Statistics Glossary → Rubin's rules (pooling imputation results)
Rubin's rules (pooling imputation results)concept
Rules that combine the results of analyses run separately on each of the m completed data sets from multiple imputation into a single estimate with the correct uncertainty.
Missing values are filled in m times by multiple imputation, each time with different plausible values, and the analysis is run on each data set. The pooled estimate is the mean of the m estimates (Q̄). The total variance has two components: the within-imputation variance Ū (the average of the squared standard errors from each analysis) and the between-imputation variance B (how much the m estimates differ from one another). Rubin's formula is T = Ū + (1 + 1/m)·B; the factor (1 + 1/m) reflects the extra uncertainty from using a finite number of imputations. Tests and confidence intervals use a t distribution whose degrees of freedom depend on the ratio of B to Ū.
The rules assume that the imputation model is compatible with the analysis model (the variables and interactions in the analysis are also in the imputation model) and that the data are missing at random (MAR). When the fraction of missing information (the share of total uncertainty due to differences between imputations) is high, more imputations are needed (e.g. 20 or more); the older advice of m = 5 is enough when missingness is low. The rules are designed for estimates that are roughly normally distributed (coefficients, means); correlations should be pooled on the Fisher z scale and odds ratios on the log scale, then back-transformed. Statistics such as R² or chi-square need special pooling methods.
An everyday example
A survey researcher creates 5 imputations for the 20% who did not answer the income question and estimates the coefficient of income on happiness in each data set: 0.30, 0.34, 0.28, 0.33 and 0.31. The pooled coefficient is 0.312. The standard error within each analysis is about 0.05, but once the fluctuation between imputations is taken into account the total standard error rises to about 0.056; using just one imputation would have understated the uncertainty.
In MerQur
The Multiple Imputation analysis pools its results with Rubin's rules (with the Barnard-Rubin degrees of freedom); the 'm (number of imputations)' option (2 to 100, default 5) sets the between-imputation variance and Rubin's total variance T = Ū + (1 + 1/m)B. The MNAR sensitivity option gives a Rubin-pooled sensitivity table for each δ shift.
Analyses where this concept appears
Related concepts
Multiple imputation by chained equations (FCS/MICE) and PMMMissing-data mechanisms (MCAR, MAR, MNAR)