployment population.
- Using personal or sensitive data without lawful basis, minimization, or de-identification.
- Failing to document transformations and lineage, making results irreproducible.