- Therapy Impact Estimation utilizing Invariant Danger Minimization(arXiv)
Authors: Abhin Shah, Kartik Ahuja, Karthikeyan Shanmugam, Dennis Wei, Kush Varshney, Amit Dhurandhar
Summary: Inferring causal particular person therapy impact (ITE) from observational knowledge is a difficult drawback whose problem is exacerbated by the presence of therapy project bias. On this work, we suggest a brand new option to estimate the ITE utilizing the area generalization framework of invariant threat minimization (IRM). IRM makes use of knowledge from a number of domains, learns predictors that don’t exploit spurious domain-dependent components, and generalizes higher to unseen domains. We suggest an IRM-based ITE estimator aimed toward tackling therapy project bias when there’s little assist overlap between the management group and the therapy group. We accomplish this by creating variety: given a single dataset, we cut up the information into a number of domains artificially. These various domains are then exploited by IRM to extra successfully generalize regression-based fashions to knowledge areas that lack assist overlap. We present positive factors over classical regression approaches to ITE estimation in settings when assist mismatch is extra pronounced.
