FIGURE 1.

Plots of the true influence function ${\tilde{\mathcal{l}}}_0\left(X_i,Z_i,U_i,{\text{Δ}}_i\right)$ against the error-prone version ${\tilde{\mathcal{l}}}_0^{\star }$ with the variables subject to measurement error noted in the graph subtitle. For example, (A) displays ${\tilde{\mathcal{l}}}_0\left(X_i,Z_i,U_i,{\text{Δ}}_i\right)$ against ${\tilde{\mathcal{l}}}_0\left(X_i^{\star },Z_i,U_i,{\text{Δ}}_i\right)$. Univariate and normally distributed X and Z were generated. Survival times were generated from an exponential distribution with rate ${\lambda }_0\exp \left({\beta }_{X}X+{\beta }_{Z}Z\right)$, where ${\lambda }_0=0.1,{\beta }_X=\log (1.5)$, and ${\beta }_Z=\log (0.5)$, with 90% independent censoring. The error was generated as $X^{\star }=0.2+X-0.1Z-0.4\text{Δ}+0.25U+ϵ,U^{\star }=U+{\sigma }_v\cdot 3-0.2X-1.05Z+v$, and ${\text{Δ}}^{\star }=\text{Bernoulli}(\text{expit}(-1.1+3\text{Δ}-0.3X-0.2U+0.1Z))$, where $(ϵ,v)$ were normally distributed with $\left({\mu }_{ϵ},{\mu }_v\right)=(0,0)$, variances $\left({\sigma }_{ϵ}^2,{\sigma }_{\nu }^2\right)=(0.5,0.5)$, and ${\rho }_{ϵ,v}=0.5$