Studies of cardiovascular disease risk estimation: how, and whether, to account for the effect of drug treatment?

Abstract

Cohort studies of individuals at risk of cardiovascular disease (CVD) aim to assess subjects’ risk to assist clinical decisions concerning whether patients should be treated with preventive drugs. Since treatment with these drugs is now widespread, it is uncertain how to best account for treatment when modelling CVD risk. Statins and blood pressure lowering agents are considered the principal preventive agents, since meta-analyses generally show between a 0 to 30% reduction in CVD events in the treated group, compared to placebo. The PREDICT cohort of subjects undergoing risk assessment in New Zealand primary care was studied. The thesis consists of five analyses of this population presented in separate chapters, with another addressing statin trials. The first explores the strength of association between initiating drug use and incident CVD. In the PREDICT cohort, use of preventive drugs was common, with 21% of the total taking statins and 32% taking at least one anti-hypertensive drug at assessment. In a cohort of 56053 untreated subjects, after adjustment for commonly measured risk factors, using a Cox model, those who started both drugs were 50% more likely to have a CVD event than those who remained untreated (95% confidence interval (CI): 3% to 117% increase). Since treated and untreated people have different risk factor profiles, the second chapter uses propensity score methods to reassess the association between statin use and CVD. The findings were, however, generally concordant with those of the first chapter. Due to uncertainty about the causal relationship between CVD risk factors, including drug use, the third chapter describes learning Bayesian networks to explore the causal relationships between these factors. The results showed likely causal influence between age and diabetes and baseline drug use; but no relationship between drug use, cholesterol ratio, systolic blood pressure and CVD. In the fourth chapter, the addition of drug use as a covariate in a Cox model did not improve the classification of the model, using varying cutpoints of risk to assign treatment, over a model which included standard CVD risk factors. The fifth chapter examines the presence of publication bias in meta-analyses of statin effects, since the preliminary chapters showed drug use was not strongly associated with CVD. In all three highly cited meta-analyses, the number of reported positive trials exceeded the expected, suggesting bias. The final chapter addresses the magnitude of the association between a novel risk factor, serum urate, and CVD. In this analysis, serum urate was convincingly associated with CVD events. A two standard deviation difference (0.45 vs 0.27 mmol/L) was associated with an adjusted hazard ratio of 1.56 (95% CI: 1.32 to 1.84), using Cox regression analysis. Any possible beneficial effects of blood pressure or lipid lowering drugs are likely to be more than compensated for by unmeasured adverse prognostic factors. This suggests that omitting drug use information is unlikely to bias models used for disease prediction. Some of the discrepancy between observational and trial evidence of drug effects may be attributed to publication bias.