On 01/28/2012 09:13 PM, MIKE OSSIPOFF wrote:
On a related subject: The other thing lacking at EM, in addition to mock elections, is support for claims that a criterion is important. We hear, “I consider this criterion to be very important”. But such assertions need to be supported by explanation of _why_ you consider that criterion important. Why should others consider it important? What practical problems are present in non-complying methods but not in complying methods? What would it be like to vote in a non-complying method?
I don't think you can find any universal reason for why a criterion is important. The reasons have to be developed on a criterion-by-criterion basis.
The logic for focusing on criteria is that no deterministic ranked method can pass all of Arrow's "ultimate criteria" (unrestricted domain, non-dictatorship, Pareto efficiency, and IIA), we have to either say "every method is equally good" (which is nonsense), or see how close we can get.
We know that if some method X passes all criteria Y does and then some, we can suppose that X is better than Y. In game theory terms, X dominates Y. But if X passes some Y doesn't and Y passes some X doesn't, then who knows? It all becomes a matter of tradeoffs, and the relative value of each criterion is not very clear. Criterion failure may have implications for how the whole system using the method evolves, so Bayesian-regret type calculations may get the relative values wrong.
Take Borda. It has a respectable Bayesian regret but is so vulnerable to cloning and voting strategy so simple that in practice, it does very badly. It invites parties to field massive numbers of candidates, so the party with the most people win.
Still, I think we can see some patterns. We have absolute criteria (majority, Condorcet, mutual majority, etc), and relative criteria (FBC, monotonicity, independence of clones, etc). The former state possibly desirable properties. The majority criterion is part of majority rule, and mutual majority extends it from sets to candidates, for instance. On the other hand, the latter tends to involve the method behaving in a self-consistent manner, or freeing the voters or candidates from strategy. When a method fails monotonicity, you have two nearly equal ballot sets, and a change from one to the other that ought to make X's claim stronger, but X loses. The method is inconsistent with itself and "gets it wrong" in one of the two cases. Further, in a real election, we don't know whether the real result or the one from the hypothetical transformed election is the wrong one, so there's a chance the real result is dubious, too. Some people think that monotonicity is about strategy, but I think it's about the performance of the method, and about how it can justify (or not) its results. FBC is a strategy-related criterion. When a method fails FBC, that means that voters may have to rank their favorite artificially low to get the result they want. That means voters can't be absolutely sure they don't have to compromise. When a method is not cloneproof, that means uncoordinated entry or exit of candidates can make a great difference. The candidates know that even if they're completely alike each other, they can make the result change by joining or leaving.
Beyond the absolute/relative distinction, I think we can also distinguish between "cheap" criteria (e.g. unanimity) and "expensive" ones (e.g. strong FBC, later-no-harm). If a method is built to pass an expensive criterion but thereby fail a number of criteria other methods pass, one should give a strong argument for why that's desirable. In your case, you give as your strong argument that by passing the FBC (which is expensive and excludes many other criteria), you free the voters from having to even think about ranking someone above their favorite.
Passing strategy-related criteria gives a method undistorted data from which to act. Passing absolute criteria and consistency-related relative ones means the method will use that data well. But how much distortion the former removes and how well each latter criterion implies is not easily discovered.
(At this point, some people try to get around methods failing certain criteria by saying "sure, it fails, but it doesn't fail where it counts". But it can easily lead to a lot of back-and-forth about what "where it counts" really means and what one really wants of an election method. Pass/fail, in contrast, is completely unambiguous. Either a method passes or it doesn't, and if a method passes a criterion everywhere, then obviously it passes it "where it counts", no matter where that might be.)
---- Election-Methods mailing list - see http://electorama.com/em for list info
