Data Literacy and Data Quality
Data literacy means reading, questioning, and using data correctly. Know the data types: structured (tables), semi-structured (JSON, logs), unstructured (text, images). Understand nominal, ordinal, interval, and ratio scales - ratio has a true zero, interval does not, so ratios of temperatures in Celsius are meaningless.
Data quality is judged on several dimensions: completeness, accuracy, consistency, timeliness, validity, and uniqueness. A dataset can be complete but inaccurate, or accurate but stale. Missing data handling differs: dropping rows (listwise deletion) biases results if missingness is not random; imputation (mean, median, model-based) preserves rows but adds assumptions. Distinguish MCAR, MAR, and MNAR at a conceptual level.
Watch for sampling bias (the sample does not represent the population), survivorship bias (only successes observed), selection bias, and label noise. Outliers are not automatically errors - investigate before removing. Correlation is not causation; confounding variables can create spurious relationships.
Traps: questions that treat "more data" as always better (quality and representativeness matter), that equate missing values with zeros, or that call ordinal data numeric without justification. Also watch for confusing data validation (format/range checks) with data quality assessment (fitness for purpose).
Worked example: a churn dataset has 40% missing "last_login." Dropping those rows removes exactly the disengaged users you want to predict - the missingness is informative (MNAR). Better: add a "was_missing" flag and impute, then validate against holdout data.
Rule of thumb: ask of any dataset - where did it come from, who is missing from it, and when was it captured?