Project Summary

Background: Patient-centered outcomes research, which is a type of comparative clinical effectiveness research, is increasingly conducted using electronic health record (EHR) databases. When estimating treatment effects using such data, we must adjust for covariates in order to isolate the effect on the outcome attributable only to the intervention rather than these covariates, i.e., the causal effect of the intervention. However, important covariates are not measured in EHR for all patients, leading to missing covariate information for some patients. Existing methods for missing covariates in causal settings can perform poorly when there are many covariates with missingness. Moreover, like all causal inference methods, they rely on untestable assumptions about the models and the nature of intervention decision process. Estimates can be highly inaccurate if these assumptions are violated, yet there is a lack of work on methods for assessing sensitivity to such violations. 

Objectives: This study will develop a set of statistical methods for estimating causal effects of interventions in the presence of missing covariates. Specifically, the team has four objectives: 

  • Develop methods that can provide accurate and stable estimates in settings with many covariates with missingness 
  • Develop flexible methods that are less sensitive to misspecification of the requisite data models 
  • Provide a framework for assessing sensitivity of estimates to potential violations of untestable assumptions 
  • Use the developed methods to generate new clinical insight about the effects of clinician-initiated deliveries on maternal and infant outcomes 

Methods: To achieve the objectives above, researchers will develop Bayesian statistical methods that: 

  • Combine data with prior information to improve estimation quality in complex settings with many missing covariates 
  • Rely on flexible Bayesian machine learning methods that relax untestable modeling assumptions, yielding estimates that are less sensitive to model misspecification 
  • Provide researchers with sensitivity analysis methods for gauging how robust intervention effect estimates are to violations of untestable assumptions about the intervention decision process 
  • Will be applied to EHRs from the Women & Infants Hospital of Rhode Island to estimate effects of clinician-initiated deliveries 

Anticipated Impact: Results from this work will yield reliable, principled statistical methods for estimating intervention effects using EHRs with missing covariate information. Given the large scope of clinical applications using EHRs, these methods have potential for broad impact on a range of studies. The research team will develop user-friendly open-source code implementing these methods which it anticipates will amplify this impact. Finally, the team’s analysis of maternal and infant outcomes in Women & Infants Hospital of Rhode Island EHRs will benefit from these methods because many important covariates in this data are missing. This will impact clinical practice in maternal care settings.

Project Information

Arman Oganisian, Ph.D.
Roee Gutman, Ph.D.
Brown University School of Public Health
$920,991 *

Key Dates

36 months *
November 2023

*All proposed projects, including requested budgets and project periods, are approved subject to a programmatic and budget review by PCORI staff and the negotiation of a formal award contract.


Award Type
State State The state where the project originates, or where the primary institution or organization is located. View Glossary
Last updated: January 24, 2024