Missing Data: When to Drop, When to Impute, When to Worry

An empty cell looks like an absence. Often it is information. Whether you should delete rows with missing values, fill them in, or stop and investigate depends entirely on why the values are missing — and that pattern is diagnosable from the data itself.

The Three Missingness Regimes

Statisticians classify missingness into three regimes, and the classification decides the remedy:

• MCAR — missing completely at random. The gap has no relationship to anything, missing or observed. A sensor that randomly skips a read. Here, listwise deletion (dropping incomplete rows) loses a little power but introduces no bias.
• MAR — missing at random. The gap depends on observed values, not on the missing value itself. Income is more likely missing for younger respondents, and age is observed. Deletion is now biased — the rows you keep are not a random sample of the whole.
• MNAR — missing not at random. The gap depends on the missing value itself. People with very low (or very high) income decline to report it. No amount of imputation from observed data fixes this; the missingness is a property of the world, not the dataset.

The practical label for most business data is: not MCAR. Exports drop fields when the source system considers them optional, forms skip questions that were conditional, and integrations silently omit columns they could not map. Assume bias until proven otherwise.

Drop vs. Impute vs. Investigate

Drop when the missingness is small in share (a few percent), MCAR-like, and confined to columns you barely use. Listwise deletion of a column at 2% missing is fine. Listwise deletion at 40% missing on a critical column is throwing away nearly half your sample.

Impute when the column matters and the pattern is MAR. But choose the imputation with your eyes open:

• Mean/median imputation preserves the central value but artificially shrinks variance and weakens relationships — a column imputed this way will show weaker correlations than reality.
• Carry-forward (last known value) suits slow-moving series like stock levels, and badly misleads on volatile ones.
• Model-based imputation (predict the missing value from other columns) is the most defensible, at the cost of complexity. For most business reporting, median imputation plus an explicit "imputed" flag in the report is a reasonable middle ground.

Investigate when the missingness is concentrated — a specific column, a specific date range, a specific source file. Concentrated missingness is not randomness; it is a system event. The classic example: a column that is fully populated until March and empty afterwards is not "missing data", it is a broken export or a changed workflow, and the correct action is a ticket to whoever owns the source system, not an imputation.

Missingness as Signal

Sometimes the gap is the finding. In survey and CRM data, a non-response is often the most informative response — people who skip the satisfaction question are systematically less satisfied. In operational data, a row with most fields empty may indicate an abandoned process, not an incomplete record. Before you fill a hole, ask what kind of hole it is. Tools like KPI Master surface this as a completeness grade per column precisely so the pattern — not just the count — gets looked at.