Estimating Causal Treatment Effects: A Theoretical and Computational Framework for Model Evaluation, Selection and Development
Abstract
The proposal addresses a fundamental question of measuring performance of various algorithms for the challenging and important topic of Causal Inference. (Causal Inference ~ statement that ~something is causing something else~ - drawing a conclusion about a causal connection based on the conditions of the occurrence of an effect). What is proposed is development of an original and novel framework for evaluating and ranking different Causal inference algorithms, and developing the capability to select the best one for the given set of data ~ a capability which does not currently exist. Developing a method for validating and ranking various causal inference models is an important step for translating advances in causal inference into practice, especially recent advances in machine learning-based causal inference. It also addresses critical issues like performance guarantees and provable optimality, and is unique in the sense that there are currently no other approaches that are capable of providing this level of performance. It is based on building upon the recent work of the PI, in which some critical measures and performance evaluation metrics (like CATE and PEHE, the main evaluation metric, see later) had been developed that are used in the proposed work to achieve the main goal ~ estimates of evaluation metric of any causal inference model via a statistic that does not depend on the counterfactual data (~ground truth~) ~ the main goal of the proposal. The proposed work expands on these results, and is founded on solid theoretical foundations vetted by the community in recent publications and conferences. In addition, some preliminary results are shown, which indicate that the approach is valid and effective. The proposal is very well written, thorough in development of various aspects of the proposed solution, with detailed discussions and with logical step by step exposition and descriptions of the proposed approach. The offeror is a leader in this nascent field, well recognized and with a significant number of relevant publications and conference presentations. Actually, Mihaela van der Schaar has an outstanding professional record. She is currently John Humphrey Plummer Professor of Machine Learning, Artificial Intelligence at Cambridge University, and Medicine Turing (Faculty) Fellow, Alan Turing Institute, London, UK. Her professional record is so extensive (with hundreds of presentations and published articles, various awards, 35 patents, etc.) that it is not possible to cite it here at any length, but details are given in the proposal on 32 pages.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 15, 2019
- Source ID
- N629091912091
Entities
People
- Mihaela Van Der Schaar
Organizations
- Office of Naval Research
- United States Navy
- University of Cambridge