We discussed this ++ with trial statisticians for iGAS trials & they were not keen on this approach. Not least because your post-hoc look isnt in fact population to which will in future be applied because post trial can still only identify 'true' cases post-hoc so have to continue to give it to all.

Sure this is fine for fairly broad questions like what should i do with this septic looking person in front of me (like Acorn) but still very difficult for more niche questions like "should I give IVIG when it might be iGAS".

Challenge with empiric trials for some syndromes is reasonable proportion don't have infection you are treating which
1) Costs £££
2) Reduces power
And almost certainly needs to be done under waiver of consent

Ongoing discussion for Toxic-shock trials but difficult to actually move forward.

Project on developing ethnicity-stratified transmission models with me and Neil Ferguson. Please get in touch with any questions!
Limited to home fees eligible students.

Think would need to be very large (many agents, would need tight control for error at trial level, would need tight NI margins).

I guess my a priori belief isn't strong enough to want to invest limited trial budget in such study. If you could embed in EHRs & do very cheap might be more plausible.

MAMS-ROCI in Bayesian trials is being discussed in the context of SNAP (with @matteoq21.bsky.social).

For the de-escalation Q I think sample size to detect a meaningful difference in a mortality outcome between individual drugs would be very big & isn't where I would invest my trial budget.

These designs are (normally) informed by simulations to understand trial operating characteristics & each piece you add makes those more complex.

So have to decide which trade-offs to take. For example 7 days less Abx is a big reduction. Not sure knowing it could really have been 8 matters as much?

Adding a PRACTical design again would increase sample size and you would need a strong a-priori reason I think to consider that there are important enough differences between say amoxicillin and cefelexin.

I think there are other considerations. First complexity of randomisation & trial in general. Secondly to date I think MAMS-ROCI has only been evaluated in a frequentist framework & methods would need adapting for a bayesian trial. Although overall efficient it would also increase sample size.

Agree to disagree. In statistical terms I don't disagree of course. But in reality 1) how often these differences are meaningful is unclear to me 2) think much larger unsalvageable errors are in study design & therefore I'm less convinced of the need to be so strict from a pure "stats perspective"

Maybe. My general feeling is that outside of a few very clear issues it would be really hard to actually develop a protocol for this (although AI could possibly do this) AND I would want to see stronger evidence it mattered enough at a real-world level.

Yes although it shifts the burden to the also unpaid editor.

Sure. Toms point was that an editor should desk reject anyone who has dichotomised a variable. My point is that you might have variables that are effectively dichotomised like age into 'in the age group affected by policy or not' and so its not realistic to have a blanket policy of rejecting.

I'd like to see (?simulations) better data on how often some of these practices lead to importantly wrong public health conclusion i.e age as continuous with a spline is more 'accurate' than groups but given how most policies are rolled out, under what circumstances does approach actually matter.

Its a great idea but I think its unrealistic.

I would also be concerned it would have negative equity impacts for authors from global south who have often less access to the level of statistical support required to meet the types of requirements being asked for.

Sure. But my point is that saying editors can just desk reject things if there someone dichotomises a continuous variable is not actually easy to implement. And I can definitely think of scenarios where there is no biological effect on an outcome only a policy level effect.

One issue is some of the rules are hard to apply in reality.
Age is continuous but say I am studying something related to pregnancy where the key external contextual factor is that people <18 are treated different than those of 18+. I think it would be reasonable to dichotomise age.

That said - I suspect we agree really.
I do agree that often Nil > Crap (my view from time to time of the GBD studies) - and I agree the trial design you describe is very difficult and often done badly - but I guess I think can be done well and is then helpful!

Does that mean all trials meet these criteria - No.
Does it mean these trials are harder than trials with a blinded intervention and an objective outcome like mortality - Yes.

But I disagree with your original suggestion that we can never trust such trials & therefore simply shouldn't do them.

I can not conceptualise a plausible natural experiment for Q 'Does CBT improve lived experience of depression'. I dont believe observational data can help nor 'expert opinion'

So my view is we do need well done trials & such trials can be done bearing in mind the issues highlighted.
3/n

You suggest we find alternative approaches or other solutions.
But
i) the meaningful outcome measures for many mental health conditions is by their nature the self-reported experience of the person
ii) most of these interventions cant be blinded
2/n

My view is you can do trials where a) unblinded b) plausible people know hypothesis & c) outcome subjective AND that those trials still generate good quality evidence.

i.e I think these trials (when done well) can generate non-misleading evidence & that is strongly preferable to no evidence.
1/n

Strong disagree. I think that these trials can (and have) been done robustly with results that are not misleading.

The approach you seem to advocate would mean we would never develop any good quality evidence at all on if behavioural interventions work for many mental health conditions.

Does this really mean you think it would be better if we had done no trials of cognitive behavioural (or equivalent) therapy?

In some areas, especially mental health, THE outcome is self-reported symptomatology and I struggle to see how we would ever build evidence if we followed this suggestion.

Do you have a background in Health Economics and want to come and do a PhD on Syphilis and STI

We have a funded studentship @lshtm.bsky.social. Come and collaborate with a great group of academics, clinicians and community groups

Open to people eligible for UK PhD Fee Status - Apply by 11th May.

I think this differs hugely by discipline. For example in trials & much of clinical epidemiology, Introduction > Methods > Results is massively preferable as it provides a clear statement of the planned study & analysis which by its nature should have 95%+ been planned in advance.

Great symposium @lshtm.bsky.social on 22nd April.
Focused on analysis of social inequalities in transmission dynamics for infectious diseases.

Can join online or in person

(if coming in person please reach out to @rozeggo.bsky.social so we can make sure we have in-person capacity)

Clinical PhD opportunity @lshtm.bsky.social through @hpruezi.bsky.social
Applications now close on 17th April.

Ideal for UK Infection registrars wanting to do a PhD in clinical application of metagenomics in returning travellers.

Collaboration with @emcat1.bsky.social & @ukhsa.bsky.social

NIHR Clinical Lectureship in Primary Care @lshtm.bsky.social

Post is open to fully qualified GPs with a PhD and an academic track record.

4 Years of funding to develop your portfolio at one of the UKs leading Public Health and Primary Care research institutions.

Apply by 13th April 2026.

Do you have a background in Health Economics and want to come and do a PhD on Syphilis and STI

We have a funded studentship @lshtm.bsky.social. Come and collaborate with a great group of academics, clinicians and community groups

Open to people eligible for UK PhD Fee Status - Apply by 11th May.