What questions should you ask before filling missing values in a dataset?

Ask what each missing value means, how it was recorded, and whether absence is related to the outcome or another variable. A missing purchase amount might mean a failed data transfer, an unavailable measurement, or a transaction that never occurred; these require different treatment. Inspect missingness by time, source, and segment. Consider retaining a missingness indicator, using a justified imputation method, or excluding unsuitable records. Fit lea ed imputation only on training data when modeling, and document how the decision could affect conclusions.