This is potentially a very good point but I think we need to see the totals of each kind of vote for each graph before implying the conclusions are wrong.

Shiva's Y axis is (% of total Trump votes among votes given to a specific candidate - % of total Republican votes among down-ballot votes). My understanding is he's plotting this value because in a pro-Trump precinct, you would expect to see people vote Trump when they vote for any individual candidate (i.e. anything not strictly down-ballot), and when they vote down-ballot you would expect to see them vote Republican.

In the video he uses roughly equal numbers between the two types of votes and his explanation makes sense. If I was shown 1000 people and was told roughly 60% of them support Trump but was not told which of the two types of votes this entire group did, I would still expect roughly 600 of them to be of whatever kind of vote meant Trump; the type of vote doesn't necessarily factor in.

If you use pretty different vote totals, though, your assessment starts to make more sense. To use a ridiculous example for the sake of expediency, let's say in a precinct 100,000 people voted down-ballot and only 10 people voted individual. Well if I'm then told this is a heavily Republican precinct I'm going to expect most of those 10 individual votes to be for Biden since statistically speaking there are a lot more Republicans than Democrats in this area; so there's more likely to be an anti-Trump Republican in the mix than a pro-Trump Democrat or Independent due to the sheer number of Republicans. Using example numbers in this case, if 80% of down-ballots were Republican but only 50% (so 5) of the individual votes were Trump you'd see a deficit of 30%!

Of course it's possible that if such a case existed it would be considered an outlier die to the large discrepancy in voting type numbers and would be excluded from the analysis as an outlier. Once the number of individual votes starts matching the number of down-ballots, Shiva's math looks more sensible. Using the same example above except this time the vote type counts are more equal (heavy Republican precinct with say 50,000 down-ballot and 50,000 individual), I'm not expecting much deviation in the percentages because in this precinct neither type of voting is significantly more popular so a Republican is equally likely to vote Trump individually as they are to vote down-ballot. Seeing a 30% differential here is much more eyebrow-raising because that implies a stark difference in voting habits based on an otherwise completely arbitrary attribute of the vote.

My hunch would be that in real life, the types of votes are closer to being of similar quantities in each precinct, which is maybe why Shiva and his team didn't even bother to address this point in their video - it's a rare enough occurrence that it isn't worth discussing. This is all based on my assumptions though, seeing vote totals of each type alongside the graphs would certainly be good for clarification.

This is potentially a very good point but I think we need to see the totals of each kind of vote for each graph before implying the conclusions are wrong.

Shiva's Y axis is (% of total Trump votes among votes given to a specific candidate - % of total Republican votes among down-ballot votes). My understanding is he's plotting this value because in a pro-Trump precinct, you would expect to see people vote Trump when they vote for any individual candidate (i.e. anything not strictly down-ballot), and when they vote down-ballot you would expect to see them vote Republican.

In the video he uses roughly equal numbers between the two types of votes and his explanation makes sense. If I was shown 1000 people and was told roughly 60% of them support Trump but was not told which of the two types of votes this entire group did, I would still expect roughly 600 of them to be of whatever kind of vote meant Trump; the type of vote doesn't necessarily factor in.

If you use pretty different vote totals, though, your assessment starts to make more sense. To use a ridiculous example for the sake of expediency, let's say in a precinct 100,000 people voted down-ballot and only 10 people voted individual. Well if I'm then told this is a heavily Repiblican precinct I'm going to expect most of those 10 individual votes to be for Biden since statistically speaking there are a lot more Republicans than Democrats in this area; so there's more likely to be an anti-Trump Republican in the mix than a pro-Trump Democrat or Independent due to the sheer number of Republicans. Using example numbers in this case, if 80% of down-ballots were Republican but only 50% (so 5) of the individual votes were Trump you'd see a deficit of 30%!

Of course it's possible that if such a case existed it would be considered an outlier die to the large discrepancy in voting type numbers and would be excluded from the analysis as an outlier. Once the number of individual votes starts matching the number of down-ballots, Shiva's math looks more sensible. Using the same example above except this time the vote type counts are more equal (heavy Republican precinct with say 50,000 down-ballot and 50,000 individual), I'm not expecting much deviation in the percentages because in this precinct neither type of voting is significantly more popular so a Republican is equally likely to vote Trump individually as they are to vote down-ballot. Seeing a 30% differential here is much more eyebrow-raising because that implies a stark difference in voting habits based on an otherwise completely arbitrary attribute of the vote.

My hunch would be that in real life, the types of votes are closer to being of similar quantities in each precinct, which is maybe why Shiva and his team didn't even bother to address this point in their video - it's a rare enough occurrence that it isn't worth discussing. This is all based on my assumptions though, seeing vote totals of each type alongside the graphs would certainly be good for clarification.