Uniform inference on continuous treatment effects via artificial neural networks in digital health

This research project will provide a general formulation and a deep learning-powered toolbox for conducting causal analysis of continuous treatment effects in observational research. Digital health innovations have paved the way for the collection of large-scale observational data. In practice, many empirical applications in healthcare programs involve continuous treatments. This project meets the immediate needs of practitioners seeking flexible and powerful statistical tools for conducting causal inference of continuous treatment effects based on large and heterogeneous digital health data. Students, especially from underrepresented groups, will be recruited to participate in the research. Easy-to-implement software packages will be developed and made publicly available. The research results will equip scientists and healthcare providers with principled analysis for making treatment recommendations, so as to improve patient care and reduce costs. Advanced digital technologies powered with a reliable deep learning toolbox will revolutionize healthcare analytics. The research will also promote collaborations with scientists from Medicine, Public Health, Engineering, and Social Sciences. In addition, the project will provide research training for graduate students. This project will develop new statistical methodologies and the associated theories for conducting uniform causal inference of continuous treatment effects via deep learning. It will pursue three specific research topics, and the developed methods will be used to solve a wide range of causal problems. Specifically, in the first topic, the project will develop a variety of neural network architectures to approximate the nuisance function for suitable data applications in digital health. In the second topic, the project will estimate the balancing weight using neural networks through generalized optimization, and construct simultaneous confidence bands for the dose-response curve for inference. In the third topic, the project will apply the proposed optimization procedure to the estimation of heterogeneous treatment effects, and to the longitudinal data setting. The research will provide a new perspective on estimating general continuous treatment effects using deep neural networks. It will provide a powerful tool for causal analysis that combines the advantages of deep learning, direct covariate balancing, and generalized optimization. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.