Causal diagrams, which are almost always directed acyclic graphs, or DAGs, have been increasingly recommended as a tool that can help identify biases, select model variables and make assumptions more explicit. Reviews and experience would suggest, however, that most health researchers, including biostatisticians, do not currently use them.
The idea for an online searchable database of example causal diagrams came partly from my own initial difficulties during my recent biostatistics PhD.
My early experience creating causal diagrams
My project was focused on causal inference but, like most people who have studied statistics to date, the concepts and methodology of causal inference were, at first, largely unknown to me. Hence, when I wanted to create a causal diagram for a study involving blood glucose monitoring, I had only seen example DAGs from introductory material. And introductions, by necessity, use simple examples to introduce new concepts.
While I understood example DAGs like this:
I was unsure how to create a diagram for my own project with all the variables available for my analysis.
Eventually, it looked like this:
Creating a DAG – but how to start?
At first, I had little idea of what the final diagram would look like. I had no examples to follow that were in any way similar to the one I wanted to create and hence, I could not form even a vague mental picture of the DAG that was my goal. And my colleagues were likewise unfamiliar with DAGs.
On top of this uncertainty about what I would create, a related problem was not having a clear way to proceed. With experience, people tend to know which methods they prefer and how to use them. But without such experience, a novice such as myself had to consider many possible options:
- Using pencil and paper quickly got messy when I wanted to change how the variables were arranged
- Hence, software seemed necessary, but which to try?
- Or other software like:
- Which variables and relationships to include?
- How should I draw each variable?
- And how might the variables in the diagram initially be organised?
It occurred to me that experiences like this may be why some researchers do not use DAGs.
The need for cognitive ease – insights from cognitive science
Later, I delved into the cognitive science literature to better understand how people make decisions and the factors that influence those decisions.
One finding is our built-in desire for cognitive ease; also called the ‘principle of least effort’ and ‘avoidance of cognitive demand’, among others.234 For example, thinking that requires genuine effort is often mildly unpleasant. So, we frequently avoid it – especially when tired or insufficiently caffeinated.
This likely evolved to help our brain manage a limited resource: energy.5 Nevertheless, goals of sufficient importance will motivate the effort required. While, on the other hand, familiarity with the steps of a task will lower the effort needed to perform that task.
Aim: to decrease the mental effort required to create a causal diagram
To create a causal diagram without prior experience involves decisions from many plausible options, thus creating a high demand on our cognitive resources.
Better familiarity with causal diagrams, however, such as from prior experience, or from relevant examples, means many of those options can be avoided which lowers the mental load and hence, the effort required.
This familiarity also makes it easier to form a mental picture of the goal.
To put this idea into practice, this website, causaldiagrams.org, has been created.
Its most important feature is a searchable database of published health research articles that include a causal diagram. Information extracted from each article includes the exposure and outcome variables of each diagram, as well as the number of variables or nodes, to help researchers locate the diagrams of most relevance to them. The list of articles can also be filtered by selecting within dropdown boxes.
Further details of each article can be seen by clicking the plus symbol, including a link to the publisher’s website; however, copyright concerns prevent the showing of images from those articles. Hopefully, this will not prevent the database being used.
Another potential issue might be the unintended promotion of DAGs that are missing important sources of confounding or selection bias. But this is still an unresolved issue, and may not be one that can be resolved.
- Watkins TR. Understanding uncertainty and bias to improve causal inference in health intervention research [PhD thesis]: University of Sydney; Sydney, Australia, 2019. http://hdl.handle.net/2123/20772
- Kool W, McGuire JT, Rosen ZB, Botvinick MM. Decision making and the avoidance of cognitive demand. Journal of Experimental Psychology: General. 2010;139(4):665-682. doi:10.1037/a0020198
- Kahneman D. Thinking, Fast and Slow . New York, NY: Straus & Giroux; 2011.
- Reber R, Greifeneder R. Processing Fluency in Education: How Metacognitive Feelings Shape Learning, Belief Formation, and Affect. Educational Psychologist. 2017;52(2):84-103. doi:10.1080/00461520.2016.1258173
- Inzlicht M, Shenhav A, Olivola CY. The Effort Paradox: Effort Is Both Costly and Valued. Trends in cognitive sciences. 2018;22(4):337-349. doi:10.1016/j.tics.2018.01.007