Figure 2.

Absolute bias of the treatment effect estimate at k = 0 (left column (panels A, D, and G)), k = 1 (middle column (panels B, E, and H)), and k = 2 (right column (panels C, F, and I)) on full data and following the use of different missing-data approaches under missing completely at random (top row (panels A–C)), constant (middle row (panels D–F)), and differential (bottom row (panels G–I)) missingness mechanisms (n = 10,000; 40% of missing data). For multiple imputation, 10 imputed data sets were generated. CC, complete cases; IPMW, inverse-probability-of-missingness weighting; LOCF, last observation carried forward; MI, multiple imputation; MPA, missing pattern approach.