Nate Silver has a post over at 538 about Washington state, and the difficulties with forecasting election results there. I don’t know how pollsters adjust their likely voter models in states with significant early voting (though this seems like an increasingly important question), but I do want to note his comments about the impact of vote-by-mail. Silver contends:
Also — probably because of mail balloting — turnout in Washington and Oregon has generally been very high, so targets that might work well in other states could fail there.
This seems wrongheaded on two fronts. First, as we’ve shown several times, vote-by-mail has caused a very small turnout increase in Oregon and Washington – on the order of a few percent in federal elections. (Although admittedly, we do see a more substantial effect in lower-tier races.) In fact, political participation is part of the culture in these two states, and they have long had high rates of voter turnout.
Second, as turnout increases in a state, the impact of likely voter models presumably becomes less important. Washington and Oregon have levels of turnout way above the national average, most other states, and the pollsters’ baselines for likely voter models.
Update: Gronke posted this reply on the Time blog:
I have to chime in here. Oren is correct about Washington. The big push for vote by mail in Washington State occurred after the Rossi/Gregoire contest (2004). Fewer than 10 counties were using VBM in 2004, so any “house effects” prior to 2006 cannot be attributed to VBM.
Second, many scholars have studied the turnout effects of VBM in federal contests, and the impact is minimal, between 2-4%. Put simply, Oregon (and Washington) were high turnout states prior to the adoption of VBM and continue to be high turnout states. I’m disappointed to see Nate propagate the myth that VBM is a magic bullet for turnout. It’s a common misconception, but one that most election officials now realize is not true.
Finally, while I understand how high turnout and VBM can make the likely voter filter problematic, I don’t understand how it would be predicted to create highly variable results. A just doesn’t link to B in this blog post.
If turnout were high, then isn’t the likely voter filter irrelevant? And if high turnout plus heavy use of non precinct place voting leads to variable polling results, then Nevada, New Mexico, and Colorado should also be displaying these patterns.