Most of us have had to vote at some stage: in primary/high school, perhaps to elect class/school captains, during university to choose the best presentation in your physics class, and obviously if you’re old enough, voting in elections for state/country leaders.
In Australia, we use preferential voting to elect officials, and for the lower house (House of Representatives), counting is done using a full preferential system, also known as instant runoff voting (IRV).
That’s all well and good, and you can read up more about preferential voting at the linked Wikipedia page, but why am I discussing voting? I was recently involved in a judging session where nine judges (myself included) ranked five possible candidates, in preferential order.
When it came time to count the votes, there was not a clear winner (actually, there was after looking at the data afterwards, but it didn’t make a difference!), so the question became how to decide on the winner? The chosen method: sum the preferences for each candidate, with the candidate that had the smallest number of total preferences being the winner.
As a statistician, this irked me. We had data in the form of ranks, but were treating them as continuous data. Just because two of the judges placed Candidate A above Candidate B, doesn’t mean the difference between the two candidates was judged the same, rather that both judges preferred Candidate A to Candidate B.
Summing, or taking the mean of the preferences, is known as the Borda count, and by the description of how it’s done (mean/sum), you can see that it is an average method: the candidate that wins the vote is often the consensus view, not that which is preferred by the majority.
So what other methods could we use? The preferential method as described above? Or the Condorcet method, which is a method that chooses the candidate that would win by majority against pairings with all the other candidates. The Condorcet method is a majoritan method. Unfortunately it doesn’t always work, because you can get cycles. A great analogy is on the previously linked Wikipedia page: there may be a situation where each candidate wins against another, like in a rock-paper-scissors game. So you need some way of breaking ties.
So how do the methods compare for the session I was judging? Let’s have a look! Firstly, I found some code to calculate both the Borda and Condorcet winner here. And here’s the data:
votes <- structure(list(Candidate = c("A", "B", "C", "D", "E"),
A = c(1L, 3L, 2L, 4L, 5L), B = c(3L, 4L, 2L, 1L, 5L),
C = c(2L, 4L, 5L, 1L, 3L), D = c(2L, 4L, 1L, 3L, 5L),
E = c(3L, 4L, 5L, 2L, 1L), F = c(1L, 3L, 2L, 4L, 5L),
G = c(1L, 3L, 5L, 4L, 2L), H = c(3L, 2L, 5L, 4L, 1L),
I = c(4L, 1L, 5L, 2L, 3L)),
.Names = c("Candidate", paste("J", 1:9, sep = "")),
class = "data.frame", row.names = c(NA, -5L))Firstly, what’s the Borda count winner?
## Names Average Rank Position
## 1 A 2.222222 1
## 4 D 2.777778 2
## 2 B 3.111111 3
## 5 E 3.333333 4
## 3 C 3.555556 5
Candidate A is the winner by the Borda count, follwed by Candidate D. What if we do the full preferential count? There were 9 judges, so after distributing preferences, a candidate needs 5 votes to be declared a winner. I did this manually, and came up with again, Candidate A, followed by Candidate D.
Finally, how about the Condorcet method? Recall this method counts pairwise winners:
## Name Rank Method
## 1 A 1 Condorcet
## 2 B 2 Condorcet
## 3 D 3 Condorcet
## 4 E 4 Condorcet
## 5 C 5 Condorcet
So here again, Candidate A is the winner, but now the runner-up is Candidate B: that is, in all pairwise elections, Candidate B came out the winner more often than Candidate D.
As it turned out, looking at the data, Candidate A had the most first preferences anyway with three votes, so to keep it simple, we should have picked them anyway! But what’s interesting, is that while Candidates D and E each had 2 first preferences, they were ranked third and fourth respectively by the Condorcet method: most people preferred other candidates.
What if the preferences had been slightly different? Let’s look at swapping two preferences for a single judge: Judge 8 will now vote Candidate D first, and Candidate E fourth:
votes2 <- votes
votes2$J8[4] <- 1
votes2$J8[5] <- 4The Borda method still places Candidate A first, with Candidates E and C swapping places:
## Names Average Rank Position
## 1 A 2.222222 1
## 4 D 2.444444 2
## 2 B 3.111111 3
## 3 C 3.555556 4
## 5 E 3.666667 5
Most importantly though, the Concordet method now chooses Candidate D as the winner, with Candidate A coming second!
## Name Rank Method
## 1 D 1 Condorcet
## 2 A 2 Condorcet
## 3 B 3 Condorcet
## 4 E 4 Condorcet
## 5 C 5 Condorcet
So whilst Candidates A and D have the same number of first preferences, more people preference Candidate D over all other candidates, than they do Candidate A.
Majority rules!