Hello Drs Muthen, I have 1 binary outcome (n=89/2561), 4 mediators (2 binary, 2 continuous), and 1 continuous predictor. Also, I want indirect effects and bootstrapped CIs, & tests of model fit. As the default estimator is WLSMV (using a probit link) with theta paramaterisation additionally specified the model runs. Would a ML estimator with logit link and montecarlo integration be better to get predicted probabilities, but this won’t generate indirect effects (I now realise won’t work with binary mediators). What model would you recommend? Checking interpretation: So between cont x and cont m– 1-unit change in the cont IV is associated with a x-unit change in the cont MV?; between cont m and binary y – 1-unit change in the cont MV is associated with a x-unit change in the latent continuous variable underlying binary y? and binary m and binary y – where being unemployed (vs. employed) is associated with a 1-unit change in latent continuous variable underlying binary?...correct? Is it better have estimates unstandardized?...given different measures, & different distributions... So, indirect effects indicate for a 1-unit change on continuous x, there is an x-unit change on the latent continuous variable underlying binary y, through whichever mediators are specified. This would mean I can compare the relative strength of indirect effects with the same IV and DV, given they will be on the same scale?
With WLSMV, all regression are between continuous variables because the binary outcomes are represented by underlying continuous latent response variables. Therefore the interpretations are the same as if you had continuous outcomes.