See this thread here --- because variance is reduced when weighting on variables that are highly correlated with the outcome, it's likely that all parameters are under estimated.
I'm not sure if there's a solution to this, though. Modeling the polls according to the correct variance would require that the reported MoE is at the same distribution level (e.g., 95%), calculated correctly (which even software like srvyr doesn't seem to do), and is correct for the model's likelihood (e.g., a multinomial moe will be different from a binomial moe and that causes issues)
See this thread here --- because variance is reduced when weighting on variables that are highly correlated with the outcome, it's likely that all parameters are under estimated.
I'm not sure if there's a solution to this, though. Modeling the polls according to the correct variance would require that the reported MoE is at the same distribution level (e.g., 95%), calculated correctly (which even software like srvyr doesn't seem to do), and is correct for the model's likelihood (e.g., a multinomial moe will be different from a binomial moe and that causes issues)