Overimputation

Overimputation is a term used in statistics and data analysis to describe the excessive use of imputed values or treating imputed data as if they were observed measurements. It encompasses situations where imputed values are given undue influence in subsequent analyses, or where the uncertainty associated with the imputations is not properly accounted for in inference.

In practice, overimputation can arise in multiple imputation workflows when analysts analyze imputed data without appropriately incorporating the variability across imputations. If the analysis ignores Rubin’s rules for combining results or treats single imputed datasets as if they were fully observed, the estimated precision can be overstated and parameter estimates can become biased. Overimputation can also occur when the imputation model is overly strong or misspecified, leading to imputed values that reflect model assumptions more than the underlying data, thereby distorting relationships among variables.

Consequences of overimputation include underestimation of standard errors, overly narrow confidence intervals, and biased estimates, particularly for parameters sensitive to the distribution or missingness mechanism. The issue is more pronounced when the missing data mechanism is not missing completely at random or missing at random and the imputation model fails to capture relevant patterns.

Mitigation involves proper use of multiple imputation, including fitting congenial imputation and analysis models, pooling results with Rubin’s rules, conducting diagnostics and sensitivity analyses for missingness mechanisms, and clearly reporting imputation methodology and uncertainty. See also: multiple imputation, missing data mechanism, Rubin’s rules.