Read on Twitter
1/
Did you know that a Bayesian approach might help enhance scientific and clinical decision-making in critical care medicine?
If you’re curious, check out this brief thread (and accompanying manuscript: https://authors.elsevier.com/a/1c6U-7tFB1JajF)
Did you know that a Bayesian approach might help enhance scientific and clinical decision-making in critical care medicine?
If you’re curious, check out this brief thread (and accompanying manuscript: https://authors.elsevier.com/a/1c6U-7tFB1JajF)
2/ Bayesian analysis combines prior information (opinions or data) with observed data to arrive at a posterior probability distribution. It’s been developing for centuries, and was used to crack the German ENIGMA machines during WWII https://books.google.ca/books/about/The_Theory_That_Would_Not_Die.html?id=_Kx5xVGuLRIC&printsec=frontcover&source=kp_read_button&redir_esc=y#v=onepage&q&f=false
3/ Bayesian analysis is often contrasted to conventional aka frequentist analysis. We used a recent systematic review of critical care trials by Abrams et al https://pubmed.ncbi.nlm.nih.gov/33031148/ to investigate the extent and nature of discordance between Bayesian and frequentist analyses
4/ We gathered clinician estimates of the minimum clinically important difference (MCID) in terms of absolute risk reduction. For example, for the PROSEVA trial the median minimum important absolute risk reduction (MCID for ARR) was 4%.
5/ We reanalyzed 82 trials: 78 were negative or indeterminate by frequentist criteria (p>=0.05 or ARR shows harm) and 4 showed benefit with p<0.05. If the Bayesian posterior ARR was more likely than not to be better than the MCID (prob > 0.50), we called it “potential benefit”
6/ Bayesian and frequentist analysis disagreed in a minority of cases. This is reassuring – Bayesian analysis is not a method for reframing “negative” trials as “positive.”
7/ But there were three ways that Bayesian analysis added value. First, Bayesian analysis helped to identify interventions where potential benefit has not been ruled out:
8/ Second, Bayesian analysis quantitatively identified scientific controversy, aka trials where the conclusions depend heavily on the prior distributions. For example:
9/ Third, Bayesian analysis measured the probability of clinically important benefits in a continuous, non-dichotomous way. Here you see that the PROSEVA study ARR is almost certain to be better than the MCID of 4%, but unlikely to be greater than 15%:
10/ You can investigate all this and more on our accompanying shiny app: https://cyarnell.shinyapps.io/BRICCS-Interactive-App/
11/ Limitations 1: We implemented Bayesian analysis with absolute risk reduction because clinicians have better intuition for ARR https://pubmed.ncbi.nlm.nih.gov/26504102/ . But a more common implementation uses log-odds ratio or beta-binomial because they are mathematically smoother.
12/ Limitations 2: Prior distributions used were not tailored to each trial. The individual results are not meant to be interpreted as definitive reanalyses! Bayesian analysis requires priors specified to reflect the plausible range of beliefs about the studied intervention.
13/ Limitations 3: Bayesian analysis can help us understand the evidence we’ve obtained, but it can’t tell us the right questions to ask. Just like any other statistical method, we still need clinical expertise to guide the research.
