tah-sci.com
Professor of Mathematical Sciences, working mainly on epidemiology although partial to a bit of non-commutative algebra, social science and basic biology.
https://personalpages.manchester.ac.uk/staff/thomas.house/about.html
1,994 posts
4,327 followers
945 following
Regular Contributor
Active Commenter
comment in response to
post
This kind of thing isn't even the most extreme behaviour the DUP's engaged in over the years.
comment in response to
post
It was dreadful, there were all sorts of confusing things like confidence intervals.
comment in response to
post
I reckon the biggest issue is lack of belief that the NHS, Universities and Civil Service can deliver "in house" without involvement of self-declared genius tech bros.
comment in response to
post
During the pandemic, we used the COVID relaxation to produce free open source tools for hospital demand prediction that run on anonymised data, and are struggling to find support to roll it out for seasonal winter pressures across the whole NHS. It was loved by trusts that had it.
comment in response to
post
"deserves a nickname"
comment in response to
post
Anyway, as you say it's that situation where HARK is the most justified, and kind of why people hold back that residual credibility, that the experiment might throw up something so surprising you have to come up with an explanation after running it.
comment in response to
post
When you elucidate priors, people usually reserve some credibility that something really unexpected happens, it's just hard to construct a sensible probability measure that captures this - various schools of thoughts like Bayes conceptually but find probability measures too constraining.
comment in response to
post
Maybe I gave up too early, this would be a reasonable use case!
comment in response to
post
Outlook tries to do that a bit by highlighting documents sent and it's actually pretty useful, so more of that is exactly where the LLM could help - low-stakes automation.
comment in response to
post
LLMs really integrated into emails could be quite useful because natural language based search would be much easier for people than "advanced search", why isn't that the direction of travel?
comment in response to
post
So what's the point? This stuff is now everywhere with the most potentially useful applications absent and usually customised with the most negative features of the tech company the only distinction (Meta violates privacy, Microsoft is corporate, Grok is fascist...).
comment in response to
post
Yes I think CIs on data counts are a crude way of accounting for uncertainty due to sampling but are basically redundant once you've done a full model of them. We use them in exploratory analysis usually.
comment in response to
post
I think my second comment said I agree here, but the way I would approach it in a discussion with someone is, what uncertainty are you hoping to capture with your interval?
comment in response to
post
I think it's silly if it's a small cohort size (i.e. the "n" for a study) but it's perfectly respectable to bootstrap admissions with something for a hospital to get an idea of how much we might expect that to vary under repeated observation with other things equal.
comment in response to
post
They needn't but they can! Or rather, if you take the empirical count as an estimator, there's no reason that estimator shouldn't have a CI. Easiest methods to produce that are usually Bootstrapping or Poisson approximation.
comment in response to
post
It's hard to know about cause and effect - the Sun newspaper here liked to claim high levels of influence on public opinion, but they did this partly by looking at who was likely to win elections then backing them.
comment in response to
post
I agree, it was a hope I think rather than assumed. My feeling is most people from modellers to clinicians felt surprised at how short the times could be between Mpox infections or vaccine and infection, but not concerned that understanding was fundamentally off.
comment in response to
post
I'm thinking out loud here around well informed people rather than arguing - it's a very odd political context. Trials are very special and should be defended but their absence needn't imply inaction.
comment in response to
post
Equally their behaviour against Mpox has been different from expected - protection against future infection seems lower than hoped rather while protection against severity holds up. Which seems similar to immunity following natural infection.
comment in response to
post
I'm assuming the context here is the US discussion around vaccines - and it's a bit of a painful one because the Salk vaccine and trials were such a pinnacle of human achievement, we mustn't forget that but equally reality is often more complicated.
comment in response to
post
There was rigorous research done to show they were safe and worth distributing against Mpox, but we did have to put together various bits of scientific knowledge to conclude that rather than run a classic vaccine trial.
comment in response to
post
Actually a somewhat serious answer for the UK is that social democracy is led by university graduates who aren't organically connected to their working-class vote, which is much more diverse and socially liberal than they imagine.
comment in response to
post
I have no idea why European politicians decided that importing the US culture war here was the route to power, I guess it seems easier than trying to solve real problems which is overwhelmingly what voters want and reward.
comment in response to
post
Similar - it's not the time per se, it's the "Do I agree with statement 1 strongly, but very closely related statement 2 very strongly? Oh look, here's all the same statements but in a different context on the next page, can I even be consistent?" that gets me.
comment in response to
post
See also, "Why won't people who are completely overloaded with bureaucracy fill in my 100 question survey that requires significant cognitive burden to distinguish between answers?"
comment in response to
post
> scientists do! It seems healthy to propose ambitious hypotheses and theories that can be tested, even to the point of raised eyebrows from colleagues, but not to insist on them somehow.
comment in response to
post
Thanks - I can't emphasise enough how, well, *silly* actual specialists found Friston's epidemic models; we all make mistakes but to be so loudly and confidently wrong, claiming in the national press to have discovered "epidemiological dark matter" as Friston did ... This is not what serious >
comment in response to
post
The paper itself seems like more of a rant / humour article, but it would be very interesting to see what neuroscientists and others generally think - do they feel able to speak freely on this?
comment in response to
post
Hi - as an epidemiologist I can say that when Friston tried to do epidemic modelling, his work was very clearly without merit. But, the idea that humans behave a bit like Bayesian reasoning would suggest is rational seems both sensible and testable. It's interesting to see this comment.
comment in response to
post
Very amateur historical take, but I just don't think immigration was conceptualised in the past in remotely the same way we do in the modern era. And even then the law changed fundamentally every few decades since the late C19 in the UK, say.
comment in response to
post
There's people I find excellent writers otherwise whose games writing aren't so enjoyable - the issue is we all see the same movie, but have very different experiences of the same game - I also prefer travel videos to travel writing.
comment in response to
post
It's unique not not really weird - reading about music, film, and visual art is great but videos are really a very natural format for games commentary, since games are participatory in a way other art forms aren't.
comment in response to
post
Went on a bit of a Shamen watching spree and I really wasn't ready for the multiple oiled up Jason Stathams dancing in his pants here.
www.youtube.com/watch?v=uWu3...
comment in response to
post
> fundamentally different even within countries - people are older, fatter, smoke less, and have very different comorbidities etc. So we have to apply knowledge from outside: e.g. vaccine trials likely not altered by this but maybe other things are.
comment in response to
post
Yes I remember these from before! I think that's pretty consistent with where I've landed philosophically - RCTs are very special but the weakest point is what you're calling "as they present". For the first big modern-style trials in the 1950s we might expect that this is now >
comment in response to
post
As I see it, the main "issue" of a trial design is generalisability - there may be genuine potentially unobserved phenotypic or genotypic subgroups that benefit or don't, which vary across time and space.
comment in response to
post
Can I just check in case I'm missing something - is the reason this is wrong that the uncertainty quantification should catch this issue, i.e. it might cause problems for statistical analysis but isn't a flaw in the design's ability to inform the research question?
comment in response to
post
Of course when things are very non-linear / lots of emergence they're weakest - but even very emergent things like swarming were susceptible to some experimental manipulation (I seem to remember a tub of locusts feeding into a lot of modelling).
comment in response to
post
Personally I'm really leaning towards triangulation / inference to the best explanation as the closest "philosophy" of causal inference, and this means there's few areas where randomised experiments don't help build up the picture - see lots of great work from Watts in complex systems for example.