The challenge of model specification

I have been playing with different models and I have reached a conundrum. My anchored primary vote model appears correct on paper, but yields a result inconsistent with other models. The anchored model seeks to locate the primary vote polls in terms of their performance at the last election. The unanchored model assumes that collectively the pollsters are unbiased (as a consequence, house effects sum to zero), and the average poll result is a good indicator of hidden voting intention.

The problem, in part, is the poor performance of the primary vote polls prior to the last election. The Greens vote was substantially down on what the immediate pre-election polls suggested. The Coalition and Other Party votes were substantially up.

Let's compare the primary vote charts for the election-anchored and sum-to-zero models. You will note that the anchored model has a wider credibility interval (and therefore offers less certainty around the median sample).

We can apply preference flows at the last election and the primary vote estimates from both models to estimate a Coalition two-party preferred result. Again we get different results.

Clearly, I have a bit more thinking to do on this.