Twitter | Search | |
Kevin McConway
Emeritus professor of statistics ; Milton Keynes inhabitant; choral singer; perpetually puzzled person, interested in anything interesting.
796
Tweets
939
Following
830
Followers
Tweets
Kevin McConway 9h
Replying to @deb_cohen
There's an ethical question on whether one is obligated to point out weaknesses in the evidence if one honestly believes that doing so might cause people to take actions that might cost lives. My feeling is that as a scientist, one is so obligated. But does that apply to doctors?
Reply Retweet Like
Kevin McConway 9h
Replying to @deb_cohen
...someone often appears to do (i) (present evidence) while actually doing (ii) (persuade the public to take a course of action). Or maybe even doing both (i) and (ii) with the boundary between them not clear. I think there was a lot of this on show on Newsnight.
Reply Retweet Like
Kevin McConway 9h
Replying to @deb_cohen
Trouble is that a view of how to do (ii) is to present evidence without any mention of uncertainties ("This is what to do, and this is why, and I'm not going to mention anything about why not to do it because that might stop you doing it"). And the consequent trouble is that...
Reply Retweet Like
Kevin McConway 9h
Replying to @deb_cohen
I think an issue here, very evident in the contributions to your Newsnight piece, is that there are 2 separate but unrelated things going on. (i) What does the evidence say (with all its uncertainties)? (ii) How does one best persuade people to take a course of action?
Reply Retweet Like
Kevin McConway Jun 25
There are several other problems as well, though less serious in my view. This part of the paper should never have got through peer review, in my view.
Reply Retweet Like
Kevin McConway Jun 25
...risk 20% *higher* at distances >3m than it is at distances <3m. That can't be right (and it doesn't really mean that anyway, because it's a phoney analysis because of the varying denominators).
Reply Retweet Like
Kevin McConway Jun 25
There's no data above 2m in their analysis. 's point about assuming a constant ratio is a good one, though in fact it looks to fit their data reasonably for distances up to 2m. But extrapolating to 3m would mean that the risk ratio at 3m is about 1.2, that is, ...
Reply Retweet Like
Kevin McConway Jun 25
...but (I think) it's spurious anyway because of the different denominators. However, the authors just carried on and said it meant bigger effects at greater distances, and used that in various summaries in the paper. They also imply that it might be OK to extrapolate to 3m.
Reply Retweet Like
Kevin McConway Jun 25
But that gets worse. The risk ratios they found were (risk at distances > x metres)/(risk at distances < x metres), so a smaller ratio (nearer 0) is a bigger effect. They found that these ratios increased with x. That means, smaller effects at greater distances. Makes no sense...
Reply Retweet Like
Kevin McConway Jun 25
Second, they meta-analysed studies that compared risks linked to specific distances, and meta-regressed the risk ratios on distance. But the risk ratios have different denominators, so this won't give you ratios between risks at different distances. See .
Reply Retweet Like
Kevin McConway Jun 25
First, there are several problems with the data they got from the studies they meta-analysed, see blog from at . And several studies have potential recall bias issues.
Reply Retweet Like
Kevin McConway Jun 25
Indeed not that paper, for Timandra's issue. But (imho) that paper is very poor on distancing, and I hope it didn't influence decisions. It also looked at masks and eye protections; I've only looked at those aspects briefly and they didn't seem so problematic. So what's wrong?
Reply Retweet Like
Kevin McConway retweeted
Chris Holmes Jun 17
decision theory, causal inference and Bayesian statistics --what more could you want in a paper! -- latest work by Phil Dawid
Reply Retweet Like
Kevin McConway Jun 16
Replying to @Acertaintom
I fear it goes back before there were VPs for academic affairs. The first of those wasn't till 2011.
Reply Retweet Like
Kevin McConway Jun 16
At one time the used not to accredit statistics degrees unless they contained some coverage of survey sampling. That requirement is long gone, but knowing a bit about sampling is indeed still really important IMHO.
Reply Retweet Like
Kevin McConway retweeted
Tom King Jun 14
It is time that the big medical magazines remembered their responsibilities, took an extra week (at least) to check consistency, and insisted on PPI where structural inequalities are a factor. The statisticians are getting suspicious:
Reply Retweet Like
Kevin McConway retweeted
Tom King Jun 14
Interesting recommendations about contact tracing, based on experience of other diseases. But the solidarity aspect of risks and empathy not recognised, and more generally effectiveness of tracing needs more specific evaluation plans
Reply Retweet Like
Kevin McConway Jun 12
Replying to @d_spiegel
I've since had yet another go at trying to explain the problem clearly - see . Of course that might still not help - it is a bit complicated.
Reply Retweet Like
Kevin McConway Jun 12
Replying to @kjm2
...in calculating the risk ratios - and you can't even guess what the 'right' figure is to divide by in 2m studies, because they give no data on how close people actually got to the source of infection if they went closer than 2m (and how close will differ between studies).
Reply Retweet Like
Kevin McConway Jun 12
Replying to @kjm2
...the upshot of using a different comparator in the 1m and the 2m studies is that you can't use the results to compare risks with a 1m minimum distance with risks at a 2m minimum distance, because the basis for the comparison has been taken away by using a different divisor...
Reply Retweet Like