<p>A Bayesian propensity score analysis for misclassified multinomial outcome is proposed to estimate causal effects in observational studies. To cope with the non-identifiability issue in modeling misclassification, an informative Dirichlet prior is placed on the parameters of misclassification probabilities. The proposed method is compared with a naive method that ignores misclassification in a simulation study. The results show that the proposed method outperforms the naive method, with smaller biases and the coverage probabilities closer to the given nominal value. It is also found that ignoring misclassification in the outcome leads to biased estimation of the treatment effect. We apply the proposed model for a study on estimating the religious impact on the choice of contraception method.</p>