Autoite – an Algorithm for Estimating Ite

Refer to the "paper" folder for rigorous documentation on my insights and why the repository works.

The general philosophy is, the effect of treatment T on outcome Y given group G of controls X is defined by latent environment U, where a latent environment is based on intrinsic similarity in how samples behave, not physical labels such as "hospital A".

As ridge is the first component in the pipeline, it has proven to be extremely rigorous against noise and increase features, and given proxies are required, a "better to be safe than sorry" approach is preferred, meaning more features which is in contrast to typical causal inference.

You fit a dataset, then get ITE predictions for each sample. A thresholding approach also exists, I advise reviewing my paper for its mechanism and utility.

Real world data is difficult to procure for testing counterfactuals, so only one dataset of real world data is used, in which we don't know the true ground truth.

Feedback is most appreciated.