“Patients given fluvoxamine within a few days after testing positive for Covid-19 were 31 percent less likely to end up hospitalized and similarly less likely to end up on a ventilator. (Death from Covid-19 is rare enough that the study has wide error bars when it comes to how much fluvoxamine reduces death, meaning it’s much harder to draw conclusions.) It’s a much larger effect than any that has been found for an outpatient Covid-19 treatment so far.”
Good to see paper published. The results were released about 2 weeks ago but first time I’ve seen it in print. There could be something here. But many unanswered questions still.
The composite endpoint includes a real softie…being in an ED for more than 6 hours. That said, the actual difference was driven by hospitalization, which is much more relevant. At the same time, the mean difference in hospital stay was only 1.22 days…and this was not statistically significant (which by itself is not surprising as this was a secondary endpoint and the study was not powered to show a difference on this metric). But still, it would suggest that the difference in hospitalization to the benefit of the fluvoxamine group was driven by the prevention of rather short admission stays. Not nothing, and certainly meaningful if hospitals were to get overwhelmed again. Interesting too that the study population had at least one comorbidity. So it was an at risk group but one that nonetheless did not get too sick with covid, and there was probably a small benefit in hospitalization person-days saved.
It’s encouraging, and at least a good start. More importantly, this shows how real science is done, unlike much of the garbage we’ve seen of late.
It should be noted that this group used this methodology previously to test ivermectin and hydroxychloroquine, with null results.
i only glanced at it but bayes with informative priors for the primary and modified itt (i hate that term), it’s interesting but you’d want to see the sensitivity analyses and i ctrl f ‘sensitivity’ and i dont see it, and it’s not clear - primary sounds like ordinal composite, but then they specify rates etc. informative priors are not really used in practice, ive seen it once and priors were down weighted, might be considered in paediatrics, orphan drug, something like that. but good for them, it’s interesting
edit: ctrl f doesnt work on that in browser
The particulars of Bayesian priors, posterior probability, and the difference btw confidence interval and credible interval, is beyond my level. You sound far better equipped to evaluate their statistical methods.
id just say theyve been pushing bayes for decades and it’s still not widely accepted in drug development, but now people are talking about ‘real world data’ etc, maybe now’s the time. if they used a noninformative prior you’d get very close to the results from the typical analysis. not sure theyre analysing the endpoint properly tho, should be using continuous probability modelling, would have to see the analysis plan. i noticed cytel in the affiliations, large CRO, i guess someone paid $$$
… still glancing, they didnt adjust alpha for the interim analysis, this is really interesting, ive never seen it, it’s an unabashed bayesian acting like a bayesian, without the usual frequentists concessions, bravo
some more info on this trial:
cytel takes credit for the design (as noted above): The TOGETHER Trial: Cytel Designs and Implements Novel Adaptive Platform Trial for COVID-19 Therapies
results presented here: https://dcricollab.dcri.duke.edu/sites/NIHKR/KR/GR-Slides-08-06-21.pdf eg slide titled “Posterior Probability of Superiority of Fluvoxamine vs. Placebo on Emergency Room Observation for > 6 Hours or Hospitalization”
Thanks for that. I am a newb with bayesian stuff. Beginner level of prior probability times likelihood ration equals posterior probability.
Fluvoxamine had a 99.6% posterior probability of superiority versus placebo, in that study population, with those endpoints. Is it fair to then characterize as “99.6% chance that fluvoxamine is better, given what was observed”…rather than to say “there is a 0.4% chance that the finding of fluvoxamine being better was from play of chance”?
i think it just represents the overlap of the distributions. the 1st phrasing is the right one
Another treatment that looks promising. However, this one won’t come cheap.
that paper had no statistician and no statistical review, i just note it to show it’s not unusual even in the major journals. Eg the authors dont know what basic terms ‘multivariate’ or ‘univariate’ mean, and what they describe as ‘univariate’ (which they presented) a statistician would never do. Other stuff like not fitting pre specified covariates regardless, reliance on significance testing, using odds ratio as the treatment effect measure etc suggest they used point-and-click software, so what are the chances the data were handled appropriately? a statistical review would have corrected this basic stuff