Handling missing numerical values

What's really being asked

The interviewer wants to see structured reasoning about the bias-variance and data-loss trade-offs of imputation, plus awareness of why missingness arose. It is a judgment question, not a recipe.

The full answer

First, deletion: dropping rows with missing values is trivial and keeps the remaining data untouched, but it discards information and produces biased estimates unless values are missing completely at random. Second, simple imputation with the mean or median: this retains every row and is fast, but it artificially reduces variance, weakens correlations with other features, and the mean is sensitive to outliers so the median is often safer for skewed columns. Third, model-based imputation such as KNN or iterative regression imputation predicts the missing value from other features, capturing relationships and giving more realistic fills, at the cost of complexity, runtime, and leakage risk if not fit only on training data. Adding a binary missingness-indicator column can also preserve the signal that a value was absent.

The mistakes people make

Always using the mean regardless of skew or outliers. Imputing before splitting into train and test, which leaks information. Dropping columns or rows without checking how much data is lost or why it is missing. Ignoring that missingness itself may be predictive.

What usually comes next

Define missing completely at random, missing at random, and missing not at random. How does imputation interact with cross-validation. When would you prefer to add a missingness flag.

A concrete example

In a salary column where high earners decline to report, the data is missing not at random. Mean imputation would pull missing entries toward the average and understate true top salaries, biasing any downstream model. A better choice is model-based imputation plus a missingness indicator, so the model learns both an estimate and the fact that the value was withheld.