Impact of specific modeling choices (fixed and varying effects) on the observed bias

Determine whether including Socio-economic Index for Areas (SEIFA) both as a non-varying (fixed) effect and a varying (group-level) effect, together with modeling sex as a non-varying effect, is causally related to the observed downward trend in estimated prevalence when additional covariates are added in the Bayesian multilevel regression and poststratification framework that jointly estimates diagnostic test sensitivity and specificity for SARS-CoV-2 seroprevalence.

Background

The paper investigates an unexpected downward trend in estimated seroprevalence as covariates are added to a Bayesian multilevel regression and poststratification (MRP) model that incorporates uncertainty in diagnostic test sensitivity and specificity.

In discussing modeling choices, the authors note differences from prior guidance (Bafumi and Gelman) and highlight that SEIFA is encoded both as a fixed and varying effect, while sex is treated as a fixed effect because it is binary. They explicitly state uncertainty about whether these choices contribute to the observed artifact and therefore run simulations to explore their impact.