OUP user menu

When ‘digoxin use’ is not the same as ‘digoxin use’: lessons from the AFFIRM trial

Sabina A. Murphy
DOI: http://dx.doi.org/10.1093/eurheartj/eht087 1465-1467 First published online: 16 April 2013

This statistical viewpoint refers to ‘Increased mortality among patients taking digoxin—analysis from the AFFIRM study’, by M.G. Whitbeck et al., on page 1481; and ‘Lack of evidence of increased mortality among patients with atrial fibrillation taking digoxin: findings from post hoc propensity-matched analysis of the AFFIRM trial’, by M. Gheorghiade et al., on page 1489

Two reports relating to the Atrial Fibrillation Follow-up Investigation of Rhythm Management (AFFIRM) Trial1 explore the association of digoxin use with mortality among patients with atrial fibrillation (AF). The AFFIRM trial randomized 4060 patients with AF and a high risk of stroke or death to rate control vs. rhythm control. In the rate control arm, different therapies were allowed to achieve adequate heart rate control, including digoxin, beta-blockers, calcium channel blockers (verapamil and diltiazem), and combinations of these drugs. In the rhythm control arm, antiarrhythmic drugs included amiodarone, disopyramide, flecainide, moricizine, procainamide, propafenone, quinidine, sotalol, and combinations of these drugs. While the strategy of rate control vs. rhythm control was randomized, the choice of specific agents used was left to the discretion of the treating physician in both arms. During a mean follow-up of 3.5 years, there were 356 deaths (23.8%) in the rhythm control group and 310 deaths (21.3%) among the rate control group, a directionally but not significantly lower mortality with rate control (P = 0.08).

In the article by Whitbeck et al., digoxin use was assessed at randomization and during follow-up.2 The association of digoxin use with mortality was evaluated treating digoxin as a time-dependent covariate in a Cox proportional hazard model.3 By using digoxin as a time-dependent covariate, patients changed from being in the ‘on-digoxin’ group to the ‘not on-digoxin’ group if their medication use changed over time in the study, and their associated time at risk for death contributed to each respective group. A simplistic way to think of this is to consider a patient who is on digoxin for the first 9 months of the study but then discontinues digoxin for the remainder of the follow-up until the patient dies at 18 months. The patient is counted in the model as alive and ‘on-digoxin’ for the first 9 months of follow-up, alive and ‘not on-digoxin’ for the second 9 months of follow-up, and then died while ‘not on-digoxin’ at 18 months. Using digoxin as a time-dependent covariate, the Whitbeck analysis found a significant increase in all-cause mortality associated with digoxin [hazard ratio (HR) 1.41, 95% confidence interval (CI) 1.19–1.67, P < 0.001] after controlling for clinical and demographic variables, as well as propensity score. Additional sensitivity analysis found similar observations when restricted to patients randomized to the rate control only cohort (HR 1.46, 95% CI 1.13–1.90, P = 0.004). Based on these data, Whitbeck et al. concluded that digoxin use was associated with a significant increase in all-cause mortality.

In the article by Gheorghiade et al., digoxin use was assessed at a fixed time point only, at the time of randomization.4 Patients were excluded from the analysis if they were on digoxin in the previous 6 months but discontinued digoxin as initial therapy at randomization (n = 465) or had missing information on digoxin use at randomization (n = 887). Using digoxin at randomization, the Gheorghiade analysis found no increase in all-cause mortality associated with digoxin (HR 1.06; 95% CI, 0.83–1.37; P = 0.640) in a propensity-matched analysis (n = 1756). Based on these data, Gheorghiade et al. concluded that there was no association of digoxin use with mortality, in stark contrast to the conclusions of Whitbeck et al.

How is it possible that two reports from the same trial can reach such divergent findings from what is seemingly a similar analysis of digoxin use? Several issues contributed to the different conclusions (Table 1).

As outlined previously, how digoxin use was defined differed greatly between the two studies. In the Gheorghiade manuscript, digoxin use was assessed at a fixed time point of initial therapy at randomization while in the Whitbeck manuscript digoxin was treated as a time-dependent covariate, accounting for changes in digoxin use over time in the study. While the latter may seem the preferred method as it provides more detailed use of digoxin, this type of approach is not always appropriate, such as when the change in treatment is related to worsening of the patients' health. For example, consider a hypothetical patient who was not on digoxin as initial therapy at randomization but who developed heart failure during the trial and initiated digoxin as treatment for the new heart failure. If the patient subsequently dies, using digoxin as a time-dependent covariate would attribute the death to the digoxin treatment. However, it is possible that the mortality may be driven by the development of new heart failure (which is known to increase mortality) and not by the digoxin itself, which was only given in response to the development of heart failure. This can lead to artificial inflation of the treatment effect when it was actually the intermediate event that was the true cause of the death. This is referred to as indication bias.3 Indication bias can be reduced by adjusting for the factors that may prompt the change in treatment. Indeed, in the Whitbeck manuscript, several variables that could be potential intermediate events were included in the Cox model as time-dependent variables, including permanency of AF, elevated heart rate, beta-blocker use, angiotensin-converting enzyme inhibitor use, amiodarone use, cumulative number of shock episodes, New York Heart Association (NYHA) functional class, and Canadian Cardiovascular Society (CCS) angina status.

