Of interest, this study was open-label. The goal was to examine treatment benefit of oseltamivir in patients with flu-like illness, primary outcome was reduction in length of illness (days).
The study found several interesting effects not seen in prior studies on this topic. This includes:
treatment benefit not limited to starting <24 hrs from symptom onset, but also seen w/ start even 48-72 hrs after onset of symptoms.
Patients with more co-morbidities & older age had greater absolute benefit (up to 2-3 day reduction).
Patient with and without Influenza had equal benefit from treatment
The really shocking finding was #3. How did patients that tested negative for influenza have equal benefit form a therapy specific for influenza?
Is their some novel unknown mechanism of action?
Is our serology testing for influenza insufficient (i.e. poor sens/spec), resulting in misclassification of influenza-status?
As study is open label, is treatment benefit all purely from placebo effect?
A lot of interesting questions. The one of perhaps statistical interest to me is evaluating the placebo effect. In an open label trial, is their any way to assess how much of treatment benefit is due to placebo vs a “real” effect? Can we look at magnitude of benefit, relative vs absolute differences, etc to gain any insights?
It is a good question. In a double-blind trial, the issue of placebo group seems solved, but some concerns have also risen. One could argue to solve the problem by using counterfactual ’ what if we had given a placebo to the treatment group’? That would be a must-read paper. About this paper, your post, I also have a question: why should the overall hazard ratio matter more than the adjusted when comparing groups of age so different as one year and older?
I’m not sure I understand the question. The hazard ratio seemed pretty consistent for all subgroups. For clinical translation, showing absolute differences for various ages was helpful.
it was a complete case analysis and not pre-specified? could explain it
“An exploratory analysis not specified in our original statistical analysis plan evaluated the interaction between the intervention and PCR-confirmed influenza status with respect to the primary outcome. These analyses were based on complete case analyses, in which patients with unknown influenza status were ignored.”
Thanks for flagging this interesting study. I’ve read the paper a few times now. A couple of observations/questions:
I’m not sure that most physicians/prescribers will know how to interpret “Bayesian credible intervals” (though arguably most of us probably don’t correctly interpret confidence intervals either). Is this the way that Bayesian trial results are usually presented? Do some trials present results differently (e.g, "the probability that osteltamivir reduces the duration of fever by more than X days is >Y%- this type of wording would probably be easier for the average clinician to understand)?
Still not quite sure I understand the authors’ rationale for the open-label design. Since the primary endpoint comprises symptoms of subjectively-gauged (other than fever) severity, lack of blinding could certainly influence the trial’s result. From my reading in this forum, I suspect there might also be some concerns about categorizing symptoms as mild/moderate/severe, but I’m not qualified to say…
Agree that the apparent lack of specificity of the treatment effect for subjects with PCR-confirmed influenza is perplexing. The only two explanations that seem plausible to me are that the apparent treatment-related benefit reflects a placebo effect OR that oseltamivir has efficacy in treatment of other (non-influenza) viruses as well. This is where blinding might have helped with interpretation of the result (though any attempts at blinding might be have been compromised by oseltamivir’s GI side effects). In the absence of blinding, knowing whether time to resolution of fever (a more objective finding) was faster in oseltamivir-treated subjects than “usual care” subjects, regardless of influenza PCR result, might help to sort this out.
Thanks for these excellent comments & thoughts. For some reason I missed them last month!
Bayesian results are not common (yet), but presentation of results is consistent with other biomedical Bayesian studies. The ability for an average clinician to correctly interpret a frequentest result, much less a Bayesian results, is an open topic.
Agree that rationale for open label study seems unclear. Especially since effectiveness is still debatable given prior literature.
This is a great point. Analyzing something like time to resolution of fever would have been very helpful.