Thread by @andersgranholm

#65trial Primary outcome of 90-day mortality not "statistically significant", but point estimate favours perm. hypotension. Can a Bayesian re-analysis help us interpret these results? #CCR20 (1/...)

https://twitter.com/VirtueOfNothing/status/1217739063027019777

First, the best way to interpret frequentist 95% CIs that I have encountered (@EpiEllie) is to consider them to be like ring tosses - either they contain the true value or they don't: medium.com/@EpiEllie/havi… (2/...)

@EpiEllie If the study is repeated indefinitely and a 95% CI is calculated each time, 95% of these intervals will contain the true value - but for a single 95% CI we don't know whether or not this is the case (either it does contain the true value or it does not). (3/...)

@EpiEllie On the contrary, Bayesian credible intervals (CrIs) are actual probability distributions, and the interpretation is more intuitive. A Bayesian 95% CrI contains the 95% most probable values (given the data, the model, and the assumptions used). (4/...)

@EpiEllie This is how frequentist 95% CIs are often interpreted, although somewhat incorrect. While these differences are somewhat technical, Bayesian analyses can be used to provide valid probability statements regarding effects. (5/...)

@EpiEllie For the simplest analyses (e.g. unadjusted RRs) if we use flat priors (= assume that all results are equally likely, as in frequentist analyses), the Bayesian 95% CrI will be similar to frequentist 95% CI. (6/...)

@EpiEllie Based on a Bayesian re-analysis, we can estimate the prob. of benefit (RR<1) or harm (RR>1) and define a region of equivalence (e.g. RRs between 0.95 and 1.05), and calculate the prob. of equivalence or clinically imp. benefit/harm (values inside/outside this region). (7/...)

@EpiEllie If we conduct the simplest possible Bayesian re-analysis and look at the unadjusted RR, we get the following results: #65trial #CCR20 (8/...)

@EpiEllie There is 92% prob. for benefit, 63% prob. of an RR<0.95 and only 1% prob. of an RR>1.05. Similar analyses could be done for the absolute diff or adjusted (if data were available). This is more informative than dichotomising the result as statistically significant or not. (9/9)

@EpiEllie On an additional note, the similarity of Bayesian 95% CrIs and 95% frequentist CIs for simple analyses with flat priors also makes the common interpretation of 95% CIs sensible in these situations from a pragmatic point of view ... (10/11)

@EpiEllie Working with posterior distributions is just easier and more formalised. The similarities between approaches are discussed in
@VPrasadMDMPH s
excellent podcast in the episode here: soundcloud.com/plenarysession… (11/11)