Integrating Causal Inference and Deep Learning for Analysis and Interpretation
Abstract
AbstractApproved for Public ReleaseCausal inference seeks to assess the expected impact of possible interventions (e.g., treatments) on a system of interest. Models developed for this purpose are typically designed based on observational data connected to prior (historical) decisions of the same type. A significant challenge in modeling based on observational data is confounding, manifested because such data are not randomized. The implementation of randomized controlled trials (RCTs) is impractical for most situations ofinterest to the Navy, and therefore there is no choice but to learn from observational data. Observed confounding is manifested when both the selected intervention and the system output are dependent on observed covariates; even more challenging, unobserved confounding occurs when the intervention and outcome also depend on unobserved latent phenomena (unobserved in the covariates, which are the only information the analyst has about the system from observational data). In prior ONR research by the PI, observed confounding was addressed through development of a new class of weights applied to the observational training data, coupled with balanced predictive feature extraction from the covariates. That work ignored unobserved confounding, which we wish to address in the proposed research. A general class of proxy data will be introduced, meant to illuminate latent confounders. Utilizing recent advances in machine learning, in the proposed research proxy data will be quite general, composed of images and documents, as well as more-traditional tabular data. The proposed research will leverage state-of-the-art models for such data: deep neural networks for image analysis and deep large language models for documents, with these coupled in novel ways to causal models. The presence of proxy data offers the potential to mitigate unobserved confounding, and the objective will be to balance the statistics of features from observed covariates and from proxy data, with balance achieved across the treated and untreated training data. A challenge to be addressed concernsthe fact that the exact form of the proxy data may vary between different observational data (e.g., some historical cases may be characterized by many images and a few documents, while other cases could have the opposite balance of proxy data); the methodology that infers latent features from proxy data must be robust to such variability. The predictive model of mean potential outcomes will be a function of features from the covariates and features from proxy data. It is hoped that the latter willbetter address outliers in the treatment assignments, which were poorly handled via prior weighting schemes, as the outliers (from the perspective of covariates) are potentially attributable to previously unobserved confounding. Since the sophisticated and possibly heterogeneous proxy data will be analyzed with generative deep neural networks and associated inference networks, an important issue concerns learning to prioritize hierarchical proxy-data features for causal inference (inferring which layers of hierarchical feature abstraction are mostrelevant for prediction of a particular causal effect). Within the proposed research, we will conduct an extensive set of experiments on both synthetic examples and available public benchmark data, with a focus on applications of interest to the Navy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2024
- Source ID
- N000142412246
Entities
People
- Lawrence Carin
Organizations
- Duke University
- Office of Naval Research
- United States Navy