What are the main fixes for missing values?

Delete rows, fill with a defensible value, model the gap, or flag the missingness as its own feature, chosen by cause.

Why not always fill with the average?

It assumes the missing resemble the typical case, which is often false and silently biases the data.

Can missingness itself be useful?

Yes: the pattern of what is blank is frequently predictive and worth a feature of its own.

How much missingness is too much?

Depends on the cause and the field's value; a mostly-empty field is usually dead weight, whatever the fix.

AI Glossary / AI and machine learning / Beginner

Missing values

Definition

Missing values are the empty cells in a dataset: measurements never taken, questions never answered, sensors that failed, customers who skipped a field, each one a decision the data refuses to make for you.

What it means

Missingness is information before it is inconvenience. A customer who leaves income blank is different from one who fills it in, and a sensor that dies during overheating is telling you something about overheating.

Three patterns matter: values missing completely at random, missing for reasons related to observed fields, and missing for reasons related to the missing value itself, such as high earners skipping income questions. The pattern decides the safe remedy.

The everyday fixes are deletion, filling with a typical value, and flagging the missingness as its own feature. Each one is a claim about the data, and the wrong claim quietly biases every downstream number.

The worst handling is silent: averages filled without a record, blanks treated as zeros, rows dropped without a count. The dataset looks complete and every conclusion built on it inherits a hidden skew.

For a business, the rule is simple: missing values are a business fact with a cause, and the cause, not the spreadsheet, decides what to do about the gap.

How it works

Measure missingness per column and per segment before choosing a remedy; the pattern matters more than the percentage.

Investigate the cause: instrument failure, optional field, process change or refusal, because each implies a different fix.

Choose a remedy that matches the cause: drop, fill with a defensible value, model the gap, or flag it as its own feature.

Record every fill and drop in the pipeline so the handling is auditable and reversible, never a silent edit.

Re-check missingness over time, since a rising rate is often the first symptom of an upstream process breaking.

Three practical examples

A survey where income is blank for fifteen per cent of respondents, and analysis shows the blanks skew toward the highest earners.
A sensor feed where gaps cluster during night shifts, pointing at a maintenance schedule rather than random failure.
A signup form where a newly added optional field is empty for ninety per cent of users, making it useless as a model input despite looking rich.

Think of it this way

Think of a hospital chart with blank boxes: the blanks are not noise to erase but facts to read, because what a clinician chose not to record is itself a clinical observation.

Worked example

the blanks that predicted churn

Suppose a subscription business finds support-ticket category blank for twenty per cent of churned customers but only three per cent of retained ones. The figures are illustrative.

The team fills the blanks with the most common category and the model shrugs; then someone adds a simple was-blank flag instead, and churn prediction improves immediately.

The gap was never missing data: it was unrecorded frustration, tickets closed without categorization during incidents, and the flag captured it directly.

The reversal is the lesson: the same blanks were noise under one handling and signal under another, and only asking why they existed told the difference.

Illustrative case study

A fictional lender filled missing incomes with the portfolio average and wondered why its risk model treated average-looking applicants oddly.

The filled group was systematically different, higher earners who skipped the field, and the average pulled their scores in the wrong direction at scale.

The invented lesson: a filled value is a claim about the data. When the claim is wrong, you have not repaired the dataset; you have edited it.

Common misconceptions and mistakes

Missing values are just a cleaning chore. They are evidence about the process that produced the data, and the evidence is often valuable.
The average is a safe fill. It is a strong claim, that the missing resemble the average, which is frequently false and always untested unless you test it.
Dropping incomplete rows is neutral. It is safe only when missingness is unrelated to anything that matters, which is rarer than it looks.

Business use

Ask why each important field is missing before deciding how to handle the gap.

Add missingness flags as model features; the pattern of absence is frequently predictive itself.

Log every fill and drop with counts and reasons, so data handling survives an audit and a staff change.

Alert on changes in missingness rates, because a rising blank count usually announces an upstream break first.

Check your understanding

Why should the cause of a missing value decide its remedy?

Answer: Because every remedy is a claim about what the blank means, and only the cause tells you which claim is true: random gaps, systematic gaps and self-hiding values each demand different handling, and the wrong claim biases everything downstream.

Related terms

Learn next

Be Productive Using AI builds the everyday data habits that keep blanks from becoming biases: https://aiforbusinessandlife.com/academy/be-productive-using-ai. [C1]

Sources

[S12] Aalto University, Dictionary of Machine Learning; [S14] KNIME, Data Science Glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Missing state variable Next: Misuse detection