THE PREDICTABILITY OF RISK ADJUSTMENT MODEL

• Main Authors: Wender Lin, 林文德
• Other Authors: Chih-Liang Yaung
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/40185948920357524828

博士 === 國立臺灣大學 === 衛生政策與管理研究所 === 91 === Objectives The objectives of this study are as followings: 1. To develop a Taiwan’s version of inpatient-diagnosis-based risk adjustment model and to evaluate its predictability; 2. To estimate the maximal predictability on individual’s future health expenditure; 3. To explore the influence of time on the predictability of risk-adjusted models. Material and Methods 1.Using a 2-percent random sample of 371,620 NHI ( National Health Insurance) enrollees, the author developed a Taiwan’s version of principal inpatient diagnosis cost groups (TPIPDCG) from 1996 claim records to predict individual’s expenditure in 1997. Weighted least square regression models were built up in an estimation sample (two thirds of the study sample), and were cross-validated in a validation sample (the remaining one third of the study sample). Predictive R2 and predictive ratios were used to evaluate model’s predictability. 2.151,315 subjects continuously enrolled in NHI from 1996 to 2000 were selected as study sample for estimation of maximal predictability. Variance component analysis was applied to estimate the ratio of within-individual variance to total variance as the measure of maximal predictability. 3.Using sample from method 2, multiple year risk-adjusted models were established by adding previous years’ risk adjusters into regression model to predict individual’s expenditure in 2000. Besides, the predictability of models to predict average expenditure in future 1 to 4 year were also estimated by averaging out the predicted expenditure obtained from year to year models. Results 1.Only 7.88% of the 1st study sample can be classified into one of 16 TPIPDCGs. Combined with demographic variables which alone can explain 3.7% of variation in individual future expenditure, risk adjustment model based on TPIPDCGs can explain 12.2% of expenditure variation. 2.For the 2nd study sample, the maximal predictability was estimated to be 49.2% for individual’s total health expenditure, 73.3% for outpatient expenditure, and 17.5% for inpatient expenditure. 3.Adding TPIPDCGs from previous 1 year, 2 year, 3 year and 4 year to predict individual’s expenditure in 2000, the predictability can improve to 15.0%, 17.5%, 19.5%, and 21.0%, respectively. Prolonging the period of predicting, that is, predicting individual’s average expenditure for future 1 year, 2 year, 3 year, and 4 year, the TPIPDCGs-based risk-adjusted models can achieve the predictability of 13.0%, 21.1%, 23.5%, and 25.1%, respectively. Conclusion Risk-adjusted capitation models based on principal inpatient diagnoses can significantly improve the predictability on individual’s future expenditure, but its predictability is far from estimated maximal predictability or that obtained from model based on previous year’s expenditure. Adding risk adjusters from previous years can marginally improve the predictability, but predicting more years’ expenditure is more recommended for improving predictability. In order to improve model predictability, further studies to incorporate diagnostic information from outpatient encounters are encouraged. Otherwise, some complementary strategies should be taken to minimize the incentive for biased selection if the capitation payment system is to be implemented.