Background

Digoxin has been used for over a century as a treatment for heart failure (HF). The DIG-trial (DIG), the only randomized trial evaluating the impact of digoxin on mortality, showed a neutral effect. However, continuing concerns have been raised about higher mortality associated with digoxin use based on observational data. The present analysis aims to illustrate the fundamental methodological error underlying causal conclusions from observational analyses by direct comparison of observational versus randomized DIG data sets, and demonstrates how the observational approach is misleading.

Methods and Results

This post-hoc analysis of the DIG main trial includes all 6800 patients (ejection fraction <45%) randomized to digoxin or placebo. 44% of these patients received digoxin before the study start. Baseline variables were well balanced among the two randomized treatment groups, but patients previously treated with digoxin differed from patients not previously treated regarding most available baseline variables including heart failure symptoms (e.g. rales, dyspnea, pulmonary congestion), NYHA classification, ejection fraction, duration of heart failure, and heart failure medication (diuretics, hydralazine), demonstrating a more advanced stage of heart failure in patients previously treated with digoxin. While randomized digoxin displayed a neutral effect on mortality in the trial (34.8 vs. 35.1% digoxin vs. placebo; HR 0.990; 95%.CI: (0.913;1.073); p=0.8013), we show that mortality was higher in patients previously treated with digoxin before randomization (40.0 vs. 30.9%; HR: 1.359; 95%-CI: (1.254; 1.473); p<.0001). Most of this increased mortality persisted despite extensive adjustments for baseline differences (HR: 1.214; 95%-CI: (1.110; 1.328); p<.0001). This excess mortality was consistently observed among patients randomized to placebo, demonstrating that the apparent excess in mortality among those prescribed digoxin in non-randomized studies is most likely due to inherent differences in the prognostic risk of patients which statistical adjustment fails to control.

Conclusion