Improving and assessing coronary heart disease risk prediction at the individual and population levels

This dissertation investigates methods for assessing and improving coronary heart disease (CHD) risk prediction at the individual and population level. CHD is the leading cause of death in the United States and most of the developed world and CHD incidence is increasing dramatically in many parts of the developing world. Models of CHD risk at the individual level are used in research settings to answer etiologic questions and in clinical settings to guide treatment decisions. Population trends in CHD risk inform policy decisions and interventions. Three approaches to improving and assessing CHD risk prediction methodology, two at the individual level and one at the population level, are presented. These approaches draw upon two sources of data from the Atherosclerosis Risk in Communities (ARIC) Study, (1) a sampled cohort; and (2) surveillance of hospitalized myocardial infarctions and CHD deaths in four US communities.; The first paper investigates the effect on CHD prediction of correcting for long-term variation in selected major CHD risk factors (systolic blood pressure, total cholesterol and high-density lipoprotein (HDL) cholesterol) in the ARIC cohort. Correction was done using regression calibration, which included the relationship between all of the risk factors to obtain estimates of individual's long-term average value for the selected risk factors. Correction had a substantial impact on hazard ratio estimates, strengthening the estimates for the varying risk factors while weakening the effect for others (especially age and medication use), but using data from two visits three years apart resulted in only a small improvement in overall risk prediction.; The second paper examines the effect of misclassification of any outcome on the area under the receiver-operator characteristic curve (AUC), a commonly used measure of the predictive ability of a model. The change in the AUC is also used in CHD risk prediction to measure the predictive value of novel risk factors. A taxonomy of misclassification is presented and a formula for the relationship between the true and observed AUC is derived. An example of the effects of misclassification on the observed AUC and the changes in the observed AUC with the inclusion of different predictors is also presented. Misclassification of the outcome was found to have a profound and quantifiable effect on prediction which should be incorporated into the interpretation of AUC analyses.; The third paper investigates the extent to which individuals' risk factor levels and trends explain community CHD rates. ARIC cohort members' information from each visit and follow-up were combined to generate models for each individual's predicted probability of a CHD event. These probabilities were then summarized across the individual risk factor distributions for each geographic, chronologic and demographic group and compared to the observed rate obtained from the surveillance and census data. The relationship between observed and cohort-predicted CHD risk varied across communities but risk factor trends translated to an expected decline in CHD.; As a whole, this dissertation addresses key issues in measurement, modeling and interpretation: characterization of the predictors, characterization of the outcome, and generalizability of the results. These results have direct applications to CHD research, as well as implications for policy and clinical practice.