I show strong evidence that a simple computer glitch can cause these trends: switching some Trump votes for Biden.
@va_shiva @Peoples_Pundit @thevivafrei @APhilosophae

Please tag anyone relevant. Here's a simulation vs reality https://twitter.com/va_shiva/status/1326220987608621056

Above shows a simulation I did, where I swap trump and biden votes, and it produces the exact patterns we see in reality.

I estimate 1 in 6 votes were flipped from trump to biden.

I saw Shivas video (top of thread) on strange trends in downballot performance, so I pulled the data myself.

Plotted here is how red a precinct is vs. trumps down ballot performance. The more red a district is, the worse trump performs against other republicans down the ticket.

We would expect these lines to be relatively flat, as republicans shouldn't behave differently based on how red the district is.

Shivas suggested this maybe a computer algorithm that changes votes based on how republican a district is. This seems complicated, so I tried a something simpler: flipping some biden results for trump.

This is data I was able to simulate using this method:

They follow the exact same trend. Here's how the simulation works:

1) Take how red the district is
2) Add a random number that is trump downballot performance
3) Move some of the votes from Trump to Biden

I get a little more technical now. Here's where things get more interesting, the parameters I use to generate the graphs come directly from a 1 variable linear regression.

The fitted slope to the data has a very nice interpretation: It is the percentage of voted that flip.

Fitted coefficients show a slope of -.167, or around -1/6.
An intercept of .0384 implies a 3.8% overperformance by trump (if unaltered).

Diagnostics point to this model being a good fit to the data.

The fact that the results of this model fit can be used to simulate the data above is very telling, imo.

Interestingly, we have non-zero intercept terms, meaning that trump actually OVER PERFORMs vs republicans. I personally suspected this, as I think there are many TRUMP voters, who don't necessarily care of the republican party. Not many republican are not voting trump.

The code to generate this is here: https://colab.research.google.com/drive/1ox0Q5FPP2TxR8kfvUJhOVM60tgQbqXin?authuser=4#scrollTo=8ZdpOl03TkGB
and the data can be downloaded here: https://results.enr.clarityelections.com/MI/Oakland/105840/web.264614/#/summary

@We_Have_Risen

I have given this more thought. The interpretation that we expect the graphs to be flat is not correct. If there is natural cross party voting, the rates in either direction would determine the slope.

In the democrat only districts, we only see democrats that vote trump, and in the republican one we only see the republicans that don't vote trump.

Let's say you expect 2% of democrats vote trump, and 2% of republicans don't, you see a line going from .02 on the left to -.02 on the right.

You *could* explain this with cross over voting, but the numbers would have to be too extreme to be believable.

It's probably not as bad a 1 in 6 votes flipped though.

There's some really interesting dynamics here. I might do a more complicated model tomorrow.