| English | 中文 |
Use causal design methods before implementation: define treatment, outcome, temporal order, physical mechanisms, confounding, alternative DAGs, overlap and sensitivity checks. Graph validity does not prove identification.
The R causal/causal.R branch is selected by analysis=causal_plr. It replaces the prediction fit/evaluate/explain branch while sharing target and preparation modules. It requires a declared DAG, manually justified adjustment set equal to features, one numeric treatment, units, treatment scale, exactly one nuisance candidate, missing.policy=drop and six justified assumption declarations. Do not check assumptions merely to pass validation.
The retained train+validation pool is cross-fitted; test rows remain excluded. Use 2–10 folds and at least max(30,10*folds) rows; group mode requires max(5,folds) groups and an independent-group assertion. These implementation guards do not establish adequate effective sample size or causal validity. R operations retains the complete invocation/configuration, coefficient-unit definition and output roles. Read causal effect, residuals, folds, nuisance models and DAG through artifacts.json, and report identification limits. Python predictive candidates do not provide this R PLR branch.