Project Summary

In comparative effectiveness research (CER) and patient-centered outcomes research (PCOR), evidence from randomized trials is generally regarded as the gold standard but can often be infeasible for either practical or ethical reasons. For example, randomizing patients to aspects of critical care (e.g., life support) violates the principle of clinical equipoise. When randomized trials are infeasible, non-randomized studies are the primary alternative but yield biased treatment effects estimates when treated and control units are not comparable. This is often due to unobserved confounders, such as unobserved variables that affect both the treatment and the outcome.

Bias from unobserved confounders can be mitigated by using quasi-experimental devices—methods design to reduce bias from hidden confounders without using randomization. For example, if sicker patients go to hospital A more often than to B, A may appear as performing worse even if, in fact, it delivers better care. If, however, distance from patients’ home to hospital affects hospital choice and is essentially as good as randomized, a principled comparison between hospitals can be done with the appropriate device. In this project, the team focuses on new statistical methods for one prominent device: the instrumental variables (IV) design. The team will address key limitations of current practice by developing, testing, and disseminating innovative statistical theory and methods that incorporate modern machine learning (ML) estimation methods.

The team will focus on weaknesses in three specific areas. (a) Robust estimation methods for IV designs. Currently, most researchers implement IV designs with methods that are prone to bias from model misspecification. The team will develop and test nonparametric estimators for IV designs that are doubly robust to model misspecification bias. The team’s estimators are based on machine learning methods but are optimally efficient. (b) Robust estimation methods for covariates-assisted bounds. Key IV assumptions can be relaxed using bounds estimators. Current bounds estimators for IV designs cannot incorporate baseline covariate information, which makes them overly conservative. Thus, the team will develop doubly robust, nonparametric methods for bounding estimators in IV designs, which can accommodate complex covariate information to provide more accurate estimates. (c) IV designs with a continuous instrument. Many IV in CER applications are continuous measures. Currently, most researchers use ad-hoc statistical methods that are either prone to model misspecification or fail to fully extract the information available in the data. The team will use the dynamic intervention framework to develop improved methods for IV designs with continuous instruments. The team will derive new causal quantities and statistical methods that are doubly robust and nonparametric.

Finally, the team would apply these new methods to real data applications. Using methods from Aim 1, the team would estimate the effect of surgical treatments for emergency conditions. Using methods from Aim 2, the team would estimate the effect of life support interventions for low-birthweight newborns. Using methods from Aim 3, the team would estimate the effect of treatment at a Level I or II Trauma Center for emergency general surgery patients.

Project Information

Luke Keele, PhD
The Trustees of The University of Pennsylvania

Key Dates

December 2021
July 2026


Award Type
Health Conditions Health Conditions These are the broad terms we use to categorize our funded research studies; specific diseases or conditions are included within the appropriate larger category. Note: not all of our funded projects focus on a single disease or condition; some touch on multiple diseases or conditions, research methods, or broader health system interventions. Such projects won’t be listed by a primary disease/condition and so won’t appear if you use this filter tool to find them. View Glossary
State State The state where the project originates, or where the primary institution or organization is located. View Glossary
Last updated: November 10, 2022