On @bengoldacre et al's new study, looking at 17 million people to establish risk factors for Covid-19; and why the exciting thing is not the findings but because it shows us a way to use one of Britain's greatest scientific resources, NHS data, safely https://unherd.com/thepost/finally-a-way-to-analyse-nhs-data-from-17-million-people/
(in which I unwisely try to explain collider bias, a statistical anomaly which I only vaguely understand myself, in a one-paragraph parenthesis in a 500-word blog post)
(I reckon it's about 70% likely I've made some cockup in that paragraph in fact. Check it! The first person to point out an error will win a prize of no money.)
someone in the DMs points out that collider bias can in fact go both ways – it can inflate or reduce the effect of a variable. Should have made that clear. But here's a good thread on collider bias which I am going to steal shamelessly for next time I write about it https://twitter.com/bristimtom/status/1259458854020165632