Fig. 4.

Bias in the treatment group mean estimate after MI with a misspecified imputation model. Missing data were generated completely at random (MCAR in red) or at random (weak MAR in green or strong MAR in blue) at a rate of 40%. From top to bottom, results are presented in increasing order of cluster sizes m within an increasing order of number of clusters k. Filled circles indicate PMM methods while open circles indicate parametric (NORM) methods