Mediation analysis investigates how certain variables mediate the effect of predictors on outcome variables. Existing studies of mediation have been limited to normal theory maximum likelihood (ML) or least squares with normally distributed data. Because real data in psychology are seldom normally distributed, classic methods often lead to biased or inefficient parameter estimates, which make the result of mediation analysis misleading. To circumvent these problems, this dissertation proposes to utilize robust methods for parameter estimation, and then to evaluate the mediated effect based on robust estimates. Theoretical details of robust procedures for mediation analysis are provided. Their properties are investigated via simulation studies, and results suggest that (a) the most commonly used Sobel z method relying on ML point and information-based standard error (SE) estimates has the worst overall performance; (b) the distribution of the product confidence interval based on robust point and sandwich-type SE estimates has the best overall performance. Applications of robust procedures are illustrated using real data examples, which show that robust procedures provide more reliable results.