|
Activity Number:
|
383
- New Developments in Sensitivity Analysis for Unmeasured Confounding
|
|
Type:
|
Invited
|
|
Date/Time:
|
Tuesday, July 31, 2018 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Statistics in Epidemiology
|
|
Abstract #326580
|
Presentation
|
|
Title:
|
Sensitivity Analysis in Multilevel Models
|
|
Author(s):
|
Nicole Bohme Carnegie* and Jennifer L Hill and Masataka Harada and Vincent Dorie
|
|
Companies:
|
Montana State University and New York University and Fukuoka University and New York University
|
|
Keywords:
|
Grouped data;
Hierarchical model
|
|
Abstract:
|
A number of approaches exist for sensitivity analysis when data are independent and identically distributed (iid), but to our knowledge none exist for grouped data structures. We introduce a two-parameter sensitivity analysis approach similar to one we have proposed for iid data (using linear models and Bayesian Additive Regression Trees) that can accommodate grouped data in multilevel models. We perform sensitivity analysis by generating potential confounders from their conditional distribution given coefficients in the treatment and response models and re-estimating the treatment effect including this potential confounder. This stochastic approach yields a distribution of treatment effects over plausible ranges for sensitivity parameters. We then compare these distributions to benchmarks generated from the observed data to assess the robustness of the treatment effect estimate to potential unobserved confounding. The method is demonstrated on applications to the effect of ethnic fragmentation on the distribution of public goods and the effect of a cooking skills intervention on youth fruit and vegetable consumption.
|
Authors who are presenting talks have a * after their name.
