Fig. 7.

Additional bias of the MI estimate of ${\beta }_{\mathit{YX}}$ when continuous exposure X is missing not at random, with missingness caused by Y and X, and the imputation model includes exposure Y and a predictor of missingness but not the missing values, Z. Simulation results shown when 50% of values are missing, varying the direct effect sizes ${\beta }_{\mathit{YX}}$, ${\beta }_{\mathit{RY}}$, ${\beta }_{\mathit{RX}}$, and ${\beta }_{\mathit{RZ}}$. The distribution of additional bias in each box-plot is averaged over the values of ${\beta }_{\mathit{RY}}$ and ${\beta }_{\mathit{RX}}$