I have a path model which included a combination of continuous, binary, and ordered categorical endogenous variables. All variables are observed. Two of the mediators are biomarkers that are highly skewed (insulin and IL6). I have run 4 different models: model 1, insulin and il6 are logged, estimator is WLSMV; model 2, insulin and il6 are not logged, estimator is WLSMV; Model 3, insulin and il6 are logged, estimator is MLR, and model 4, insulin and il6 are not logged, estimator is MLR. Most of the relationships don't change across models, except for two of the paths between two mediators. In Model 1 (logged mediators, WLSMV), the relationship between a binary mediator and il6 and another continuous mediator are NOT significant (p~.13). In Model 2 (mediators not logged, WLSMV), path from binary mediator to IL6 is significant, while path to the other continuous mediator is still not significant). In Models 3 and 4(MLR), those same two paths are significant (p=.006). These two paths are a big part of the story and I am not sure which results to run. I know WLSMV and MLR are the best options for categorical outcomes, but which one do I trust here? There is no missing data. Thank you!