Results were similar in our secondary analyses comparing full vs highly-selected ALSPAC samples, with no meaningful differences in interpretation despite large differences in sample size and risk of selection bias.

Turns out, the answer is: practically no difference whatsoever!

While there was a small amount of bias – a 2%-point difference in the odds ratio (from 1.00 to 1.02) – the confidence intervals still contained the true value ~95% of the time, so for all practical purposes the bias was non-existent.

For instance, if both our exposure and outcome predict participation (ie, collider bias), this can bias the relationship between these variables.

(Image from doi.org/10.1038/s414...)

We all know that – alongside global warming, Elon Musk, Vladimir Putin and other evils – selection bias is one of the largest threats to humanity*

But how bad exactly is selection bias?

This is what we aimed to explore in our recent paper: doi.org/10.1080/2153...

*Perhaps a slight exaggeration…

Recently took part in Carbon Literacy training & my action to try and reduce carbon emissions was to share recipes encouraging plant-based foods (as lower impact vs meat/dairy). Thought I'd share here in case of interest

Gnocchi w/ tomato sauce - Enjoy! 😋

danmajor-smith.netlify.app/2024/05/20/c...

Gave a talk on 'Ethical Academic Publishing' earlier, so thought I'd share the slides in case of interest to anyone: osf.io/urpx5

Basically, the system is f**ked, so we need to publish more ethically (e.g. support Diamond OA & not-for-profit journals)

This slide is a summary of my current thinking

Say hello to our delicious new foster kitten Tinks!

She's six weeks old and was found under a bush in a park (her mum was a feral, we think) 😥

But she's super lovely and playful and ruddy adorable 😍

In addition to promoting open-science practices, these synthetic datasets can also be used as educational tools to learn about new methods or how complex data works in reality – We provide an example of longitudinal modelling comparing observed vs synthetic ALSPAC data in the paper

/8

Researchers can then use these synthetic datasets to follow the analysis code, replicate the results in the paper (albeit not perfectly, as the data is not 100% identical), explore the methods used, and more generally check the analysis for any issues or errors

/7

We provide an easy-to-follow example using synthpop to generate synthetic ALSPAC data, with an openly-available subset of ALSPAC data (osf.io/8sgze). We also describe ALSPAC guidelines on how to make synthetic ALSPAC data publicly-available (for use in any study)

/5

For my first post on this site, here's a scrumptious photo of our foster cat Bella - Enjoy! 😊