Project Summary
One of PCORI’s goals is to improve the methods that researchers use for patient-centered outcomes research. PCORI funds methods projects like this one to better understand and advance the use of research methods that improve the strength and quality of comparative effectiveness research.
What is the project about?
A randomized controlled trial, or RCT, is often the best way to learn if one treatment works better than another. RCTs assign patients to different treatments by chance. When RCTs aren’t possible, researchers can use electronic health record, or EHR, data to study treatment effects over time. But EHR data often don’t capture differences in treatment patterns between patients. For example, they may not capture when patients take more than one treatment or switch treatments at different times. Also, EHR data may not have complete information about patients, which can bias study results.
In this study, the research team is developing and testing new methods for analyzing EHR data to improve how well they work to study treatment effects on patient outcomes over time. The new methods account for differences in treatment patterns between patients and missing patient data. The team is using machine learning to develop the methods. In machine learning, computers use data to learn how to perform different tasks with little or no human input.
How can this project help improve research methods?
Results may help researchers study treatment effects over time using EHR data.
What is the research team doing?
The research team is developing machine learning methods to study treatment effects. These methods account for:
- Differences in treatment start time
- Taking more than one treatment or switching treatments over time
- Patient traits that change over time
- Bias due to patients leaving the study early
- Errors in the statistical model
- Bias due to missing patient data that could affect treatment outcomes
The research team is testing the new methods using data from the Yale New Haven Health System. Data are from research studies of two health problems with treatments that may vary over time: COVID-19 and hypertension. The team is also creating free software for using the methods.
Research methods at a glance
Design Element | Description |
---|---|
Goal | Aim 1. To develop a new robust, marginal structural quantile model to estimate the causal effects of longitudinal treatments across the distribution of outcomes and further develop its machine learning extensions. Aim 2. To develop a Bayesian likelihood-based machine learning method that can accommodate time-varying covariates to estimate a set of weights for correcting time-varying confounding or selection bias due to informative censoring. Aim 3. To develop a flexible and interpretable sensitivity analysis framework for estimating causal effects adjusted for the posited amount of unmeasured confounding over time. Aim 4. To apply the methods developed in Aim 1 and Aim 2 to two conditions with treatments that may vary over time: COVID-19 and hypertension, and to develop open-source software packages within the R computing platform to assist the implementation of the new methods. |
Approach | Causal inference, survival analysis, Bayesian analysis, machine learning, sensitivity analysis, software development |