Tuesday, September 11
I just finished proofing what has got to be nearly the final version of Levente's book, the scintillatingly titled Corruption and Democratic Performance. I would like to share a brief passage:
The procedure that marries the advantages of producing unbiased estimates and introducing the appropriate level of uncertainty in the modeling while being highly efficient is multiple imputation. Multiple imputation (MI) requires the researcher to impute missing values several times, creating M number of independent complete datasets. Imputations can have different imputed values in all datasets, as they are simulated values that consider both the expected value for the missing item and the uncertainty. The imputation is a random draw from the plausible distribution of the item. These M datasets have to be analyzed independently, with the same analytical procedure and their results combined using a set of formulas which are collectively known as Rubin's rules. Simulation studies have shown that M ≥ 10 imputations produce sufficiently accurate results in longitudinal models. This is how I determined the number of imputations I used.
Tugs at the heartstrings, don't it?