Home Statistical Dictionary Imputation
Imputation
The general practice of filling in missing survey or dataset values with substituted, plausible values, so that standard complete-data analysis methods can be applied without discarding entire cases due to partial nonresponse.
In Plain English
Imputation is the umbrella term for any method of filling in the blanks when a dataset has missing values, a survey respondent skipped the income question, a sensor failed to record one reading, a field was left blank on a form. Rather than throwing out that entire respondent's data (which wastes everything they did answer, and can introduce its own bias if the people who skip a question differ systematically from those who don't), imputation substitutes a plausible value for the gap, using anything from a simple overall average to a sophisticated statistical model, so the dataset can be analyzed as if it were complete.
Definition
Imputation is the process of substituting plausible values for missing entries in a dataset, most commonly item nonresponse in a survey (a specific question left unanswered, as opposed to unit nonresponse, an entire respondent not participating at all), so that standard complete-data statistical methods can be applied to the resulting filled-in dataset without discarding partially observed cases. Imputation methods range widely in sophistication, from simple deterministic approaches like mean imputation and hot-deck imputation, to model-based approaches like regression imputation, to modern approaches like multiple imputation that explicitly propagate the uncertainty introduced by the imputation process itself into the final analysis; the appropriateness and validity of any imputation method depends critically on the underlying missing-data mechanism (whether data are missing completely at random, missing at random, or missing not at random), a mismatch between the imputation method's assumptions and the true missingness mechanism is a common source of bias in imputed analyses.
Properties
- Imputation addresses item nonresponse specifically, missing individual data fields within an otherwise-participating case, this is distinct from unit nonresponse, where an entire sampled unit fails to participate at all and is typically handled instead through nonresponse weighting adjustments rather than imputation.
- Simple, single-value imputation methods (mean imputation, single regression imputation) share a common flaw, treating the imputed value as if it were genuinely observed data understates the true uncertainty in any subsequent analysis, standard errors computed on singly imputed data are generally too small, this is exactly the problem multiple imputation was developed to solve.
- No imputation method can correct for missingness that depends on the unobserved value itself in a way unrelated to any observed variables (missing not at random), all standard imputation approaches implicitly or explicitly assume some form of missing-at-random mechanism, and can produce biased results if that assumption is violated.
At a Glance