First steps to identify and handle missing values

WHAT THIS TESTS: This question probes your ability to treat missingness as a first-class modeling problem rather than a preprocessing chore. Interviewers want to see that you investigate why data is absent before deciding how to repair it, because the mechanism behind missing values often matters more than the imputation technique itself.

A GOOD ANSWER COVERS: Four phases in order. First, detection and quantification: compute the percentage of missing values per column and per row, check for complete duplicates, and flag columns that are entirely empty. Second, pattern analysis: use heatmaps or correlation matrices of null indicators to see if missingness clusters by feature or by subgroup, which hints at whether the data is missing completely at random, missing at random, or missing not at random. Third, mechanism-driven strategy selection: if the missingness is MCAR and under five percent, listwise deletion may be acceptable, but if it is MAR or MNAR, deletion introduces bias. Fourth, method comparison: contrast a univariate method like mean or median imputation against a multivariate method such as KNN or iterative imputation. Mean imputation is computationally cheap and easy to implement, but it collapses variance, weakens correlations, and produces biased standard errors. KNN or iterative imputation preserve relationships between variables and yield more realistic distributions, yet they demand more memory and processing time, and they assume that the observed rows are sufficient to reconstruct the missing ones.

COMMON WRONG ANSWERS: Jumping straight to filling nulls with zeros or the global mean without plotting the missingness pattern. Recommending dropping all rows with any nulls without checking how much data is lost or whether the remaining set is biased. Treating missingness as purely a software bug rather than a statistical signal. Failing to mention that imputed values should be tracked with missingness indicator columns so downstream models can learn whether a value was originally absent.

LIKELY FOLLOW-UPS: How would your strategy change if the missingness were concentrated in a single demographic group? What would you do if fifty percent of a critical feature were missing? Can you explain the difference between MAR and MNAR with a concrete business example? How do you validate that your imputation did not distort the target variable distribution?

ONE CONCRETE EXAMPLE: Imagine a customer churn dataset where income is missing for twenty percent of users. A heatmap reveals that income is missing far more often for younger customers who joined via mobile app sign-up, suggesting MAR. Mean imputation would insert the average income across all ages, artificially inflating the purchasing power of the young mobile cohort and potentially masking churn drivers. Instead, iterative imputation using age, tenure, and app usage as predictors would infer income conditional on those covariates, preserving the true relationship between income and churn risk while adding a binary indicator column showing which rows were originally missing.