Hello @Pau_lina,
The SEM platform can be used to specify mediation models with or without latent variables (LVs). If you don't have LVs, then the 2-min video in this link shows how to use the Model Shortcuts menu to specify the model. If you do have LVs, then you have to specify the model using the From and To Lists (that is, specify the LVs first, then link them with the single and double headed arrow buttons). Note that the main red triangle menu offers an "Inference > Bootstrap" option to obtain bias-corrected confidence intervals for your indirect effect. This is in addition to the indirect and total effects you can get in the model-specific red triangle menu (those have standard errors that use the delta method). Moreover, if you're interested in testing the interaction between predictor and mediator (aka "causal mediation"), then you can take advantage of a cool add-in that was developed for this purpose. The add-in produces a very nice Johnson-Neyman plot to probe the interaction (if it's significant), in addition to other useful output.
The key issue is modeling your time-to-event outcome. One approach for this is to create a new variable that takes the log of your outcome (right-click on the header of the column in your data table, go to New Formula Column > Log), and then use that in the SEM platform. This would be similar to assuming a lognormal distribution of the outcome and using accelerated life testing. This would also make it possible to use the SEM platform, as we don't otherwise have special handling of time-to-event data yet.
HTH!
~Laura
Laura C-S
