More on election fraud claims in Pennsylvania
After looking at Walter Mebane’s paper about potential election fraud in Pennsylvania and being unable to figure out exactly what he was arguing, I asked Andrew Gelman to take a look at it. It turned out somebody else had asked him to do so earlier, and he blogged about this morning:
Pointing to this new article by Walter Mebane, eforensics Analysis of the 2024 President Election in Pennsylvania, Stefan Gramatovici writes:
I think you might find it interesting. I think their claims are wrong, but this is not my area of expertise, and I am having difficulty fully following the paper due to the lack of details in the models they are fitting.
They specify some precinct level model for turnout and vote share, fit the data and find some sort of pattern in the residuals. Then they add some variables for fraud, and that “The statewide total across precincts of eforensics-fraudulent votes, 225440.2 [207757.1, 252978.1], exceeds the statewide gap of 120266 votes between Trump and Harris.” So now we even have rather tight “fraud vote” CIs.
The key issue I have with the paper is that these “fraudulent” votes not explained by the original precinct model are probably due to other factors. On page 8, the author says: ” there is a signal that likely the incremental stolen votes at least in part come from malevolent distortions in Philadelphia and Huntingdon, but generally–including in these two counties–the incremental stolen votes are unknown admixtures of malevolent distortions and electors’ strategic behaviors.”. I am not a political scientist, but I could name several likely “strategic behaviors” around the 2024 election, especially in the Philadelphia area. I see no method suggested by the author to distinguish between patterns due to model misspecification and patterns due to fraud.
I took a look at the linked paper and I can’t understand what’s going on here at all. I also looked at the earlier paper by Mebane et al. describing their “eforensics” method, and I still can’t figure out what they’re trying to do. Mebane is a respected quantitative political scientist, which can be taken as evidence that we should try to read these papers more seriously; conversely, without his name attached to them I don’t think we’d be reading them at all.
The bottom line here for me at least is that I think claims about serious fraud in the US electoral system face a very high evidentiary barrier for good reasons, and a statistical analysis that is too confusing for someone like Gelman to understand is failing to meet that standard basically by definition.