Read on Twitter
Quick thread: There is some discussion over whether Trump will benefit from the increasing rate of likely voter polls as we get closer to November. We can use our polling data to provide a helpful and IMO pretty robust answer to this question. 1/?
In the past, likely voters have tended to be a bit more Republican-leaning than the registered pop. But there are problems with assuming the LV-RV gap will be constant across elecs. For 1, RVs and LVs change demographically. Covid also introduces noise 4 history-based LV filters.
This is a tricky problem to answer. It's not typically enough to just compare LV and RV polls; when you do so, you also conflate some effects of mode and pollster firm. What if all the LV polls are conducted with automated dialers? What if all RV are online? Etc etc...
So the best clue we have at the LV-RV difference is when one pollster releases numbers for both filters of the same poll. But here, there is mixed evidence; polls from ABC/WaPo have shown large, 5-point fx (toward Trump) of switching the filter. So does Monmouth. But not NYT!
We can add to this conversation using the Economist/YouGov polls since May. We use a statistical method called MRP to parse out the relationships between demographics and behavior and then make projections for the nation, and each state, based on census data.
So we grab about 22k interviews and run a set of models to predict each person's likelihood of...
(a) Registering to vote
(b) Voting
(c) Supporting Biden
The DV for (b) is self-reported turnout, but we add a correction to the predicted probabilities based on bias in 2016.
(a) Registering to vote
(b) Voting
(c) Supporting Biden
The DV for (b) is self-reported turnout, but we add a correction to the predicted probabilities based on bias in 2016.
Then we match up the models with updated census data on how many black and white, rich and poor, educated and uneducated etc. people live in The US. Then we have predictions for how many registered and likely voters there are, where they live and how they'll vote.
Running the analysis in this way lets us extract the demographic and political differences of likely versus registered voters (and non-voters!). In effect, we create three sets of different "polls" by adjusting the Biden-Trump numbers based on who we think will turn out.
So, for all eligible voters, if the election were held today we'd predict....
For all adults: Biden 47, Trump 35
For registered voters: Biden 50, Trump 42
For likely voters: Biden 49, Trump 42
That equals a small 1-point shift toward Trump from going RV -> LV. Some caveats...
For all adults: Biden 47, Trump 35
For registered voters: Biden 50, Trump 42
For likely voters: Biden 49, Trump 42
That equals a small 1-point shift toward Trump from going RV -> LV. Some caveats...
(1) We know people who answer online polls are more engaged than they "should" be. So our LV-RV gap could be artificially deflated by RVs who are likelier to be LVs than a live-caller poll would show.
(2) Our adjustments for self-report bias might not be robust across elecs.
(2) Our adjustments for self-report bias might not be robust across elecs.
After acknowledging of the caveats, I still think this analysis is useful for pointing out that the presidential race will probably tighten just a bit when we get more likely voter polls (to say nothing of separate trends in the contest that could push it toward 50/50).
/end
/end
cc @Nate_Cohn i know you had some tweets about this last week iirc
