October 22, 2025 | Noon to 1:00 PM EST
Michael A Bailey (Georgetown)
Many survey researchers worry that contemporary surveys can be biased due to non-ignorable nonresponse (NINR), which can persist even after quota sampling, weighting, or other covariate-based adjustments. NINR bias occurs when the propensity to respond is correlated with the outcome of interest, conditional on covariates. Because panel survey data provides multiple observations for individuals across panel waves, we can identify the correlation between response propensity and the outcome of interest. To date, however, most panel survey data is assessed as cross-sectional data. This paper makes three contributions. First, we present a random effects panel selection model that explicitly accounts for NINR. Second, we present simulations that demonstrate that the model outperforms conventional methods when NINR is a problem. Third, we show that our panel selection model outperforms conventional methods when applied to survey data from the 2020 election cycle.