Modeling Individual Health Care Expenditures in China

In recent years, China has made great efforts in improving its health care system and reforming methods of payment. There is an increasing interest in modeling health care expenditures to provide a basis of policy analysis and design for purposes of payment system reform. It is well-known that health expenditures have special distributional characteristics, including a restricted range, a “spike” of zero values, and skewness, which may lead Ordinary Least Square (OLS) estimates to be biased and inefficient. A large number of studies have examined spending patterns in the US, Europe and other countries with well-developed provider information systems, but studies using supply side data rather than survey-based data are rare in countries with less well developed insurance and claims data. Our study is the first to use large sample data on insured individuals in China to examine such issues in the Chinese system.

We examine a variety of econometric approaches to model health care expenditures, using a rich dataset covering over 280,000 insured individuals in China from 2006 to 2014. The data set contains detailed information on spending and diagnosis, as well as individual characteristics, such as age, gender and education. In this study, we model expenditures in the following three dimensions.

The first part of our paper considers the regression methodology. We compare OLS model to two common alternative approaches: two-part model and the Generalized Linear Model (GLM). We find that OLS model has a similar predictive power comparing to the other two models when using large sample data.

Second, we consider the proper control variables in the regressions. We will include age, gender and diagnosis information in the basic model. We will try several diagnosis classification systems to choose the most suitable method for Chinese residents, including MDCs and the conventional HHS-HCC risk adjustment models applied in the Health Insurance Exchanges in the US.  Similar as using data in the US, we find that adding diagnosis variables in the model significantly improve its predictive power. In addition, for a subset of individuals, we have their survey information merged to the claims data, and we explore control variables, such as marital status and education, to analyze the fit of those models.  We do not find these survey-based variables to significantly change model’s predictability.

Third and finally, we will consider the appropriate “unit of payment” to characterize our outcome variables. We categorize total spending into outpatient, inpatient and prescription drug and consider model performance for different groupings of health care expenditures. We find that outpatient spending is easier to be predicted than inpatient spending.

We aim at searching for the “best” model in predicting health care costs in the context of China’s system. Our work is the first study in analyzing econometric models of medical costs for Chinese residents. Our findings will help policymakers to better understand the determinants of health care spending, leading to a solid foundation for future payment system reform in China.