Presenting our ontology proposal for geospatial exposome data at the Int'l Society of Exposure Science workshop in Lisbon. Nice to see that it already attracts great interest from other groups, highlighting the need for metadata standards in the exposome field. Feedback is welcome.

osf.io/sy256

Last week, I attended a course on Target Trial Emulation by @miguelhernan.org at @cemfi.es in Madrid.

www.linkedin.com/feed/update/...

Stillness and foul air
Together, they dim the mind
Data whispers faint

Our new paper investigated the syndemic effects of air pollution and physical inactivity on cognitive decline.

doi.org/10.1123/jpah...

www.thenational.scot/news/2517213...

I also want to assess residual effect (spatial + unstructured) after controlling for deprivation. But, their distribution (see figure) doesn't seem okay.

I have also checked similar studies but I am still unsure if I am calculating the results from the posteriors correctly.

However, when I exponentiate the coefficient posteriors to get an RR, it gives an unreasonably big number. The issue persists with different priors. What am I doing wrong here?

Should I instead exponentiate only the mean of the posteriors? With this, the RR is reasonable but HDIs are still not.

The model specifications are below. I used non/weakly informative priors. Traces has converged and reached a fair precision.

Is there a friendly Bayesian fellow out there who can help me figure out my first Bayesian hierarchical analysis? 😇

I am trying to model the number of COVID-19 cases across 33 neighbourhoods. I want to account for spatial autocorrelation using an ICAR model (Besag-York-Mollie) and deprivation.
...

Today was my last day at Queen's. Grateful to have worked with an inspiring PI, Ruth Hunter, and brilliant teams on several exciting projects. Last 3 years have been truly a period of growth!

Lately, I was asked to draw what my last project means to me 👇

The project: www.qub.ac.uk/sites/ground...

🔦 How can such AI be used?

Our model enables surveillance of tobacco-promoting content both for research purposes and enforcement of tobacco control measures. Beyond that, we suggest a range of health promotion opportunities this tool can help.

📈 Quantitative 2 - We predicted tobacco promotion in social media content for a period of a month and explored its distribution by content characteristics.

🧠 Quantitative 1 - We fined-tuned a pre-trained language model (BERT) to classify tobacco-promoting content. It achieved 87.8% recall and 81.1% precision.

🏷️ Qualitative 2 - We manually coded a small sample of social media posts. Some posts fell easily into one category (Fig1), yet others were more complicated and required us to use a matrix (Fig2).

We used mixed-methods

💬 Qualitative 1 - We inductively identified four main ways of tobacco promotion in social media: modelling the behaviour, expressing positive attitudes, recommending use, and marketing brands or vendors.

Hi Peter, nice to come across you in the (blue)sky :)