STAR has been extremely divisive on the EM-list. Lots of people have some very harsh things to say about it. This wasn't just the usual jockeying from perfectionists, and I think I've come around to the other EM-listers on this. The opposition to STAR seems to be common to both cardinal and Condorcet supporters. Everyone seems to agree either Smith//Score or plain Score would be better than STAR.
Most of these criticisms boil down to STAR managing to break every criterion in the book. Adding a runoff destroys favorite betrayal, later-no-help, participation, . It does this in to try and prevent a strategic voting dilemma that:
- Has never been observed in an actual election (and is even named after an election where it notably didn't happen),
- Is probably rare,
- Would incentivize party elites to clear the field in the first place to avoid this, and
- Seems almost impossible to accomplish without any other candidate noticing and retaliating.
Personally, I was supportive of STAR until @SaraWolk's comment about the possibility of candidates not running in cloned pairs made me think through its potential turkey-raising problems. Burt Monroe has argued (very convincingly, in my opinion) that any system that fails turkey-raising will eventually be repealed. It doesn't matter that the expected value of a strategy is negative: empirically, parties push turkey-raising strategies even in systems where risks of blowback are extremely strong, e.g. the Democratic party's support for extreme Republican nominees in primaries, or their refusal to put up an alternative to Gavin Newsom in the recall.
It's true that criteria are worst-case guarantees, and what we care about is the average case. But for something as complex as a political system, we don't know the average case. Simulations are unrealistic and we've never run elections with STAR. All we know is in the worst case, STAR might do very badly.
If nothing else, these dramatic criteria failures severely limit the possibility of receiving endorsements from economists and social choice theorists. Few, if any, voting theorists who hear about STAR's criteria failures are eager to endorse it.
This leads me to suggest the following alternatives to STAR in its current form, which attempt to eliminate the small (but potentially disastrous) risk of turkey-raising.
- Provide a small reward to ballots that rate multiple candidates above 0. D2.1 does this by granting a single antiplurality vote to voters who give multiple approvals. A full antiplurality vote sounds too strong, but maybe something else would improve on it, e.g. down-weighting bullet votes to give only 4 stars to the favorite instead of 5?
- Using trimmed-means instead of the average rating.
- Smith//Score and Score DSV both received strong praise on the EM list as alternatives to STAR, especially if combined with a tied-at-the-top rule.
- Limit STAR runoffs to allied candidates (described elsewhere).
- Limit STAR runoffs to candidates who have strong mutual support (i.e. actual chickens or clones). For example, the runoff could consist of the score winner, paired against the candidate whose score is most positively correlated with the score winner's (across ballots).