Tracking Causal Inference Experiments in W&B

Start with hotel booking, then maybe adapt diabetes readmission dataset

Model evaluation table: 1 row = causal model, target estimands, estimated effect, and refutation

3-4 different rows

Sweeps: sweep across different estimation methods and hyperparameters, possibly multiple sweeps per model due to differing hyperparameters

Separate Sweeps for each: Baseline linear matching and stratification. train_DML train_DR train_Meta

Artifacts: graph, train, test, interventions using do-Sampler

Log a table of: Causal graph, estimation method, train_ATT, train_ATE, train_ATC, CATES for effect modifier, internal model scores, and results on validation set, Refutation strategy, resulting sensitivity

Stretch: deep learning algo with keras logger

conda install packages - python 3.8 brew install graphviz conda install pygraphviz brew install lightgbm conda install numba pip install econml