Propensity Score Methods for Studies with Hierarchical Data Structure and Continuous Treatments: Application to Childhood Obesity Interventions in Los Angeles County

• Activity Number: 337 - Causal Inference for Complex Data Challenges
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Justin R Williams* and Catherine Crespi and May Wang
• Companies: UCLA and University of California, Los Angeles and University of California, Los Angeles
• Keywords: propensity score; causal inference; hierarchical data; continuous treatment

Abstract

Generalized propensity score (GPS) methods are used to estimate the causal relationship between a continuous exposure and an outcome of interest in nonrandomized experiments. The GPS quantifies the likelihood that a subject received a specific level of treatment given a set of potential confounders. Using the GPS, an average dose-response function is estimated across fixed treatment values. Most current GPS methods rely on the assumption that there is independence between units. However, in many complex data settings this assumption does not hold. To estimate a dose-response function for clustered data, for example, correlation between units that share the same cluster must be considered. We present a method for GPS estimation with clustered data that uses a linear combination of the GPS value along with the level of exposure to estimate an average dose-response function via a linear mixed model. We apply this method to a three-level hierarchical dataset with a continuous exposure applied at the upper-most level to estimate the effect of neighborhood level obesity interventions on obesity outcomes among low-income pre-school aged children in Los Angeles County.