Private vs Public Polls

During the 2024 presidential election campaign, Mark Halperin regularly shared with his audience what he learned from the campaigns’ private polls. Private polls are better funded and far more accurate than public polls. By listening to Halperin, I knew for six months ahead of election day that Trump was ahead in the key battleground states and yet in the news media, there was almost no coverage of private polls that showed the battleground sunbelt states were out of reach for the Democrats in the presidential campaign and that Trump was highly likely to win the presidency.
Andrew Gelman does touch on private polls, though he does not make them a central focus. The most relevant passage appears in his 2021 paper “Failure and Success in Political Polling and Election Forecasting,” where he notes that well-funded campaigns and advocacy groups can do more effective survey adjustment using the voter file, which contains information including past turnout history on nearly 200 million Americans. That is his acknowledgment that a gap exists between what campaigns know and what public pollsters produce, though he states it briefly and moves on.
The broader literature around Gelman’s work is more explicit. G. Elliott Morris, his collaborator on the Economist model, notes that campaign pollsters and private pollsters have been doing mixed-mode surveys for some time, because their reputation relies on being accurate and they lose clients if they are not. That is the core of what I observed with Halperin: private pollsters face a direct accountability test that public pollsters do not. A campaign that misreads the battleground states loses money and influence. A public pollster who gets it wrong faces a news cycle of criticism and then moves on.
Gelman’s deeper concern is the structural failure of public polling. He points out that with response rates in the 10 percent range, the select group who happen to respond to surveys are nothing like a random sample of the population of adult Americans or even of likely voters. His diagnosis of the 2016 and 2020 errors centers on differential nonresponse rather than a “shy Trump voter” effect, arguing that differential nonresponse and differential turnout are more plausible explanations of polling error than the hypothesis that Trump voters systematically concealed their preferences.
What Gelman does not do is examine why media organizations systematically amplify the public polls that showed closer races while largely ignoring or marginalizing private poll signals that showed Trump ahead by comfortable margins in the Sun Belt. That is a question about institutional incentives and coalition maintenance, not statistical methodology, and it sits outside what Gelman studies. Public polls that show tight races generate more coverage, more engagement, more fundraising for both sides, and more relevance for the forecasting industry. A race that is functionally over by September is bad for business across the entire election media complex. The private polls existed. The information was available to people like Halperin. The mainstream press largely chose not to make it central to their coverage, and Gelman’s work, focused as it is on improving the statistical models, does not explain why.
I asked Gelman for a response on the above and he said: “I would only say that ‘private polls that showed the battleground sunbelt states were out of reach for the Democrats in the presidential campaign’ is too strong. In retrospect, sure. But ahead of time, maybe not. Private polls can have systematic errors too.”