Casual discovery: Data-driven witchcraft or a useful tool for constructing casual models for cohort studies?

In observational sciences (e.g., epidemiology), causal models such as DAGs are often used for describing causal mechanisms and informing effect estimation strategies. Especially in life course studies, where the same individuals are followed for long periods of time, constructing such models can be a cumbersome and difficult activity. Moreover, one must refer to established empirical or theoretical results to construct the DAGs, which gives the procedure little room for uncovering new causal pathways: It is fundamentally confirmatory.

Causal discovery constitutes a data-driven alternative to this traditional approach. Here, machine learning algorithms are applied to empirical data to deduce what information about the causal data generating mechanism can be recovered, and the results are typically represented as a family of possible DAGs. In recent years, we have developed modified causal discovery algorithms, including temporal PC and temporal GES, that are specifically tailored towards efficiently using the longitudinal temporal structure in cohort studies. 

In this presentation, I provide a general introduction to (temporal) causal discovery algorithms. I furthermore present a study investigating the utility of causal discovery in life course epidemiology by comparing their performance with a traditional approach to constructing causal models: namely experts drawing directed acyclic graphs with reference to previous study and general theory. This investigation focused on a life-course study concerning etiology of depression and heart disease in early old age for a cohort of Danish men followed from birth (1953) until 2018 (The Metropolit Cohort). We compared the resulting suggestions for causal models and used this application example to discuss when and how causal discovery algorithms may be useful in life course epidemiology and beyond. If time permits, I will conclude by showcasing a few additional applications of causal discovery to cohort datasets.