The video definitely has some shortcomings, but I don't think your assumptions about what the Y-values should look like is correct. I just paste in from a tweet I saw:
Simple model: Dems vote straight D, Reps vote straight R, and Independents vote a mixed ticket. x-axis measures number of R compared to D. y=i-x, where i is % of Indep who vote for Trump. i should be approx indep of x, so slope should be about -1. Actual slope is approx -0.6.
https://twitter.com/Clustify/status/1326553412356382723
If the people that are not straight ticket are true independents who choose their candidate without caring who their neighbors are voting for, their vote should have no dependence on X, so the "- X" part of Y causes a slope of -1.
I really hope they have something better than that. Dr. Shiva's analysis is badly misguided. For example (this is only a small part of the available criticism):
I don't have time to look at the data, but I'll say this: Be careful about using Benford's Law -- it only applies when the numbers span several orders of magnitude, which may not be the case for election data. Here is a good discussion:
Even better, the website should detect a dupe and show the poster all of the previous copies so they can decide to go upvote one of them instead of posting it yet again.
Someone who could run a Benford Law analysis of all our completed research.
Be careful about applying Benford's Law -- it only applies when the numbers span several orders of magnitude, which may not be the case for election data.
I'm on T-mobile and I've had trouble sending links to Verizon customers at times while not having trouble sending the same links to someone else (not sure what network she is on). The links do go through if you try again a little later. I'm guessing (could be wrong) that Verizon is checking to see if the URL goes to a page that spreads a virus and won't let it go through until it confirmed to be OK. They are annoying slow about it, though.
Go big or go home! ;-)
Yeah, someone needs to make a video of them saying that spliced together with Trump supporters being beat up. Sadly, I don't have the skills...
The Clinton Foundation's bank account disagrees.
It gets better...
Since his interpretation of the Wayne County graph was off, I'm not inclined to assume his interpretation of any other data is correct without seeing it for myself.
Anyway, even if some counties do show a legitimate zero slope, there is now a strong argument that negative slope does not imply fraud:
It's only a graph of -x if you assume the percent of Trump-only voters is independent of x.
Not sure if you are referring to the "ymin = -x" (which is just a mathematical identity that doesn't rely on any model) in the tweet or something farther down in the thread (sounds like the latter). Dr. Shiva is graphing Y=I-X, where I is the percentage of split-ticket voters that are voting for trump. He is assuming (modeling) that I-X should be indep of X (otherwise, there is fraud). A model that takes the opposite extreme would be that I itself is indep of X, which would give a slope of -1 instead of 0. That argument is made here:
https://twitter.com/Clustify/status/1326553412356382723
The actual data shows a slope of -0.6, which is totally reasonable (lies between the two extreme models) without any fraud.
The argument that negative slope does not imply fraud is shown to be even stronger here:
the overall total vote for the Yaxis
That part is incorrect. Check the video at 15:00 for definitions, and then at 19:35 where he says how they are defining the graph. He is computing the Y value by taking the percentage of split-tickets that voted for Trump and subtracting the X value. This post argues that it's not crazy to think the split-ticket percentage is independent of X, which would imply a slope of -1:
If you follow the twitter thread, the argument has actually gotten much stronger that the analysis is just wrong. Early part of the thread claims that any sensible model would give a negative slope (which Dr. Shiva incorrectly interprets as a sign that votes are stolen) instead of a horizontal line. Later analysis, based on no model at all, just the assumption that Rep % + Dem % = 100% (no significant amount of third party votes), shows that if you take the exact same set of data and graph it in terms of "Biden votes" instead of "Trump votes" you would get a result showing points clustered around a line with the same slope. A negative slope for Trump means Biden must also have a negative slope, so interpreting a negative slope as evidence that votes were stolen from that candidate makes no sense.
That is basically the argument that is made elsewhere in the thread: https://twitter.com/Clustify/status/1326553412356382723
If you are talking about the Wayne County graph, virtually all of the data has X much less than 25% -- there really isn't enough high X data to see what is going on. He didn't keep the scales the same for the different counties, which makes it easy to misinterpret.
I don't think that's right. Check the video at 15:00 for definitions, and then at 19:35 where he says how they are defining the graph.
You could have any precinct has 0-100% straight ticket have remaining R voters be 0-100% defectors. The issue is there should be no trend up or down at all.
Maybe I am misunderstanding your analysis, but are you losing track of what he is graphing on the vertical axis? Unless I'm mistaken, the vertical axis is the percentage of non-straight-ticket voters that voted for Trump minus the value from the X-axis. So, the farther to the right you go on the graph (bigger X), the smaller the Y value will tend to be.
To put it a little differently, the two most extreme (but somewhat reasonable) modeling assumptions you could make are:
- Assume I-X is indep of X
- Assume I is indep of X
The slope of the lines in his graphs imply that reality is somewhere in between those two extremes (somewhat closer to (2) than to (1)). He uses assumption (1) and concludes that all deviation from that must be fraud, and calculates 69,000 votes stolen from Trump based on that assumption. If you applied the same thinking with assumption (2), you would conclude that tens of thousands of votes were stolen from Biden. To make such a strong claim based on such an extreme and unjustified assumption is just crazy.
That's not really the point. The thing he is graphing is of this form: Y=I-X where I is the percentage of individual votes that go to Trump. He is assuming that I-X should be independent of X. To the degree that I is independent of X, rather than I-X being independent of X, you would expect a very strong downward slope (even if there is no cheating). The downward slope, which he attributes purely to fraud, it just a consequence of him choosing to graph I-X instead of I on the vertical axis.
But part of me feels like its too good to be true.
It is too good to be true. He's assuming the line should be horizontal when any reasonable model would have it sloping downward even if there was no cheating. Pushing this analysis is going to end up in embarrassment, unfortunately.
Image search is not showing the blue part anywhere.
https://duckduckgo.com/?q=dominion+voting&t=ffab&iax=images&ia=images