What it costs to get the causal structure wrong, measured on data where the true effect is planted and therefore known exactly.
| project | what is measured | result |
|---|---|---|
| 01_adjustment_choice | bias from adjusting for a confounder, collider, mediator or instrument | the same adjustment removes a 25% bias in one structure and creates a 75% bias in another; R² prefers adjusting in all four and is right in one |
Data comes from linear structural equation models with chosen coefficients, so "bias" is
the estimate minus a number that was set, not a comparison against another estimator.
Every claim in a project README is checked against that project's results.json before
it ships.
No neural networks, no GPU. NumPy and SciPy.
MIT, see LICENSE.