BACKGROUND/RATIONALE:
Appropriate inferences on veterans' health care costs require a regression model that can address several problems that typically arise in cost data - many observations with zero costs, skewness and heteroscedasticity - that are not fully accounted for in current (parametric) methods. As a result, these problems may generate biased estimates of Veterans' health care costs and wrong inferences about the distribution of veterans' health care costs. For example, in our two real example data sets, one from a V A study on V A Community Based Outpatient Clinics (CBOCs) and another from a VA study on 30 common chronic diseases, total inpatient health care costs share all the characteristics mentioned above.
OBJECTIVE(S):
The study objective was to develop, test and validate new regression methods for the analysis of V A health care costs that allows one to account for skewness, heteroscedasticity and preponderance of zero values in VA cost data to get more accurate cost prediction.
METHODS:
We used mathematical and statistical tools to develop new statistical models for predicting the health care cost of a patient with given covariates. We used simulation to assess the adequacy of the proposed prediction models.
FINDINGS/RESULTS:
We developed three new statistical methods for the analysis of health care costs. We also performed the empirical study to assess the bias and efficiency of the eight approaches to modeling the second part of the two-part model and examined bias reduction obtained by moving from a simple linear model to a more completely specified model that reduces heteroskedasticity and non-linearity. The newly proposed methods include: (1) confidence intervals for the mean ofthe zero-inflated lognormal distribution., (2) a semi parametric heteroscedastic two-part transformation regression model to deal with skewed costs with additional zeros, and (3) a kernel-smoothing based nonparametric test for assessing the goodness-of-fit of a variance function in a general heteroscedastic regression model.
Our approximate generalized pivotal approach to account for skewness, heteroscedasticity and zero values performs better than other methods. Using the newly developed method we re evaluated data about costs associated with CBOCs which addressed previous methodologically issues. Our findings suggest that costs for areas with CBOCs are similar to areas without CBOCs.
IMPACT:
The study results provide better regression methods for predicting health care costs of a patient given his/her characteristics that can handle the special features of health care costs. Our new methods can be used by health policy and health economics researchers to develop (1) better risk adjustment models for capitation payments, and (2) better case-mix adjustment profiling models for cost performance measures.
PUBLICATIONS:
Journal Articles
- Zhou XH, Lin H. Semi-parametric maximum likelihood estimates for ROC curves of continuous-scale tests. Statistics in Medicine. 2008; 27(25): 5271-90.
- Zhou XH, Liang H. Nonparametric single-index regression models for health care charges with zero values. Computational Statistics and Data Analysis. 2006; 50: 1378-1390.
