Is mean filling model based imputation?

Under the broad OECD terminology, yes.

Does multiple imputation recover the true missing values?

No. It represents plausible completions under assumptions and supports uncertainty-aware analysis.

Should the same imputation procedure serve every use?

Not necessarily. Prediction, statistical inference, and operational reporting can impose different information and modeling requirements.

AI Glossary / Statistics and probability / Intermediate

Model based imputation

Definition

Model based imputation replaces missing values using an estimated or specified relationship in the data. Methods range from simple mean or median substitution to regression and more elaborate conditional models. The replacement is an estimate or a plausible draw, not a recovered observation. Its usefulness depends on the model, the missingness process, and the purpose of the subsequent analysis.

What it means

When a field is blank, other recorded information may help estimate what could plausibly belong there. A model can use those relationships to fill the gap while retaining the record for analysis.
The OECD glossary uses this term broadly, including averages, medians, and regression equations. Modern applications distinguish deterministic replacements from stochastic draws and multiple imputation because these approaches handle uncertainty differently. Merely completing a table does not establish that its statistical properties have been preserved.

How it works

First distinguish genuinely missing measurements from structural absences, inapplicable questions, and valid zero values. Examine missingness patterns and the data-collection process. The observed data alone generally cannot prove that missingness assumptions are correct.
Choose predictors and an imputation model appropriate to the variable's type, range, relationships, and analysis purpose. A continuous measurement may require a different model from a category, count, or bounded proportion. Grouping, time dependence, and interactions can matter.
Deterministic regression imputation substitutes a fitted conditional prediction. It can preserve some mean relationships but removes residual variation from the replaced observations. Stochastic imputation adds appropriate variability; a single stochastic completion still does not automatically represent all uncertainty about estimated parameters and missing values.
Proper multiple imputation creates several plausible completed datasets using a procedure that reflects relevant uncertainty. Analyze each dataset and combine estimates using suitable pooling rules. For common estimands, uncertainty includes both within-imputation and between-imputation contributions. Averaging the completed tables and treating that average as observed data defeats this purpose.
For predictive machine-learning evaluation, learn preprocessing from training data within each validation split and use only information available at prediction time. Statistical inference may appropriately include an observed outcome in an imputation model; that differs from using unavailable future outcomes in a deployed prediction pipeline.
Retain missingness indicators and imputation provenance. Inspect plausibility, distributions, relationships, and sensitivity to assumptions. If missing values depend on unobserved quantities in important ways, sensitivity analyses or explicit missingness models may be needed.

Three practical examples

Survey processing: estimate a missing item from related responses while retaining flags that identify imputed values.
Operational analytics: fill incomplete sensor readings using a model that respects time ordering and the measurement's feasible range.
AI preprocessing: learn an imputation transformation from training data and apply it consistently during validation and deployment.

Think of it this way

Think of restoring missing tiles in a mosaic. Nearby patterns suggest plausible replacements, but the new tiles are still inferred. A restoration that looks smooth can conceal uncertainty or erase genuine variation, so the restored areas should remain identifiable.

Worked example

Consider an original dataset with observed values 4, 8, and 12, plus one missing value. Mean imputation replaces the missing entry with 8.
The observed mean is 8, and the completed-data mean remains 8. However, the observed sample variance is [(4 - 8)^2 + (8 - 8)^2 + (12 - 8)^2]/2 = 16.
After adding the imputed 8, the same squared deviations total 32 but the sample-variance denominator becomes 3, giving 32/3, approximately 10.67. Treating the imputation as an ordinary observation mechanically reduces this descriptive variance; it has not supplied a new independently measured value.
Alternatively, suppose a separately fitted regression predicts the missing quantity as 2 + 3x and its recorded predictor is x = 4. Deterministic regression imputation gives 14, producing a completed mean of 9.5.
The different answers reflect different assumptions and information. Neither calculation identifies the true missing value or establishes a valid standard error for later inference. The regression coefficients here are stipulated, not fitted to the three displayed outcomes.

Illustrative case study

Fictional case: Brook Services analyzes customer response times with missing values concentrated in one reporting channel. Analysts initially consider filling every gap with the overall mean.
They investigate the channel's recording process, compare plausible conditional models, and preserve imputation flags. For an inferential report, they use a justified multiple-imputation procedure and examine sensitivity to assumptions about the missing cases. The report explains how conclusions depend on those assumptions rather than presenting the completed values as measurements.

Common misconceptions and mistakes

Imputation does not make missingness harmless. A sophisticated model can still be wrong about why values are absent or how missing cases differ from observed ones.
Do not confuse multiple regression imputation with multiple imputation: the former may use several predictors but still create only one completed dataset. Repeating a deterministic procedure without meaningful variation also does not capture missing-data uncertainty.
A good reconstruction score on artificially hidden observations may not transfer to real missing values if the masking mechanism differs. Likewise, plausible marginal distributions do not guarantee preservation of correlations or subgroup patterns.
Never silently overwrite the distinction between observed and imputed information.

Business use

Businesses can use model based imputation to support analyses that would otherwise discard incomplete records, but they should evaluate whether the assumptions fit the data source and decision. Track missingness rates by group and time, document the method, and assess downstream effects. Improving collection practices may provide more reliable value than increasing imputation complexity indefinitely.

Check your understanding

Why did sample variance shrink after mean filling?

Answer: An added value at the mean increased the denominator without adding squared deviation.

What must remain documented?

Answer: Which entries were missing, how replacements were generated, and the assumptions governing their use.

Related terms

Missing data
Data leakage
Sensitivity analysis
mean imputation
regression imputation
multiple imputation
missing at random
missing not at random

Learn next

compare deterministic and multiple imputation on a small incomplete dataset, checking both estimates and their uncertainty rather than only whether blanks disappear.

Sources

Sources: OECD Glossary of Statistical Terms, model based imputation, printed page 337: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf . Van Buuren, Flexible Imputation of Missing Data: https://stefvanbuuren.name/fimd/sec-nutshell.html and https://stefvanbuuren.name/fimd/sec-linearnormal.html . Scikit-learn: https://scikit-learn.org/stable/modules/impute.html . Relevant source sections were checked. Calculations and fictional case are original.

Previous: Model-based clustering Next: Model-based reasoning