Read on Twitter
New paper just published in @IJEeditorial. It’s about when to analyse cohort study data, an issue I’ve had in both my PhD and now Post-doc. Paper is quite dense at times ;) so here is a short thread. A blog post will follow on @IJEeditorial shortly too. https://doi.org/10.1093/ije/dyz212
We’ve used @uk_biobank data to look at the association between phys act and mortality/CVD. There were 7 years of follow up time available. We reanalysed the data multiple times, artificially shortening the follow up time. Shorter follow up = stronger associations with outcomes.
We thought this might be due to reverse causality i.e. ill people reporting low activity introducing bias. So also used different strategies to account for this: adjusting for disease status & various exclusions. The more you adjust/exclude = weaker associations. (Supp Fig below)
(At this stage it is worth saying that longer follow up also gives people more chance to change their activity levels from what they reported, which would also weaken the relationships.)
But the strategies used to deal with reverse causality made more difference at the shorter follow up times.
(realise you might need a key for many of these figs, but go see the paper https://doi.org/10.1093/ije/dyz212)
(realise you might need a key for many of these figs, but go see the paper https://doi.org/10.1093/ije/dyz212)
We hope this helps inform people making decisions about when to analyse data. We don’t aim to give a number, but to help understand what the consequences might be.
We've also got a more simple summary blog coming out on @IJEeditorial soon so will post that link when it comes.
We've also got a more simple summary blog coming out on @IJEeditorial soon so will post that link when it comes.
Just to add the link to the blog post https://ije-blog.com/2019/10/29/
