Emma Thornton, Psychology, University of Liverpool (2017 Cohort)

After 18 long months of working from home with only my dog for company, at the beginning of September I was lucky to attend The Alan Turing Institute Summer School in Causal Inference with Observational Data at the University of Leeds. In person! Now I know some of you may be wary about attending your first events in person (and having to socialise) but hopefully I can put your mind at ease. By attending this course in person, I had the opportunity to meet new people for the first time, and also meet friendly faces that I had only ever seen over Zoom (and what better way to do this than eating different cuisines every night such as Thai, Japanese and Indian). Although going out for dinner used to be the norm – this felt like such a treat and socialising in person really made this experience great.

This summer school was an intense 5-day introduction to analysing observational data to determine causal relationships. Causal inference is inherently different from prediction, which seeks to estimate the likely value of an outcome, given information about risk factors. The course was deeply philosophical, delving into common pitfalls in causal data analysis such as interpretation of all regression coefficients in a table (known as Table 2 fallacy), who knew this was problematic?!). It also covered other pitfalls such as collider bias, conditioning on the outcome bias, outcome truncation bias, selection bias, composite variable bias, you can see where this is going… if you can think of a type of bias, I am sure it was covered here! Most importantly, the takeaway message of this course was to draw a DAG (directed acyclic graphs), something that I had not previously heard of but wish I had! DAGs are a vital tool to help recognise these biases and more. They are a non-parametric diagram used to assist the planning and execution of estimating causal effects. By creating a DAG, identifying confounders, colliders, mediators and much more becomes very straight forward and if you are analysing relationships, this will really be a lifesaver!

Overall, I would really recommend this course. It used great, varied examples to demonstrate every concept such as a relationship between shark attacks and ice cream sales, to COVID-19 research, including a paper on COVID-19 risk factors which resulted in the French government withdrawing protection from vulnerable workers, and the (many) problems with UK Biobank. The team delivered every lecture with great enthusiasm and passion, making tricky concepts (such as Propensity Scores and g-methods) easier to understand, and convincing everyone in the room that mediation analyses may not be the best approach.

If you want to learn more about causal inference, or what any of these concepts actually mean, keep an eye out for the Winter School in Causal Inference. You can follow the organisers on twitter: @PWGTennant @statsmethods @GeorgiaTomova @kellyn_arnold. But beware, once you are aware of these pitfalls, there’s no going back; see slide above for a word of warning!

If you have any questions, please don’t hesitate to email me – e.r.thornton@liverpool.ac.uk

Twitter: @emma_thornton95