We welcome the provision of additional data in the field of atrial fibrillation (AF) and heart failure (HF), a combination of conditions that results in difficult management decisions and worse outcomes, both for patients with reduced and preserved left ventricular ejection fraction. However, it remains important to adhere to some fundamental aspects of evidence-based medicine. Recent years have seen an abundance of subgroup analyses attempting to answer questions related to treatment effects in nonrandomized studies. These analyses, often using propensity matching or other statistical adjustments, have become commonplace and are often mistakenly considered to be as important as randomized controlled trials (RCTs). Regardless of analysis method, observational data should only be used to generate hypotheses about treatment effects (1,2), and we should resist the temptation to analyze datasets simply because of availability. Ignoring the weaknesses of such studies, and incorrect interpretation, can subsequently lead to confounded conclusions. In the case of the paper by Cadrin-Tourigny et al. (3), the patients were not randomized to receive beta-blockers, and hence there was confounding at both the patient and physician level that hampers external validity. Physicians typically give beta-blockers to patients at lower risk, confirmed in this study as being younger in age, with more nonischemic cardiomyopathy, less time in AF, and higher use of anticoagulation and defibrillator therapy (all factors associated with lower mortality). Although a propensity-matched analysis can be useful to mitigate minor differences in demographic characteristics, it was not designed to account for different patient populations or for exposures that interact with the outcome. The limitations of propensity-matched analysis have been demonstrated for digoxin therapy (the inverse of beta-blockers, in which clinicians tend to prescribe to higher risk patients), where propensity-matching was unable to replicate the results of RCTs (4). Confounding may also explain why Cadrin-Tourigny et al found such discrepant findings for death and hospitalization. Furthermore, only 57% of their population was actually in AF at the time of analysis. We have already shown how effective beta-blockers are in preventing AF (and therefore subsequent adverse outcomes) for patients with AF in sinus rhythm (5). A few methodological issues also arise on detailed review, including the divergence in matched groups, omission of paroxysmal versus persistent AF from the standardized difference plot (with clearly >10% difference), and misrepresentation in the abstract about the sample size (n = 655 [not 1,376], with just 95 deaths without beta-blockers and 136 deaths on beta-blockers). We considered the requirement for AF on the baseline electrocardiography a strength of our previous analysis, which demonstrated a significant interaction in beta-blocker efficacy according to heart rhythm using data from double-blind, placebo-controlled RCTs in patients with HF and predominantly reduced left ventricular ejection fraction (5). Using systematic, carefully checked, and harmonized individual patient data from 10 trials, we identified no significant benefit from beta-blockers in >3,000 patients with concomitant AF, consistent across all outcomes studied and based on a published design paper and pre-specified analysis plan. Where, as a body of clinical scientists, do we go from here? The answer lies in new RCTs, rather than more analyses of observational data.