GitHub

Output quality

DAGMakie aims for notebook-scale causal diagrams (roughly 5–200 labelled nodes) that read closer to Graphviz dot / dagitty than to a generic spring layout.

Defaults

  • layout_mode = :auto — layered for DAGs; SCC-aware for cyclic digraphs
  • :spring — explicit NetworkLayout fallback
  • Pass layout = … for fixed positions or another NetworkLayout algorithm
  • dagplot_do_comparison / intervention helpers reuse a shared DAGLayoutResult

Robustness features

  • Edges are straight by default. Per-edge :curved, CurvedEdge(), or a bow fraction bows edges with quadratic Bézier samples (long_edge_routing = :quadratic); other styles include :natural_cubic, :none, :rounded, :tangents, :curve_distance
  • Feedback edges inside an SCC use curved overlays with boundary-aware arrowheads
  • Bidirected confounding arcs are clipped to node markers
  • Axis limits include markers, labels, and overlay waypoints
  • Highlight specs use first listing wins when the same node or edge is repeated (put exposure / outcome / adjustment before path colours)

Checks

  • Geometry regressions live in test/test_layout.jl, test/test_interventions.jl, and test/test_highlighting.jl
  • Optional timing budgets: set DAGMAKIE_PERF=1 when running the test suite (see test/test_perf.jl)
  • Manual timing gallery: julia --project=. test/benchmark_layouts.jl

Non-goals

  • No hard Graphviz dependency (optional local dot comparison is fine in scripts)
  • Not a WebGL explorer for huge graphs