Imputation

Imputation is a statistical technique for handling missing data by replacing absent values with estimated substitutes. Methods range from simple approaches such as mean or median replacement to sophisticated techniques such as multiple imputation, which generates several plausible datasets to account for the uncertainty introduced by the missing values.

Missing data are nearly universal in research. Clinical trials lose participants to follow up, survey respondents skip questions, and sensors malfunction. How missing data are handled can dramatically affect study conclusions. Naive approaches such as complete case analysis (analyzing only participants with full data) reduce statistical power and can introduce bias if the missing data are not missing completely at random. Regulatory agencies, journal editors, and systematic reviewers increasingly expect researchers to describe their missing data patterns, justify their imputation strategy, and conduct sensitivity analyses to test the influence of missing data assumptions on results.

The choice of imputation method depends on the mechanism generating the missing data. Three mechanisms are recognized. Data missing completely at random (MCAR) are absent for reasons unrelated to any observed or unobserved variable. Data missing at random (MAR) are absent for reasons related to observed variables but not to the missing values themselves, after conditioning on those observed variables. Data missing not at random (MNAR) are absent for reasons related to the unobserved values themselves, which is the most problematic scenario.

Simple imputation methods include mean imputation (replacing missing values with the variable's mean), last observation carried forward (LOCF, using the most recent observed value), and regression imputation (predicting missing values from other variables using a regression model). These methods have significant limitations: mean imputation reduces variance and distorts relationships between variables, LOCF assumes no change over time, and single regression imputation underestimates standard errors by treating imputed values as if they were observed.

Multiple imputation (MI), a widely recommended approach, addresses these limitations through a three step process. First, the imputation step generates M complete datasets (typically 20 to 50) by drawing plausible values from the predictive distribution of the missing data, conditional on observed data. Second, the analysis step applies the planned statistical analysis to each of the M datasets separately. Third, the pooling step combines the M sets of results using Rubin's rules, which account for both within imputation and between imputation variability. The result is a single set of estimates with standard errors that reflect the additional uncertainty due to missing data, provided the data are MAR and the imputation model is appropriate.

For MNAR data, sensitivity analyses are essential. Pattern mixture models and selection models allow researchers to explore how results change under different assumptions about why data are missing. Tipping point analyses ask how extreme the missing values would need to be to change the study's conclusions.

Use imputation when missing data exceed a trivial fraction of the dataset (often more than 5%) and when complete case analysis would substantially reduce sample size or introduce bias. Multiple imputation is generally preferred when data are plausibly MAR, because it preserves variability and, with a well specified imputation model, produces valid standard errors. Simple imputation methods are acceptable only when the fraction of missing data is very small and when sensitivity analyses confirm that results are insensitive to the imputation approach. MNAR sensitivity analyses should be conducted whenever the assumption that data are MAR is questionable.

A frequent error is applying mean imputation without recognizing that it artificially narrows the distribution and distorts associations with other variables, typically weakening correlations toward zero. Another mistake is using too few imputed datasets in multiple imputation; early guidance suggesting 5 to 10 datasets has been replaced by recommendations for 20 or more, particularly when the fraction of missing data is substantial. Researchers also sometimes fail to include the outcome variable and key predictors in the imputation model, which can bias the analysis. Including variables in the imputation model that are not in the analysis model (auxiliary variables) generally improves the quality of imputations and is recommended when those variables predict the missing values. Finally, not reporting the amount and pattern of missing data or the imputation method used is a transparency failure that makes it difficult for readers to assess the validity of the results.

In a multicenter RCT of 600 patients evaluating a new treatment for chronic pain, 18% of participants are missing their primary outcome (pain score at 6 months) due to dropout. The statistical team examines baseline characteristics of completers versus dropouts and finds that dropouts have higher baseline pain scores, which shows the data are not MCAR; MAR given the baseline variables is a plausible working assumption, although observed data cannot rule out MNAR. They implement multiple imputation with 30 imputed datasets, using baseline pain, age, sex, treatment group, and site as predictors in the imputation model. The pooled treatment effect (mean difference: 1.4 points, 95% CI: 0.6 to 2.2) is slightly attenuated compared to the complete case analysis (mean difference: 1.7), and a tipping point analysis shows that results remain significant unless dropouts in the treatment group had pain scores about 3.3 points or more worse than predicted (if about 18% of that group is missing, each point of difference lowers the treatment effect by about 0.18 points, so the lower confidence limit of 0.6 reaches zero at about 3.3 points).