A comparison of ARIMA Model, BP Neural Network Model and combined model in health policy evaluation:An empirical study of public hospitals pricing reform / 中国卫生政策研究

ABSTRACT

Objective:To study the effectiveness of different time series models in the prediction of financial data in public hospitals,with the aim of obtaining a more reliable counterfactual in health policy evaluation. Methods:ARI-MA model,BP neural network and their combination were used for the estimation and prediction of drug revenue and medical service revenue based on a dataset for the period from November,2011 to October,2016 for hospital X before and after Nanjing medical pricing reform. Root mean square error (RMSE) was used to estimate the model accuracy. Results:RMSE of drug revenue from the three models were 692.82,501.44 and 380.80,and of medical service were 184.04,215.63 and 168.65. The findings shows that the combination model was proved to be the most efficient one a-mong the three. The combined model was used to calculate the net loss of drug revenue which was estimated to be 120, 440 million,and the net increase of medical service was estimated to be 185,326 million after the reform,which was 1. 539 times of the drug loss. Conclusions:The revenue data of public hospitals are usually complex with a both linear and non-linear trend. The combination model of ARIMA and BP neural network could solve the problem for once with an acceptable accuracy. However,ARIMA model is simpler to operate as compared to other two models, and also more consistent with the forecasting trend,therefore ARIMA is also recommended in the evaluation for health policies.