Causal Graphs vs. Causal Programs: The Case of Conditional Branching

We evaluate the performance of graph-based causal discovery algorithms when the generative process is a probabilistic program with conditional branching. Using synthetic experiments, we demonstrate empirically that graph-based causal discovery algorithms are able to learn accurate associational distributions for probabilistic programs with contextsensitive structure, but that those graphs fail to accurately model the effects of interventions on the programs.