Thanks! I agree that the approach you describe is optimal under very reasonable assumptions, but I think in practice few people used it (the median ratio between someone’s top choice and their second choice was 2, the mean if you throw out one outlier was ~20; only 7 people voted for at least 2 candidates and had ratios between their top two that were at least 20). Moreover, we had some[1] voters who didn’t vote the way you describe, but who did assign a fairly big number of projects similar small point values — I think kind of throwing in some points for charities they don’t favor that much, and I didn’t want to overweight their votes in the way I tallied the RCV-translated (or approval-translated) scores.
Still, I agree that my translations are bad — I should at least represent scores from people who basically approximated RCV in the current voting method the way they would be counted in RCV. I might try this (and think about what translation actually makes sense — just the top 10 charities people voted for?) later, but might not prioritize doing it.
For approval voting, you could also just look at the number of voters who gave a charity any (positive) number of points; these counts are included in this post and wouldn’t have changed the top 3.
Quickly estimating: there were 90 voters who voted for at least 4 candidates whose last two votes differed by a ratio of less than 1.5. There were 27 if instead of requiring at least 4 candidates, you require that the smallest point assignment is <5.
Or looking at it another way: across all ratios (across all voters) between what a given voter gave the candidate they ranked N and the candidate they ranked N+1, if we remove only the top 1 percentile of ratios (removing because a few people did use an approximation of RCV—equivalent in this case to removing ratios higher than 100:1), the mean is 2. Across all ~12K ratios, about 500 are exactly 1.
Thanks! I agree that the approach you describe is optimal under very reasonable assumptions, but I think in practice few people used it (the median ratio between someone’s top choice and their second choice was 2, the mean if you throw out one outlier was ~20; only 7 people voted for at least 2 candidates and had ratios between their top two that were at least 20). Moreover, we had some[1] voters who didn’t vote the way you describe, but who did assign a fairly big number of projects similar small point values — I think kind of throwing in some points for charities they don’t favor that much, and I didn’t want to overweight their votes in the way I tallied the RCV-translated (or approval-translated) scores.
Still, I agree that my translations are bad — I should at least represent scores from people who basically approximated RCV in the current voting method the way they would be counted in RCV. I might try this (and think about what translation actually makes sense — just the top 10 charities people voted for?) later, but might not prioritize doing it.
For approval voting, you could also just look at the number of voters who gave a charity any (positive) number of points; these counts are included in this post and wouldn’t have changed the top 3.
Quickly estimating: there were 90 voters who voted for at least 4 candidates whose last two votes differed by a ratio of less than 1.5. There were 27 if instead of requiring at least 4 candidates, you require that the smallest point assignment is <5.
Or looking at it another way: across all ratios (across all voters) between what a given voter gave the candidate they ranked N and the candidate they ranked N+1, if we remove only the top 1 percentile of ratios (removing because a few people did use an approximation of RCV—equivalent in this case to removing ratios higher than 100:1), the mean is 2. Across all ~12K ratios, about 500 are exactly 1.