Table 3.

Discrimination, Model Fit, and Calibration Estimates for Total Coronary Heart Disease Events (n = 235 Events) in Older Adults During a Median Follow-up Period of 8.2 Years (Maximum, 10.2 Years), According to Traditional Cardiovascular Disease Risk Factors, With or Without the Addition of Markers of Inflammation or of Atherosclerosis (n = 1,515), Health, Aging, and Body Composition Study, 1997–2007^a

Model	Discrimination		Model Fit				Calibration: Modified Hosmer-Lemeshow Testb
	C Indexc	Adjusted C Indexd	Likelihood Ratio Teste		Measures Penalized for Model Complexity
			χ2 (df)	P Value	Akaike's Information Criterionf	Bayes Information Criteriong	χ2	P Value
Traditional cardiovascular disease risk factors	0.631	0.611	35.31 (8)	<0.001	3,102.6	3,145.2	4.23	0.90
Plus IL-6	0.650	0.629	58.71 (11)	<0.001	3,085.3	3,143.8	8.19	0.52
Plus C-reactive protein	0.638	0.622	41.87 (11)	<0.001	3,104.8	3,163.3	13.86	0.13
Plus tumor necrosis factor-α	0.647	0.621	43.49 (11)	<0.001	3,089.5	3,148.0	14.26	0.11
Plus AAI	0.650	0.624	45.80 (12)	<0.001	3,100.2	3,164.1	15.01	0.09
Plus aortic pulse wave velocity	0.629	0.605	41.42 (11)	<0.001	3,102.6	3,161.1	3.76	0.93
Plus IL-6 + AAI	0.662	0.634	68.58 (15)	<0.001	3,083.5	3,163.3	13.01	0.16

Abbreviations: AAI, ankle-arm index; IL-6, interleukin-6.

^aAll estimates reported are from multivariate analyses, adjusted for traditional cardiovascular disease risk factors, with or without addition of each marker in quartiles or clinically defined categories (for AAI) (see Methods section in text), among 1,515 participants with complete data on all markers (mainly because of 422 missing values for aortic pulse wave velocity).

^bParzen and Lipsitz's adaptation (25), based on martingale residuals, of the Hosmer-Lemeshow test of goodness of fit in the logistic model to the Cox model, comparing observed and expected failures within deciles of predicted risk. Larger P values indicate better calibration (24).

^cHarrell’s C Index, an adaptation of the C statistic (equivalently, the area under the receiver operating characteristic curve) in logistic models. Higher values indicate better discrimination.

^dHarrell's C index, corrected for optimism using bootstrap resampling (24).

^eOmnibus test of the overall statistical significance of all predictors in the model. A higher value indicates better fit.

^fAkaike's Information Criterion (10) is a likelihood-based measure in which a simple measure of goodness of fit (−2 times the log-likelihood) is penalized for the number of predictors in the model. A lower value indicates better prediction.

^gThe Bayes Information Criterion (24) is similar to Akaike's Information Criterion. This measure of predictiveness imposes a more severe penalty for the number of predictors. A lower value indicates better prediction.