Background: Prophylactic steroids are often used to reduce the systemic inflammatory response to cardiopulmonary bypass in infants undergoing heart surgery. The STRESS trial found that the likelihood of a worse outcome did not differ between infants randomized to methylprednisolone vs placebo in a risk-adjusted primary analysis (adjusted odds ratio [OR], 0.86; 95% CI, 0.71 to 1.05; P=0.14). However, secondary unadjusted analyses showed possible benefits with methylprednisolone. We re-analyzed the STRESS trial using Bayesian analytics to assess probability of benefit with methylprednisolone. Methods: We used a covariate-adjusted proportional odds model using the original STRESS trial model covariates and primary outcome (a ranked composite of death, transplant, major complications and post-op length of stay). We assessed effect thresholds from OR 0.6 to 1.25 (OR <1 conveys benefit, OR >1 conveys harm). We assumed a neutral probability of benefit vs harm with weak prior belief (SD of the normal prior distribution = 0.425). In sensitivity analyses, we evaluated pessimistic (5%-30% prior likelihood of benefit), neutral and optimistic (70%-95% prior likelihood of benefit) prior beliefs, and controlled strength of prior belief as weak (SD = 0.425), moderate (SD = 0.215) and strong (SD = 0.135). We compared posterior distribution of the OR under these priors with the reference results under the vague prior distribution. Analyses consisted of 10 Markov Chain Monte Carlo simulations each consisting of 2000 iterations with a 1000 iteration burn-in to ensure proper posterior convergence. Results: In primary analysis, the posterior probability of benefit from methylprednisolone was 92% and the probability of harm was 8%. The mean absolute benefit was 12%. In sensitivity analyses, the probability of benefit was ≥ 79% for all informative priors except the most pessimistic (Table/Figure). Conclusion: In Bayesian re-analysis of the STRESS trial, probability of benefit with prophylactic methylprednisolone is high and harm is unlikely. Assessing probability of benefit or harm may be more informative than frequentist analytics relying on a p-value threshold. Another advantage is the ability to consider a range of prior evidence.
Read full abstract