Thread by @shylockh

Voter Birthday Explainer Thread
revolver.news/2020/12/pennsy…

This came out in Revolver recently. It’s a new twist on identifying voter fr**d: Instead of starting with weird vote patterns, find *other data* that look weird (here, voter birthdays), and then relate it to votes.
(1/N)

It’s surprisingly hard to generate fake birthdays without leaving some trace in the data. The piece considers two broad ways that pull in opposite directions. First, you’ll probably pick too many round numbers – 1st, 15th & 31st of the month, Jan and Dec etc.
(2/N)

So, you think, I’ll be clever. I’ll use a uniform distribution over months of the year. Bzzt! Months have different numbers of days. Okay, hmm. I’ll choose uniformly over days of the year. Bzzt! Wrong again. It turns out that actual birth data aren’t uniform here either.
(3/N)

To get truly convincing data, you’d actually have to go through administrative datasets and find out the distribution of months of the year among births in your county, then find a different dataset to also match the distribution of days of the month.
(4/N)

Who’s got time for that? Apparently, the author of the piece! This is why fake data is hard – you never know who might think up some odd way to test it down the line. The piece does a lot of ways of aggregating weird measures of birthdays.
(5/N)

And when you do that, there’s a bunch of counties in PA that look highly suspicious.
“(Northumberland, Delaware, Montgomery, Lawrence, Dauphin, LeHigh, and Luzerne) have numbers of suspicious birthdays above the 99.5th percentile of plausible distributions”
(6/N)

On its own, this is an interesting list. The biggest outlier votes majority Republican! But also very high on the list is our old friend Montgomery PA, home of the most suspicious vote update in America
revolver.news/2020/11/explos…
(7/N)

The big question is, are these benchmarks right? The piece considers a number of them, and they’re pretty compelling to me, but it’s tough to say. I think it’s hard to think of obviously better ones you could get from actually available data, but it’s still an assumption.
(8/N)

Importantly, the fact that these counties look weird *doesn’t actually depend on their votes*. It just comes from birthdays, regardless of how they vote. But you can then use this to find out the relationship of suspicious birthdays to votes.
(9/N)

And second, even if you don’t think the benchmarks aren’t totally correct, this implies that the *levels* might be wrong. But this wouldn’t obviously explain why the *variation* that does exist should nonetheless predict vote outcomes.
(10/N)

And when you do that, Northumberland notwithstanding, having more suspicious birthdays is strongly related to more votes for Biden, with a p-value of 0.000008 (the probability of observing a relation this strong by chance).

(11/N)

This effect is also big! A one standard deviation increase in suspicious birthdays is associated with higher Biden vote share by 6.8 percentage points.
(12/N)

“Out of the 13 counties in Pennsylvania who voted majority Biden, 9 are above the 95th percentile of suspicious birthdays.”
(13/N)

Mores suspicious birthdays also positively predict more votes for Biden relative to Democrat performance in all elections since 2000, with a p-value of 0.003. So it’s not just that these counties always vote Democrat, but they do so at unusually high levels this election.
(14/N)

They’re also positively associated with more Jorgensen votes relative to Trump, a secondary indicator of suspicious vote patterns that has been looked at in a few places, such as the Montgomery analysis.
(15/N)

Finally, the piece tries to quantify how important this stuff might be. It does this by relating the magnitude of suspicious birthdays to Biden vote share, and asks what would happen if the birthday distribution were a little less suspicious
(16/N)

This is the most speculative part, and is probably best thought of as a back of the envelope estimate. But the numbers seem to be big enough to potentially swing the state to Trump. We might have guessed this from the large effect sizes, but it’s interesting to quantify.
(17/N)

I think that the strongest case for voter fr**d comes from the piling up of coincidences, across lots of datasets, and lots of ways of measuring suspicious outcomes. Individually, any one has weaknesses.
(18/N)

But when they keep identifying the same odd places in different ways, that adds to an increasing cause for concern. Pennsylvania vote outcomes are looking pretty damn strange, and Montgomery County is looking the strangest of the strange.
(/fin)