Updated November 3, 2026
Missing values are not all the same. A blank cell may mean a measurement was not taken, a question did not apply, a system failed to record a value or the value was intentionally withheld.
Measure the pattern
Calculate missing counts and percentages by column and, where useful, by important groups or time periods. Check whether missingness is concentrated in particular segments.
Review whether a column is essential to the task and whether missingness itself may carry information.
Choose a defensible treatment
Possible treatments include leaving values missing for models that support them, imputing with a statistic or domain rule, adding a missingness indicator, or excluding rows/columns when justified. Document the choice and its limitations.
Avoid leakage
For predictive modelling, learn imputation values from the training split only. Apply the fitted transformation to validation and test data without recalculating statistics from those sets.