| name | lemma-causal |
| description | Rigor for causal questions and A/B tests (the effect of acting on X): confounding, post-treatment bias, valid control groups. |
| homepage | https://github.com/tkpratardan/lemma |
| license | MIT |
Causal inference
Deliver
State the design, target population, estimand, effect with uncertainty,
diagnostic evidence, identifying assumptions, and an action-limited
conclusion. If the assignment story is unknown, report association only.
Check
- Establish treatment, outcome, unit, timing, assignment mechanism, and
counterfactual comparison.
- Preserve randomized assignment and exclude post-treatment variables from
adjustment.
- For observational work, state identification assumptions before choosing
covariates or an estimator.
- Inspect balance, overlap, interference, attrition, compliance, outcome
measurement, and treatment timing as applicable.
- Challenge the weakest assumption with a pre-trend, placebo, balance, or
sensitivity check.
Do not call correlation causal, control for mediators or colliders, choose
groups after seeing outcomes, or bury failed overlap or pre-trends.
For design diagnostics and sensitivity methods, read
references/deep-guide.md.
