Figure 3.

Simulation results (for each sensitivity scenario, as the proportion of active arm deviations increases, each panel shows the evolution of the mean estimate of the time 3 treatment effect variance (over 1000 simulations) calculated in four ways: , Rubin's multiple‐imputation variance, from reference‐ or δ‐based MI; , information‐anchored variance ($\widehat{\mathbf{E}}[{\widehat{V}}_{\mathsf{anchored}}]$); , applying the primary analysis variance estimator in sensitivity scenarios; , variance when post‐deviation data are alternatively fully observed under the given scenario ($\widehat{\mathbf{E}}[{\widehat{V}}_{\mathsf{full},\mathsf{sensitivity}}]$)): (a) sensitivity scenario, copy reference, treatment effect 0; (b) sensitivity scenario, jump to reference, treatment effect 0; (c) sensitivity scenario, copy increments in reference, treatment effect 0; (d) sensitivity scenario, last mean carried forward, treatment effect 0; (e) sensitivity scenario, copy reference, treatment effect 1.0; (f) sensitivity scenario, jump to reference, treatment effect 1.0; (g) sensitivity scenario, copy increments in reference, treatment effect 1.0; (h) sensitivity scenario, last mean carried forward, treatment effect 1.0