29): Luke Keele, “Approximate Balancing Weights for Clustered Observational Study Designs”

Posted on April 14, 2022October 17, 2024 by Newton, Sarah

RMME/STAT Joint Colloquium

Approximate Balancing Weights for Clustered Observational Study Designs

Dr. Luke Keele
University of Pennsylvania

Friday, April 29, at 3:00PM ET

https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m35b82d4dc6d3e77536aa48390a02485b

In a clustered observational study, a treatment is assigned to groups and all units within the group are exposed to the treatment. Clustered observational studies are common in education where treatments are given to all students within some schools but withheld from all students in other schools. Clustered observational studies require specialized methods to adjust for observed confounders. Extant work has developed specialized matching methods that take key elements of clustered treatment assignment into account. Here, we develop a new method for statistical adjustment in clustered observational studies using approximate balancing weights. An approach based on approximate balancing weights improves on extant matching methods in several ways. First, our methods highlight the possible need to account for differential selection into clusters. Second, we can automatically balance interactions between unit level and cluster level covariates. Third, we can also balance high moments on key cluster level covariates. We also outline an overlap weights approach for cases where common support across treated and control clusters is poor. We introduce an augmented estimator that accounts for outcome information. We show that our approach has dual representation as an inverse propensity score weighting estimator based on a hierarchical propensity score model. We apply this algorithm to assess a school-based intervention through which students in treated schools were exposed to a new reading program during summer school. Overall, we find that balancing weights tend to produce superior balance relative to extant matching methods. Moreover, an approximate balancing weight approach tends to require less input from the user to achieve high levels of balance.