Strategy-Free Elections, And The True Measure of a Voting System

If you get involved in the debate on voting system reform, it's never long before someone brings up the concept of "strategic voting." The idea is simple enough: we all want to be as honest as possible on our ballots, but because the voting system is imperfect, we vote otherwise. This is easy to see in plurality; lots of people who claim to really want a third-party candidate end up voting for one of the two major-party candidates on election day. But it happens, to a lesser or greater extent, under all voting methods.

Actually, that's a cleverly-constructed lie of omission.

It turns out that there are a number of voting systems which are 100% strategy-free, so your honest vote will always also be your best vote. But you're not going to like them. Here's an example: voting is performed like plurality, i.e., each voter picks a single candidate. The winner is whichever candidate is named on a ballot chosen at random. Clearly, you should always vote the one candidate you think is best for the job, because there's no reason to fear that you're "throwing your vote away" or making it easier for a candidate you dislike to win. The clever part in the constructing of the lie was omitting the word deterministic, which means "no random components". There are no strategy-free deterministic voting systems.

Constructing other (and better) strategy-free methods is easy enough. So if strategy is so vitally important that it invariably comes up in every voting-system discussion, why don't we use one of them? The answer is: average performance. Random ballot voting, as your intuition probably tells you, is an absolutely terrible system. But intuition is sometimes wrong, so it's important that we can back it up with data by running computer simulations to calculate Bayesian regret. And the data shows that, based on the number of candidates competing, random ballot is two- to four-times worse than plurality voting; which we all know from experience to be a pretty bad system.

Strategy, and a voting system's susceptibility to it, are an important thing to be aware of. But immunity to strategy, even though it sounds like a great thing to strive for, isn't the goal of a voting system; if it were, we'd have an easy answer to the problem, in the form of non-deterministic voting methods. And there are a host of other reasonable-sounding things for a voting system to accomplish, many of which have been codified as voting system criterion. But, besides many of them being mutually-exclusive (i.e., you can't meet them all), using any of them as a litmus test obscures the true objective, in the same way that focusing exclusively on being strategy-free obscures the true objective. The only true measure of a voting system is it's expected performance: how well it delivers a desirable candidate to the electorate. Average performance, as measured by Bayesian regret, smooths over all the coarse edges of criterion, implicitly assessing all of them for frequency as well as impact.

Why should it matter that, for instance, approval voting fails the majority criteria, if the failure rate is vanishingly-infrequent and has minimal impact? When it performs significantly better than a host of other systems that do meet this criteria, but fail some other, equally-reasonable criteria? It shouldn't. Holding the percentage of strategic voters constant, approval voting has significantly better performance than just about any other voting method. Range voting (AKA score voting) can be even better. Which criteria are passed are secondary to that fact.