Primary Submission Category: Sensitivity Analysis
Covariate Measurement Errors as An Omitted Variable Bias Problem
Authors: Minh Duy Pham, Chad Hazlett,
Presenting Author: Minh Duy Pham*
A necessary but rarely stated assumption in causal inference is that all variables, especially the covariates we condition on to identify the causal effects, are measured without error. However, in reality, measurement errors are endemic, and we often adjust for error-prone proxies or measurements of the true confounders instead. In this work, we approach the problems of covariate measurement errors as one of residual unmeasured confounding: conditioning on an error-prone proxy of a confounder will most likely not fully adjust for its confounding effects on the treatment and the outcome, leaving (a portion of) it unobserved. If so, how unreliable must our measurements be for the residual confounding to ruin our treatment effect estimates? We extend and reparameterize the omitted variable bias approach in Cinelli and Hazlett (2020) to describe how measurement (un)reliability influences the bias in an intuitive and interpretable manner. In doing so, we highlight the merits of and recommend a novel approach of applying sensitivity analysis to covariate measurement error problems.