The provocative article by Serstè et al.1 about the negative impact of propranolol on the survival of patients with cirrhosis and refractory ascites ends with the recommendation that beta-blockers should not be used in patients with refractory ascites. Some of the methodological concerns of the study are addressed in the accompanying editorial: the lack of causality between the main cause of death (hepatocellular carcinoma) and beta-blocker therapy, the lack of consecutive patient enrolment, and the misbalance between groups with sicker patients in the beta-blocker group.2 The editorial also points out that a randomized controlled trial would have been the most appropriate tool for evaluating the effects of beta-blockers. Although we agree, we think that carefully analyzed observational data may provide us with firm clues about causality even without balanced randomization.
Serstè et al.1 linked beta-blocker therapy to increased mortality with a proportional hazards model. Through automatic backward modeling, the criterion for selecting the final model strategy was statistical significance. We think that an explanatory strategy for building the Cox model, using time-dependent covariates and a propensity score to adjust for the potential confounding factors, would have enriched the study.3–5 In this way, instead of being driven by significance tests, covariates would have entered and remained in the explanatory model as a result of their modification effect on the association of therapy and mortality.3 Moreover, they could have checked for confounding and likely interactions to explore whether the observed effect was the same in different subsets of patients, as the editorialists claimed. Besides, the use of time-dependent covariates would have allowed fine-tuning of the beta-blocker therapy duration and would have better addressed its influence on outcomes.4 Finally, a propensity score, which defines the probability that an individual will receive a specific treatment based on his or her pretreatment characteristics, is useful for overcoming the imbalance between groups when treatment assignment is not random.5 Specifically, in Serstè et al.'s study, the propensity score would have corrected the effect of beta-blockers for patient characteristics such as the presence of varices, which heavily conditions their prescription. With such an analysis, the focus of the model would have been the influence of beta-blockers on survival rather than the identification of factors influencing survival; hence, it would have offered more clues to the causal effect. The proposed approach would add robustness to the interesting results provided by Serstè et al.