Is mean imputation always appropriate?

No. Distribution shape, variable type, missingness and analytical purpose all matter.

Should the test set determine the fill value?

Not in an evaluation that is meant to test a transformation learned only from training data.

Does an imputed value become an observed fact?

No; its origin remains an estimate or specified substitute.

AI Glossary / Mathematics and optimization / Advanced

Data Imputation

Definition

Data imputation replaces missing values with estimates or specified substitutes so that incomplete records can be used in an analysis or machine-learning workflow. The replacement may come from a simple statistic, such as a training-set median, or from a model using other observed information. An imputed value is a constructed value, not a newly recovered observation or a guarantee of what was originally missing.

What it means

A blank field can mean that a sensor failed, a respondent skipped a question or a measurement did not apply. Before filling a gap, establish what the missing marker means and whether a numerical replacement is sensible.
Imputation can retain useful information from partially observed records. It also introduces assumptions. The goal is to support the intended analysis with a defensible treatment of missingness, rather than to make a spreadsheet appear complete. Preserve the original missingness information and the method used to generate replacements.

How it works

Simple methods use a column's observed mean, median or most frequent category, or a defined constant. Elaborate methods use relationships between features or information from similar records. The chosen rule must respect variable types and any relevant constraints.
For predictive evaluation, learn imputation rules using the training data and apply those fitted rules to held-out data. Within cross-validation, fit them separately within each training fold. A pipeline can enforce that separation.
Single imputation produces one completed dataset. Multiple imputation produces several plausible completed datasets and combines appropriate analysis results to reflect missing-value uncertainty under its assumptions. Repeatedly applying the same deterministic fill does not create that uncertainty representation.

Three practical examples

A customer model uses a training-set median for missing numeric account attributes. A separate indicator records which values were absent, allowing the analyst to distinguish a median-valued observation from a replacement.
An equipment-monitoring team estimates an absent sensor reading from related sensors. It checks whether those sensors are available at the moment a prediction is required; future measurements cannot silently enter a real-time workflow.
A survey analyst uses multiple imputation for incomplete responses under a statistical model. The analyst combines estimates and uncertainty appropriately instead of treating every completed copy as additional independent respondents.

Think of it this way

Think of restoring a missing tile in a mosaic. Nearby tiles can suggest a plausible color, but the replacement does not reveal the lost tile with certainty. Marking the restoration and documenting the reasoning preserves the distinction between what was observed and what was inferred. Different plausible restorations may support different interpretations of the picture.

Worked example

Suppose a training column contains 2, 4, 6 and one gap. Mean imputation uses the observed values: (2 + 4 + 6)/3 = 4. The completed training column becomes 2, 4, 6, 4.
Now consider two held-out records with values 100 and missing. Apply the fitted training rule: retain 100 and replace the missing value with 4. Do not recompute the rule from these held-out records. If someone combines all observed training and test values first, the mean becomes (2 + 4 + 6 + 100)/4 = 28. The held-out value has now changed the training transformation, violating the intended evaluation separation.
The example also shows a statistical effect of filling with the mean. The three observed training values have sample variance [(2 − 4)² + (4 − 4)² + (6 − 4)²]/2 = 4. Treating the four completed values as ordinary observations gives sample variance 8/3, approximately 2.67. The inserted central value reduces the computed spread without adding a genuine measurement.
Neither calculation tells us the missing value's actual value. A completed table can be convenient for a predictor while still being unsuitable for a naive uncertainty calculation that treats every cell as observed. These are separate uses and need separate justification.

Illustrative case study

In a fictional service company, an optional field records the number of previous support contacts. One export uses blank for an unanswered question, while another uses zero when the customer confirms having made no contacts.
A first draft of the preparation process converts both to zero. This makes customers with unknown histories look identical to customers known to have no previous contacts. The team instead preserves observed zeros, represents blanks as missing, and records a missingness indicator.
It compares a simple training-only imputation baseline with an alternative using available account information. Evaluation includes records with and without missing values, because an overall score could conceal poor performance on incomplete records. The team also checks whether the optional field will be collected consistently after deployment. Its final choice depends on evidence from the intended prediction task, not on which method produces the neatest-looking table.

Common misconceptions and mistakes

Imputation does not necessarily remove bias caused by systematic missingness. If omitted values differ from observed ones in ways the model does not capture, a sophisticated algorithm can still produce misleading replacements.
Zero is not a universal missing-value code. A missing outcome should not be filled and then treated as an independently observed evaluation label. Finally, single imputation is not equivalent to multiple imputation, and filling several columns once is not what the word multiple means in that statistical procedure.

Business use

Document which fields are imputed, why they are missing, what information the method uses and when that information becomes available. Keep the original data and fitted transformation reproducible. Compare downstream performance with baselines and examine results for incomplete records. If the intended output is a statistical estimate with uncertainty, check whether the analysis accounts for imputation rather than treating replacements as additional evidence.

Check your understanding

What fills the missing training value in the example?

Answer: Four, the observed training mean.

Why is 28 inappropriate for that fitted evaluation workflow?

Answer: It incorporates held-out information when learning the transformation.

Why retain a missingness indicator?

Answer: It distinguishes an absent original measurement from an observed value that happens to equal the replacement.

Related terms

Learn next

Study missing-data mechanisms and how they affect the assumptions behind an analysis. Then compare simple and model-based imputation inside a correctly separated validation workflow.

Sources

Aalto Dictionary of Machine Learning, data imputation and missing data: estimating absent values. https://aaltodictionaryofml.github.io/ADictML_Main.pdf Scikit-learn, Imputation of missing values: simple, multivariate and multiple imputation, and missingness indicators. https://scikit-learn.org/stable/modules/impute.html Scikit-learn, Common pitfalls and recommended practices, section 12.2: training-only fitting and preprocessing leakage. https://scikit-learn.org/stable/common_pitfalls.html The mean, variance and leakage calculations and fictional service-company case are original teaching illustrations.

Previous: Data hub Next: Data independence