Missing data basics

Missing values are gaps in your table: empty cells, None in Python, NULL in SQL, or sentinel codes like -1 meaning “unknown.” How you handle them changes model behavior and metrics.

Types of missingness

• MCAR — missing completely at random (rare in practice)
• MAR — missing depends on observed columns
• MNAR — missing depends on unobserved or the value itself (hardest)

Example MNAR: high earners skip income survey questions more often—dropping rows biases averages downward.

Audit missingness first

1. Count missing per column
2. Cross-tab missing flags with target or segment
3. Check if “missing” is informative (create indicator features)

Common strategies (preview)

• Drop rows — only if few rows and MCAR-like
• Impute — median/mode, or model-based (advanced)
• Separate category — “unknown” for categoricals

Cleaning lessons cover imputation workflow; never impute on full data before splitting train/test.

Important interview questions and answers

1. Q: Why MNAR matters?
   A: Imputing without modeling why data are missing can bias conclusions.
2. Q: Missing indicator feature?
   A: Binary column marking imputation—sometimes improves models when missingness is informative.

Self-check

1. What does NULL mean in SQL?
2. Name two strategies for missing numeric data.
3. Why audit missingness before imputing?

Tip: Ask why data is missing before filling—MNAR is common.