View this table:
Table 1

Summary of differences between AFFIRM studies in the primary methods used for evaluating the relationship between digoxin use and all-cause mortality

Whitbeck et al.Gheorghiade et al.
Study designNon-randomized, observational analysis using data from randomized AFFIRM trialNon-randomized, observational analysis using data from randomized AFFIRM trial
Time point digoxin used assessedTime-varying covariate, throughout studyFixed, at baseline only
CohortFull cohort (n = 4058)Selected cohort (n = 1756)
Propensity methodAdjustmentMatchinga
Primary HR for digoxin and all-cause mortality associationHR 1.41, 95% CI 1.19–1.67; P < 0.001HR 1.06, 95% CI 0.83–1.37; P = 0.640
Main conclusion from authorsDigoxin associated with significant increase in all-cause mortality in patients with AFNo evidence of increased mortality associated with digoxin use as baseline initial therapy in patients with AF
  • AF, atrial fibrillation; CI, confidence interval; HR, hazard ratio.

  • aPropensity adjustment used for sensitivity analysis.

Another key difference between the manuscripts is the cohort used. Since the Gheorghiade manuscript evaluated digoxin use as initial therapy at randomization, subjects could only be considered for inclusion if data were not missing at this single time point (n = 1266 in the rhythm control group and n = 1957 in the rate control group); all other patients had to be excluded from the analysis (n = 767 in the rhythm control group and n = 70 in the rate control group).1 This difference in available data between treatment groups was due to changes in the design of the case report forms during the trial, with details of initial medication use only collected in the rhythm control group for part of the trial.1 Missing data can introduce selection bias, particularly if the data are not missing at random,5 which is the case here. Indeed, mortality was higher among subjects with missing data on digoxin use at randomization as compared with subjects with digoxin data available.6 The Gheorghiade manuscript also excluded 465 subjects receiving digoxin in the 6 months prior to the trial who then discontinued digoxin use at randomization.

Given the non-randomized selection of treatment with digoxin, both studies employed a propensity analysis in order to control for selection bias. Propensity scores attempt to account for differences in the pre-treatment characteristics between patients treated with experimental and control interventions, resulting in a less biased treatment effect estimate.7,8 Differences exist between the studies in the propensity analysis, both in the time point for predicting digoxin use and in the approach used for the propensity score analysis. Propensity scores can be applied using several different methods, such as matching, stratification, adjustment, or inverse probability of received treatment weighting.7 Whitbeck et al. used propensity adjustment, where the propensity score is added to the treatment effect model as a covariate. Caution should be applied with propensity adjustment, which is often subject to incorrect assumptions about the functional relationship of the propensity scores and outcome, e.g. non-linear relationships and mis-specification of the propensity model, both of which can result in biased treatment effect estimates. Gheorghiade et al. used propensity matching, which selects one patient from the treated group (i.e. the ‘on-digoxin group’) and matches them to one patient from the control group (i.e. the ‘not on-digoxin group’) with a similar propensity score. Propensity matching allows for baseline covariate balance but faces the limitation of a reduced sample size since matches between experimental and control therapy cannot be found for all patients.9 Indeed, 950 of 2706 subjects (35%) were dropped from the Gheorghiade analysis due to not having an adequate match between the ‘on-digoxin group’ and ‘not on-digoxin group’. However, similar results were observed in a sensitivity analysis using propensity adjustment rather than propensity matching, which allowed for inclusion of the larger cohort (n = 2706).

What conclusions can be drawn from these two analyses? Given the non-randomized, observational design of both studies, the findings should be considered hypothesis generating. Even sophisticated statistical methods such as propensity analysis cannot replace randomization. When such observational studies are published, it is crucial for the reader to understand the cohorts and the how treatment groups are defined, because sometimes digoxin use is not the same as digoxin use.

Conflict of interest: none declared.


  • The opinions expressed in this article are not necessarily those of the Editors of the European Heart Journal or of the European Society of Cardiology.