Traditional mediation models may either overlook the multilevel structure of the data set, or the temporal sequence of X, M, and Y for causation to take place. In the dissertation, extending from the cross-lagged panel models (CLPMs) in Maxwell et al. [2011], I proposed the multilevel autoregressive mediation models (MAMMs) by allowing all the coefficients to be random across individuals. In addition, Level-2 covariates can be included to partly explain the inter-individual differences of mediation effects. Given the complexity of the model structure, Bayesian estimation was used. To illustrate the estimation procedure, both the CLPM and MAMM were fitted to a daily diary data set. It was found that the two models yielded different statistical conclusions regarding the average mediation effect. In light of this, a simulation study was then conducted to examine the estimation accuracy of Bayesian estimation for MAMMs and consequences of model mis-specifications. Factors considered include sample sizes (N), numbers of time points (T), fixed indirect and direct effect sizes, and Level-2 variances and covariances. Results indicated that estimation of fixed effects for the indirect effect components (a and b) and the conditional indirect effects with a Level-2 covariate were accurate for N ≥ 50 and T ≥ 5. In terms of the estimates for Level-2 variances and covariances, they were accurate provided a sufficiently large N and T (e.g., N > 400 and T ≥ 50), and less accurate when N and T became smaller. Estimates of the average mediation effect were mostly accurate with moderately large N × T (e.g., 1,000). Furthermore, with sufficiently large N and T (e.g., N = 400 and T = 50), estimates of a, b and the average mediation effect were accurate when Level-2 variances were zero and MAMMs were fitted, and they could be inaccurate when random effects existed and CLPM was fitted; in addition, DIC can be used to select the correct MAMM with non-zero Level-2 variances. Limitations and future directions were discussed.