The main reason FiveThirtyEight's 2016 model gave Trump a better chance than other models was its assumption that polling errors are correlated.

This is a claim about the design of FiveThirtyEight's own 2016 model and why its Trump probability differed from other forecasters' (e.g., NYT Upshot, PEC, HuffPost, PredictWise). True: published model documentation and/or independent comparative analyses show that FiveThirtyEight assumed correlated state polling errors more strongly than the other models, and that this is the largest single source of the gap in Trump's probability. Mostly true: correlated error is shown to be a major driver, perhaps tied with or barely ahead of other factors (higher overall uncertainty, undecided/third-party handling, trend adjustment), so 'single most important' is only slightly overstated. Mixed: analyses find correlated errors were one of several roughly equal drivers, with no clear primacy. Mostly false: other models also assumed substantial correlation, or other factors (e.g., larger national uncertainty, treatment of undecideds) clearly account for most of the gap, with correlation playing only a minor role. False: the evidence shows FiveThirtyEight did not assume more correlated errors than the others, or that correlation played no meaningful part in the difference.

Set before the research began.

1. E1 Unskewing the Election rjlipton.com (Richard J. Lipton blog) · 8 November 2016

   “The main difference cited by Silver is the higher number of voters telling pollsters they are undecided or supporting third-party candidates compared to 2012.”

2. E2 After 2016, Can We Ever Trust the Polls Again? The New Republic · 14 December 2016

   “FiveThirtyEight’s Nate Silver, who successfully predicted the 2008 and 2012 presidential races, was among the few pollsters to at least account for correlated error, giving Trump a 28 percent chance of winning.”

3. E3 The Polls and the US Presidential Election in 2020 ….and 2024 MIT Sloan (working paper by A. Barnett) · 30 December 2025

   “praising its awareness that outcomes in Pennsylvania, Michigan, and Wisconsin—the three key states Clinton was expected to carry but which went to Trump—were positively correlated.”

4. E4 Chapter 20: When Models Fail: 2016, 2020, and Beyond datafield.dev (online textbook chapter) · 1 January 2026

   “The difference between 5% and 29% was almost entirely explained by different assumptions about correlated versus independent state errors.”

E2 reports that FiveThirtyEight was among the few forecasters to account for correlated error and gave Trump a 28 percent chance; E3 praises an awareness that outcomes in Pennsylvania, Michigan and Wisconsin were positively correlated. E4 says the difference between a 5% and a 29% forecast was almost entirely explained by different assumptions about correlated versus independent state errors. E1 reports that Silver cited a higher share of undecided and third-party voters compared with 2012 as the main difference, which suggests other factors may also have shaped the model's output. Correlated error appears to be a major driver, but no rigorous comparison in the evidence shows that it was the largest single source of the gap.