Interpreting mediation effects on survival outcome using the mediation package in R I am interested in performing a mediation analysis using R's mediation package. I am fitting the following two models:
mod.y <- survreg(Surv(time, event) ~ M + X, data = d)

mod.m <- lm(M ~ X, data = d)

med.m <- mediate(mod.m, mod.y, treat = "X", mediator = "M")

summary(med.m)


And I get the following output:
Causal Mediation Analysis 

Quasi-Bayesian Confidence Intervals

                         Estimate 95% CI Lower 95% CI Upper p-value    
ACME (control)             0.5964       0.3212         0.90  <2e-16 ***
ACME (treated)             0.3778       0.2074         0.59  <2e-16 ***
ADE (control)             -6.7323      -8.3373        -5.10  <2e-16 ***
ADE (treated)             -6.9509      -8.6410        -5.26  <2e-16 ***
Total Effect              -6.3546      -7.9865        -4.74  <2e-16 ***
Prop. Mediated (control)  -0.0922      -0.1542        -0.05  <2e-16 ***
Prop. Mediated (treated)  -0.0576      -0.1049        -0.03  <2e-16 ***
ACME (average)             0.4871       0.2670         0.75  <2e-16 ***
ADE (average)             -6.8416      -8.4985        -5.17  <2e-16 ***
Prop. Mediated (average)  -0.0749      -0.1291        -0.04  <2e-16 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Sample Size Used: 4884 


Simulations: 1000 

I know how to interpret this table with continuous/binary responses in my outcome model, but am unsure as to what exactly these effects mean in the context of a survival model. For example, is ADE (control) the expected average direct effect on years lived in the control arm, had they switched to the treated arm? (i.e., are the units of the Estimate here time lived?).
 A: The code for mediate() indicates that, for a survreg outcome model, it estimates the linear-predictor values and then applies the inverse transformation associated with the model's distribution before it takes the differences used to report the various effects. The (default) Weibull survreg model is:
$$\log(T) = \eta + \sigma W $$
with $\eta$ the linear predictor, $\sigma$ the scale factor, and $W$ distributed as standard minimum extreme-value. The inverse transformation to return to units of $T$ in your situation is thus the exponential.
With mediation model reports based on (the exponential of) the linear predictor, $W = 0$ is implicitly assumed. (That's also what you get for "response" predictions from a survreg object; see Section 4.6.1 of the survival vignette.) So what's reported for a location-scale survival model like the above in general are effects in terms of the difference in estimated survival time when $W =0$.  The extreme-value distribution isn't symmetric, so the outcome for $W=0$ with a Weibull model isn't a "mean" or "median" survival time; it's about the 63rd percentile. If you use a survival distribution based on a symmetric $W$ distribution, like log-normal or log-logistic, the effects would be differences in estimated median survival.
