Jeff Lax points to this post from Nate Silver and asks for my thoughts.

In his post, Nate talks about data quality issues of national and state polls. It’s a good discussion, but the one thing he unfortunately *doesn’t* talk about is multilevel regression and poststratification (or see here for more). What you want to do is fit a multilevel regression to your raw data so as to estimate your outcome of interest (for example, support for Hillary Clinton) in demographic/geographic slices of the population, characterized by age, sex, ethnicity, education, state of residence, maybe some other variables, maybe party identification as well. Then you poststratify using some combination of census and poll data that give you the number of people in each category within each state.

That’s the way to go.

See here for further discussion, particularly on the subject of state-level opinion.

And, hey! We wrote this paper in 2005 on state-level opinion from national surveys.

And here’s the original paper, “Poststratification into many categories using hierarchical logistic regression.” It was published in 1997 in the journal Survey Methodology.

It takes a long long time for research to make its way from academia to journalism. Slowly but surely, though, it will happen.

Here is an interesting application of Andrew’s “It takes a long long time for research to make its way from academia to journalism. Slowly but surely, though, it will happen.” The article is entitled, “Here’s How HuffPost Averages The Polls And Figures Out Who’s Ahead

It’s actually a bit more technical than an ‘average.'” Actually, it is a great deal more technical. Note that Kalman died in July, 2016 and the eponymous filter dates from the 1960s.

http://www.huffingtonpost.com/entry/huffpost-pollster-poll-averages-methodology_us_57d1a3b2e4b06a74c9f361cb?section=&

You may be happy to know we use MRP at A Place for Mom, a senior housing and care referral service, to estimate senior living costs all the way down to the municipal level and with surprising success. It’s not Bayesian and we aren’t journalists, but it is shared with the public, and I used the method because you convincingly evangelized it.