• Realistic performance evaluation of a large set of causal discovery methods. • Methods struggle with data that has a low signal-to-noise ratio. • Application of causal structure discovery to data from UK-Biobank study. Causal analyses are predicated on known or hypothesized causal structures, which are challenging to obtain for exposome studies. A possible solution is to use data-driven causal structure discovery (CSD) approaches, but their performance for exposome studies is understudied. We explored the potential of CSD algorithms to infer causal structures from typical exposome data. We compiled an overview of available algorithms and evaluated their performance on synthetic data resembling that of the UK-Biobank study. We used bootstrapping and model aggregation to account for sensitivity to variable ordering. We applied top-performing algorithms to UK-Biobank data and explored the causal effect identifiability of particulate matter PM 2.5 on the incidence of cardiovascular disease (CVD) as a potential application of the estimated structure. We evaluated 20 methods and noted significant variation in performance, with average F1 scores ranging from 0.1 to 0.9. The best-performing algorithms were TABU (average F1 95% Quantile Interval = 0.64 0.30,0.86), HC (0.63 0.31, 0.87), H2PC (0.62 0.30, 0.84), MMHC (0.61 0.31, 0.84), and GRaSP (0.60 0.33, 0.86). Combinatorial methods tended to outperform continuous optimization-based algorithms. Causal structures estimated for the UK-Biobank data were highly interconnected and included both established causal relations and some that appeared implausible. Structures estimated using methods with diverse inferential techniques were significantly different and implied distinct adjustment sets for estimating the causal effect of PM 2.5 on CVD. Most CSD algorithms rely on strict assumptions that are challenging to meet in observational studies. On realistic exposome data simulated with a relatively large sample size but a low signal-to-noise ratio, all implementations struggled to perform well. Widely available causal structure discovery algorithms could be used as an exploratory tool in exposome studies, e.g., to generate potential causal structures and inform causal hypotheses. Evidence triangulation could likely improve performance, i.e., including prior information in the discovery process and validating estimated structures.
Kakhaia et al. (Fri,) studied this question.