Election forecasting has, in recent years, come to rely on models like those described in Jackman (2005):
y_{i} \sim \mathrm{Normal}(\mu_{i}, \sigma_{i}) \\ \mu_{i} = \alpha_{t[i]} +\delta_{j[i]} \\ \alpha_{t} \sim \mathrm{Normal}(\alpha_{t-1}, \omega), t = 2, ..., T
Here, poll y is modelled as the function of some grand mean \alpha that varies over each day in the data and a set of house effects, \delta, due to persistent biases present in each polling company’s methods.
Importantly, \alpha_{t} is itself modelled as a function of \alpha_{t-1}. I know that brms can handle autoregressive elements, but can it do so for anything other than the dependent variable?
It’s worth noting also that often there are missing days in the data too.