Comparison of alternative risk adjustment measures for predictive modeling: high risk patient case finding using Taiwan's National Health Insurance claims

Table 2 C-Statistics, sensitivity, and positive predictive value by five risk adjustment models and three outcome thresholds

		Model 1: Demo (sex, ages)	Model 2: ACGs & Demo	Model 3: ADGs, Sel. EDCs & Demo.	Model 4: 2002 expenditures & Demo.	Model 5: ADGs, Sel. EDCs, 2002 expenditures & Demo.
Outcome: 2003 top 0.5% user (N = 242)
C-Statistics		0.773	0.849	0.893	0.904	0.913
Top 0.5% predicted group	% identified	0.52%	0.49%	0.48%	0.49%	0.48%
	sensitivity	0.021	0.037	0.343	0.450	0.450
	PPV	0.019	0.037	0.347	0.447	0.456
Top 5% predicted group	% identified	5.03%	4.85%	4.89%	4.94%	4.83%
	sensitivity	0.169	0.360	0.665	0.698	0.707
	PPV	0.017	0.036	0.067	0.069	0.072
Outcome: 2003 top 1% user (N = 492)
C-Statistics		0.797	0.860	0.893	0.900	0.907
Top 1% predicted group	% identified	0.98%	0.98%	1.01%	1.00%	0.99%
	sensitivity	0.051	0.138	0.313	0.396	0.402
	PPV	0.051	0.140	0.310	0.396	0.403
Top 5% predicted group	% identified	4.84%	4.86%	4.88%	4.94%	4.89%
	sensitivity	0.228	0.384	0.598	0.622	0.650
	PPV	0.047	0.079	0.122	0.125	0.133
Outcome: 2003 top 5% user (N = 2,467)
C-Statistics		0.815	0.869	0.884	0.885	0.897
Top 5% predicted group	% identified	4.89%	4.89%	4.97%	4.89%	4.93%
	sensitivity	0.258	0.365	0.417	0.467	0.476
	PPV	0.264	0.373	0.419	0.477	0.482

Demo: demographic information (sex and age groups); PPV: positive predictive value

ISSN: 1472-6963