In a mediation model the effect of a dependent variable (DV) on an outcome is (partially) due to the DV's effect on one or multiple mediator(s) that consequently have an effect on the outcome. The use of such models as the theoretical background guiding empirical studies is widespread. Mediation models are fundamentally causal models that specify causal sequences. Unfortunately, the necessary causal assumptions are in practice often violated. In the current paper, we discuss possible improvements of causal mediation analyses, and highlight some potential pitfalls. We discuss the benefits gained by analyzing indirect effect between latent variables specified with measurement models. The validity of statistical findings can also be improved by using experimental designs. We discuss the cross-over design and the cross-over encouragement design, and how they can help improve causal conclusions. We also discuss recent advances on sensitivity analyses in the context of mediation models. Specifically, we explain how this analysis can be used to argue for the severity of unobserved confounding. Lastly, we discuss the practice of reversing the direction of the arrow between variables in a mediation model. We argue that if such reversals result in equivalent models, this practice cannot be recommended.