Causal mediation analysis is widely used to study the causal mechanisms underlying an observed treatment-outcome relationship. Recently, machine-learning methods have been increasingly incorporated into mediation analysis, but their finite-sample performance remains less well understood. We conduct a simulation study comparing several commonly used approaches for causal mediation analysis in observational studies, including g-formula and efficient influence function (EIF)-based estimators with nuisance functions estimated using either parametric regression models or data-adaptive machine-learning methods, as well as a targeted minimum loss-based estimator (TMLE). We evaluate the estimators in terms of bias, Monte Carlo variability, and 95% confidence interval coverage for the natural indirect effect, natural direct effect, and mediation proportion (MP) across settings that vary sample size, nuisance-function complexity, pathway strength, and propensity-score overlap. Overall, the machine-learning EIF estimator provides the most favorable balance of bias and coverage across most simulation settings. TMLE also exhibits small bias but shows undercoverage in several settings. We further find that estimation of the MP can be substantially less stable when the sample size or the total effect is small. Finally, we apply the methods to NHANES 2011-2018 data to evaluate the extent to which the effect of higher weight-adjusted waist index on systolic blood pressure is mediated through hyperuricemia.