Is between-imputation variance the whole standard error?

No; total variance also includes within-imputation uncertainty and the finite-M adjustment under Rubin's rules.

Can the same formula be used for every imputation method?

No; the framework and procedure must justify it.

AI Glossary / Statistics and probability / Intermediate

Imputation variance

Definition

Imputation variance is the component of an estimate's variability associated with replacing missing or unusable observations by imputed values. Its exact definition and estimation depend on the imputation procedure and inferential framework. In multiple imputation, variation between estimates from different completed datasets helps represent uncertainty about the missing information and is combined with within-dataset variance.

What it means

Filling a blank cell makes a table look complete, but it does not make the replacement a known observation. Another plausible completion could produce a different mean, regression coefficient, or other result.
Imputation variance concerns that uncertainty in the estimate, not simply the spread of the filled-in values. It is also different from imputation bias: a method can produce stable replacements that are systematically wrong, or variable replacements under a well-justified model.

How it works

Identify the target estimate and the missing-data procedure. Survey imputation, random donor selection, deterministic replacement, and Bayesian multiple imputation have different variance properties. There is no single formula that applies unchanged to every completed dataset.
In proper multiple imputation, create M plausible completed datasets that reflect relevant uncertainty under the imputation model. Analyze each using the same substantive procedure. Retain each estimate Q-m and its complete-data variance estimate U-m.
The pooled point estimate is Q-bar, the average of the M estimates. The within-imputation variance U-bar is the average of the U-m values. It represents the uncertainty that the completed-data analyses would report conditional on each completed dataset.
The between-imputation variance is B = sum of (Q-m - Q-bar)^2 divided by M minus 1. Under the usual Rubin combination rules, the total variance estimate is T = U-bar + (1 + 1/M)*B. The pooled standard error is the square root of T, not the average of the separate standard errors.
The additional B/M term reflects finite-imputation simulation variability. Increasing M reduces that part and improves stability of the estimated components. It does not eliminate the underlying uncertainty represented by B or create the information that genuinely observed values would have supplied.
Rubin's rules require suitable imputation and analysis procedures. Compatibility with the substantive model, appropriate treatment of parameter uncertainty, and the relevant missingness assumptions matter. Special analyses or unconventional imputation schemes can require different variance methods.
For complex surveys, account for sampling weights, clustering, stratification, and the imputation procedure together. A completed-data variance calculation that ignores the sample design may remain wrong even after a between-imputation term is added.
Deterministic imputation can give identical completed datasets when rerun, yet uncertainty about the missing values still exists. Zero variability across identical reruns does not establish zero missing-information uncertainty. Repeating a deterministic fill is not equivalent to proper multiple imputation.
Report how variance components were obtained and assess sensitivity to plausible missing-data models. Imputation variance generally does not cover every uncertainty about model choice, nonignorable missingness, measurement error, or deployment shift. A precise-looking standard error can still be conditional on questionable assumptions.

Three practical examples

An analyst combines regression results across multiply imputed customer datasets instead of reporting the standard error from only one completion.
A survey team estimates uncertainty that reflects both its sampling design and donor-imputation procedure.
An AI evaluation team examines whether missing labels materially affect uncertainty in an estimated performance measure.

Think of it this way

Think of completing several plausible versions of an unfinished puzzle and measuring the same feature in each. Each completed puzzle has its own measurement uncertainty, and differences between plausible completions add another source. A single neat completion hides that second source.

Worked example

Suppose four proper imputations produce estimates 10, 12, 11, and 13 for the same scalar quantity. Each completed-data analysis reports variance 4. The pooled estimate is (10 + 12 + 11 + 13)/4 = 11.5, and U-bar = 4.
The squared deviations from 11.5 are 2.25, 0.25, 0.25, and 2.25. Their sum is 5, so B = 5/(4 - 1) = 5/3.
Rubin's total variance is T = 4 + (1 + 1/4)*(5/3) = 73/12, approximately 6.083333. The pooled standard error is sqrt(73/12), approximately 2.466441. Using only the within-imputation variance would give standard error 2 and miss the additional component.
The finite-M contribution is B/M = 5/12. Even if that simulation contribution became negligible with many imputations, the limiting expression would still include U-bar + B. Four imputations are used only to make the arithmetic transparent, not as a recommended number for reliable practice. An interval also needs an appropriate reference distribution and degrees of freedom.

Illustrative case study

In this fictional case, Elm Insights fills missing survey answers and reports an average with a narrow interval from one completed dataset. A statistician repeats the analysis using a justified multiple-imputation model and pools both estimates and variances. The point estimate changes little, but the uncertainty increases. Management revises its confidence in a small segment difference and asks whether additional data collection would be more useful than simply generating more imputations.

Common misconceptions and mistakes

The variance of imputed cell values is not the same as imputation variance of a target estimate. Averaging standard errors is not Rubin pooling. Stacking completed datasets as if they contained new independent respondents inflates apparent information.
More imputations reduce Monte Carlo error but do not remove missing-data uncertainty. Agreement across imputation models does not prove their shared assumptions correct. A small variance component also does not establish absence of imputation bias.

Business use

This concept helps teams avoid overstating confidence in reports built from incomplete records. Preserve missingness indicators, document the imputation model, and report uncertainty appropriate to the final analysis. When decisions hinge on small differences, compare the value of obtaining missing observations with the limited benefit of additional computational imputations.

Check your understanding

Why is the example's standard error greater than two?

Answer: The pooled variance includes variation across plausible completed-data estimates.

What does increasing the number of imputations mainly reduce?

Answer: Simulation error from using a finite number of completions, not the information lost through missing observations.

Related terms

Multiple imputation
Rubin's rules
missing information
within-imputation variance
between-imputation variance
survey variance

Learn next

Learn missing-data mechanisms and proper multiple imputation. Then study pooling rules, small-sample inference, complex survey variance, and sensitivity to assumptions about unobserved values.

Missing data

Sources

OECD Glossary of Statistical Terms, Imputation variance: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf Stata Multiple-Imputation Reference Manual, mi estimate methods and formulas: https://www.stata.com/manuals16/mi.pdf

Previous: Impulse response weights Next: In-crowd algorithm