Dr. Shiva is a smart guy, but I disagree with his analysis. As best I can tell, the way he set up his data, dipping below the line does NOT mean fewer Trump votes, since "straight party" Republican votes are all Trump votes too. If you chart (Trump only votes as a percent of total any-candidate-only) as the Y and (straight Repub as a percent of total straight-party) as the X, you get a nice linear fit that goes from bottom left to top right as expected - i.e. the more Republican-straight-ticket as a fraction of all straight ticket votes in the precinct, the higher the fraction of Trump only votes out of all some-candidate-only votes. I did find that it leaned toward straight ticket votes, i.e. the slope of my curve fit was like 0.85 instead of 1. However it seemed totally legit.
In addition, it is my opinion that he has it a little bit backwards here. WAYNE COUNTY IS HORRIBLY CORRUPT. We should expect that Detroit poll workers frequently threw Trump ballots in the trash. That's why Wayne County looks different. It barely has any recognizable Republican presence because they are a den of corruption and iniquity.
With all the above said, I agree 100% with Dr. Shiva's comments about how primitive our elections are, and how it would be an amazing and totally feasible improvement to publish the ballot image scans for public consumption. In one fell swoop that would eliminate the need for manual recounts - anybody could recount the ballots who wanted to.
Dr. Shiva's analysis might have a problem along the lines you describe, but I think you misunderstood his Y axis, or maybe he miscommunicated it.
One piece of information you're missing is:
Straight Republican votes are NOT also all Trump votes. You can mark the straight party bubble but make exceptions for individual races, and the exceptions take precedence.
Dr. Shiva isn't entirely clear here, but I think he actually constructs the Y axis with the expression:
rather than:
trump_only_vote_percent - straight_republican_bubble_filled_vote_percent
That would explain why Shiva expects these values to center in a horizontal line around 0, give or take, depending on whether Trump is more or less popular than the party as a whole.
From another perspective, I suspect Y actually represents:
EDITED, because I failed to grasp something that you did the first time around:
That doesn't really matter though. Either way the substance of your criticism is probably more right than wrong: On the left side of the graph you have people who hate the GOP but might like Trump. On the right side of the graph you have people who love the GOP but might hate Trump. In a precinct that voted 0% GOP except maybe the President, the worst he can do is break even with the GOP. In a precinct that voting 100% straight GOP except maybe the President, the best he can possibly do is break even with the GOP. After all, a GOP voter who likes Trump "even more" than the GOP can't just vote for Trump twice to show it...but a never-Trumper Republican can say "all GOP except Trump." This pretty much requires a trend line that goes down to the right.
MUCH LATER EDIT:
We could both be wrong though. There are three possibilities for the Y axis:
a. total_trump_vote_percentage - republican_straight_party_bubble_vote_percentage
b. trump_vote_percentage_on_mixed_ballots - republican_straight_party_bubble_vote_percentage
c. trump_vote_percentage_on_mixed_ballots - trump_vote_percentage_on_republican_straight_party_bubble_ballots
Dr. Shiva's analysis doesn't make sense for 'a' or 'b' (where we'd actually expect a down-right line, just of a shallower slope), but it may make sense for 'c'.
Note: It's no longer clear to me that "straight Republican votes are NOT also all Trump votes." At least in accordance with this data set, they very well may be, despite the articles I read about being able to make exceptions.
Dr. Shiva is a smart guy, but I disagree with his analysis. As best I can tell, the way he set up his data, dipping below the line does NOT mean fewer Trump votes, since "straight party" Republican votes are all Trump votes too. If you chart (Trump only votes as a percent of total any-candidate-only) as the Y and (straight Repub as a percent of total straight-party) as the X, you get a nice linear fit that goes from bottom left to top right as expected - i.e. the more Republican-straight-ticket as a fraction of all straight ticket votes in the precinct, the higher the fraction of Trump only votes out of all some-candidate-only votes. I did find that it leaned toward straight ticket votes, i.e. the slope of my curve fit was like 0.85 instead of 1. However it seemed totally legit.
In addition, it is my opinion that he has it a little bit backwards here. WAYNE COUNTY IS HORRIBLY CORRUPT. We should expect that Detroit poll workers frequently threw Trump ballots in the trash. That's why Wayne County looks different. It barely has any recognizable Republican presence because they are a den of corruption and iniquity.
With all the above said, I agree 100% with Dr. Shiva's comments about how primitive our elections are, and how it would be an amazing and totally feasible improvement to publish the ballot image scans for public consumption. In one fell swoop that would eliminate the need for manual recounts - anybody could recount the ballots who wanted to.
Dr. Shiva's analysis might have a problem along the lines you describe, but I think you misunderstood his Y axis, or maybe he miscommunicated it.
One piece of information you're missing is: Straight Republican votes are NOT also all Trump votes. You can mark the straight party bubble but make exceptions for individual races, and the exceptions take precedence.
Dr. Shiva isn't entirely clear here, but I think he actually constructs the Y axis with the expression:
total_trump_vote_percent - straight_republican_bubble_filled_vote_percent
rather than: trump_only_vote_percent - straight_republican_bubble_filled_vote_percent
That would explain why Shiva expects these values to center in a horizontal line around 0, give or take, depending on whether Trump is more or less popular than the party as a whole.
From another perspective, I suspect Y actually represents:
trump_only_never_rino_vote_percent - never_trumper_rino_vote_percent
EDITED, because I failed to grasp something that you did the first time around: That doesn't really matter though. Either way the substance of your criticism is probably more right than wrong: On the left side of the graph you have people who hate the GOP but might like Trump. On the right side of the graph you have people who love the GOP but might hate Trump. In a precinct that voted 0% GOP except maybe the President, the worst he can do is break even with the GOP. In a precinct that voting 100% straight GOP except maybe the President, the best he can possibly do is break even with the GOP. After all, a GOP voter who likes Trump "even more" than the GOP can't just vote for Trump twice to show it...but a never-Trumper Republican can say "all GOP except Trump." This pretty much requires a trend line that goes down to the right.
MUCH LATER EDIT: We could both be wrong though. There are three possibilities for the Y axis:
a. total_trump_vote_percentage - republican_straight_party_bubble_vote_percentage
b. trump_vote_percentage_on_mixed_ballots - republican_straight_party_bubble_vote_percentage
c. trump_vote_percentage_on_mixed_ballots - trump_vote_percentage_on_republican_straight_party_bubble_ballots
Dr. Shiva's analysis doesn't make sense for 'a' or 'b' (where we'd actually expect a down-right line, just of a shallower slope), but it may make sense for 'c'.
Note: It's no longer clear to me that "straight Republican votes are NOT also all Trump votes." At least in accordance with this data set, they very well may be, despite the articles I read about being able to make exceptions.