Conference Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 386 - SPEED: Statistics in Epidemiology Part 1
Type: Contributed
Date/Time: Wednesday, August 10, 2022 : 8:30 AM to 10:20 AM
Sponsor: Section on Statistics in Epidemiology
Abstract #323510
Title: Prioritize Variables with Joint Variable Importance Plots in Observational Study Design
Author(s): Lauren Liao* and Amanda Ngo and Yeyi Zhu and Sam Pimentel
Companies: University of California, Berkeley and Kaiser Permanente Northern California and Kaiser Permanente Northern California and University of California, San Francisco and University of California, Berkeley
Keywords: causal inference; weighting; health outcomes; observational study design; matching; propensity score
Abstract:

When measuring treatment effects in observational studies, it is essential to remove any confounding between the treated and control groups. Typically, likely confounders are identified by measuring each variable’s imbalance across treatment groups. However, not all variables are equally important in the outcome model, and imbalance-focused approaches do not address outcome importance systematically. We propose joint treatment-outcome variable importance plots based on pilot samples to identify variables important for both treatment assignment and outcome. These plots can be augmented with bias curves (computed using the omitted variable bias framework) to better guide prioritization of variables in study design; they also suggest natural choices for tuning parameters in existing matching and weighting methods. We demonstrate this tool on a population-based cohort with gestational diabetes to evaluate the impact of different treatment modalities (pharmacotherapy vs. medical nutrition therapy) on birth outcomes.


Authors who are presenting talks have a * after their name.

Back to the full JSM 2022 program