hey benji, real question: I know there's nothing to suggest declining response rate has affected the accuracy of national polling, but what about along socioeconomic demos? Has primary polling ever been "accurate", even in the aggregate?
Most polling in general is only as accurate as what the campaigns use it for (trends tracking), then at the end the polling accuracy shoots up because you're polling the circumstances the actual election will take place in. The primary process being split among multiple states that used to vote with large gaps between them with outsized power given to Iowa and New Hampshire meant that national primary polling wasn't polling anything that would actually happen. This cycle's Super Tuesday is slightly different because states like California have moved into it so more people will be voting in the first month than usual.
State polling has generally been okay but the trend lines often matter more, the infamous example (for many reasons) is 2004, where the results were a "surprise" though Dean had been collapsing and Kerry/Edwards ascending for weeks beforehand. Once the early states voted campaigns that had maintained national clout (~10% or more at times) like Lieberman and Clark and Gephardt and Dean became nearly instantly irrelevant to the process. 2012 never even tested Santorum against the nation properly once his polling became a threat to Romney because his campaign ended for other reasons, and like Gingrich's campaign they weren't active in any other states after the early ones, sometimes not even getting on the ballot like Virginia. 2008 and 2016 saw caucus states polling poorly, but generally the trends did eventually move towards the national polling because people back winners.
The way they sample and then adjust the results is supposed to correct for any problems in the demographics. But there's always the chicken and egg problem, are we seeing more of [Y] because they're going to turn out more or is it due to the polling process and we should weight them down? I don't really know what the polling companies are getting for socioeconomic data versus what it should be as the census is almost ten years old now, in academic research if the sample is off in that you can just do more rounds of it since you aren't looking at a tight timeline to publish. Also the sample sizes in general used to be larger but now the polling companies are able to afford much larger samples. Like the Politico one supposedly does 5000 a day which is still a huge sample for other polls:
Our Democratic primary results are reported using 16,274 interviews with registered voters who indicated they may vote in the Democratic primary or caucus in their state. For those who say don’t know or no opinion, they are asked to pick a candidate they are leaning toward. Results are reported among first choice and those who lean toward a candidate. The interviews were collected from September 23-29, 2019 and have a margin of error of +/- 1%. The “Early Primary State Voters” demographic consists of 734 voters in Iowa, New Hampshire, Nevada, and South Carolina, and has a margin of error of +/- 4%.
The Economist/YouGov poll that has 28 Warren 22 Biden 13 dead guy is 1500
adults, much much less (although not that much less in terms of margin of error) than Politico/Morning Consult claims to poll a day. Adults is less accurate than registered voters which is less accurate than likely voters, but at this point in the process I'm not sure that really matters too much as long as it's consistent. Especially if the race effectively ends before those likely voters can vote for the other candidates. General Elections usually don't have that situation occur.
