Nate Silver gives five reasons that Republicans might do better than his model predicts and five reasons that Democrats might do better.
I add two more possible reasons that Democrats might do better after the jump. They are numbered 3.1 and 3.2, because they are variants on his third reason
More on likely voter filters.
3.1 Likely voter filters include to many people over 65 compared to people age 30-40.
3.2 Likely voter polls are too far from registered voter polls when they are very far from registered voter polls.
Long arguments for both below. First 2 facts support LV polls. 1 They didn’t have a meaningful pro-Republican bias in the past. RV voter polls therefore had a pro Democratic bias in the past of around 4% (or is that calculated with this elections congressional generics ? I got the number here). 2 They are on average closer to the outcome than RV polls.
So why might it be different this time ? One is Silver’s argument 3.
3.1 Another reason is that much of the success of LV filters might be due to their indirectly conditioning on age. Age has a huge effect on the probability of voting. Many LV questions clearly relate to age. For example \”I always vote\” does not mean \”I have voted in every election since I turned 18\” but \”I have voted in almost all general elections since so long ago it doesn’t matter.\” It must be a slowly changing function of past voting. This means that, since actual voting increases sharply up till age 35 then more slowly thereafter, answers to the question will tend to remove too many 35-50 year olds. Taking the word \”always\” literally, the answer can only move from \”I always vote\” to \”I don’t always vote\” the probability of answering yes would never increase. I’d guess that it increases but much too slowly.
I’d guess that following elections closely will have even more to do with being retired than actual voting.
Some Gallup questions very clearly select on age. One is have you voted at your current polling place (another is do you know where it is?). This is partly about past voting and partly about \”have you moved since the last election\”
Now I guess that LV filters would not be fixed if it turned out that \”likely voters\” were significantly more likely to be over 65 than actual voters. I guess that LV filters are mainly chosen and changed based on the performance of past polls with the LV filter in predicting the result (that is the bottom line).
This means a major change in the probability of voting as a function of age and change in the probability of supporting the Democrats as a funciton of age will both make the past data on the reliability of LV polls invalid.
We know that support for Republicans currently increases steeply with age all the way up. It wasn’t always this way. It used to be about level age 40 and up with an actual hump of self identified Republicans in generation X. The huge shift in the association of age and partisan support is part of the explaination of the huge increase in the LV-RV gap and also a reason that it might be a failure of LV filters.
3.2 A third reason it might be different this time, is just that current gaps are unusual, and maybe when the gap is unusual likely voter polls do badly. This can happen if say paying attention to the election is moderately associated with voters but huge swings in partisan differences in paying attention are not correlated with the predicted very large swings in voting. I’d guess there is more of a bandwagon effect in paying attention. I sure wouldn’t blame Democrats who don’t want to hear about the elections this year provided they vote.