The present study set out to develop an efficient and robust Kalman filter based statistical air quality model using Bayesian approach, with its application to the daily PM10 concentrations in Macau as practical verification. Two types of time-varying statistical air quality models, namely the TVAR(p) model and the TVAREX model were developed and the application of the Bayesian approach was made in two areas to enhance the efficiency and robustness of the developed model. First, the Bayesian approach was used to estimate each model’s process noise variance and the measurement noise variance, which are uncertain parameters affecting the state estimation performance of Kalman filter. The Bayesian approach was proved to be a good method in providing these estimations as well as their corresponding uncertainties. Second, the Bayesian approach was used to select the most plausible air quality model from the six proposed candidates by making a tradeoff between the prediction accuracy and the sensitivity to the modeling error. The method was found to be efficient and the TVAREX(1) model was found to be the most plausible model in the present study. Performance of the selected TVAREX(1) model was then compared against a well known approach: a multilayer perceptron (MLP) based air quality model. Under v the condition of same input variables, it was found that TVAREX(1) model was more efficient than the MLP based model in terms of the general modeling performance and the capability of capturing pollution episodes; hence the present proposal in developing an efficient and robust Kalman filter based statistical air quality model using Bayesian approach was successful. Better performance achieved by TVAREX(1) model led to a supposition that it was due to the dynamic nature of an air quality system. To validate this supposition, an effort was then made to turn the MLP model into adaptive by implementing the Kalman filter in its learning algorithm. The resulting time-varying multilayer perceptron (TVMLP) mode was found more efficient than the MLP based air quality models under the conditions of same input variables and the functional form; therefore confirming the dynamic nature of the air quality system.