Reassessing Nate Silver’s claim from last month that Democrats’ successful (in retrospect) 1-and-Done campaign against the California recall was “self-destructive . . . bad advice . . . a mistake”

Roy shared this take from political consultant Nate Silver last month:

It’s self-destructive more than self-interested. Pretty decent chance Newsom gets recalled. Democrats could potentially keep the seat if they urged their voters to consolidate behind an alternative Democrat but instead they’re telling them not to vote on the replacement!

Silver was referring to the “1 and Done” campaign by Democrats encouraging Californians to vote No on the initiative to recall the governor and to leave question 2 (the vote on the potential replacement) blank.

The logic of the “1 and Done” plan was clear enough: If Democrats coordinate to not vote on question 2, that makes it super-clear that the alternative to the governor will be a Republican (and, in this case, not just any Republican, but an extremist who, it turns out, was willing to endorse fabricated claims of fraud). This then makes a vote to keep the governor more appealing, when voters consider the alternative.

OK, that’s easy to say now, given what happened—but even in August this was clearly a possibility.

There’s also the more direct motivation that “1 and Done” is a cleaner message and does not involve voters having to fish through a pile of unappealing alternatives in order to find a Democrat to vote for.

But here was what Nate said:

Pretty much always, if someone tells you not to vote, they are giving you bad advice. If you live in California and leave the recall line blank, you are partly disenfranchising yourself and are making a mistake. To be clear, this *might* be in the best interest of Newsom—on the unproven and but at least vaguely plausible notion that voters might be more likely to vote “yes” on the recall if there’s another D to vote for—but it’s likely not in the best interests of the Democratic Party. After all, the DIRECT effect of this strategy is that it’s MUCH more likely a Republican replaces Newsom, conditional upon the recall succeeding. You’d have to be VERY confident that this strategy makes the recall less likely to succeed to outweigh that. If, for instance, this strategy increases Newsom’s chances of winning the recall from 70% to 75%, but reduces the Democrats’ chance of winning the replacement race from 50% to 10%, the chance of ending up with an R governor rises from 15% to 22.5%.

I guess it all depends on the numbers. As of mid-August, the polls seemed to show that the governor might be recalled. It turns out that this calculation was a mistake, based on a poll so flawed that the organization that produced it disowned the result—but it wasn’t until a few weeks later that the org revealed the problem, so we can’t blame Nate for thinking in August that the race was closer than it was. I guess the California Democrats were doing internal polling, and the true picture was clearer to them.

The interesting thing here is to try to figure out where Nate’s guesses of “from 70% to 75%” and “from 50% to 10%” came from.

A change in win probability from 70% to 75% corresponds to a tiny, tiny shift in the expected vote. I’ve done this sort of calculation before . . . let’s just do it again right here. Start with the 70%. You can get that probability if the expected vote proportion for No is 52% with an uncertainty (standard error) of 3.8%:

> 1 - pnorm(.5, .52, .038)
[1] 0.70

To bump this up to 75%, how much do we need to change the expected vote proportion? Let’s play around . . . ok, it.s

> 1 - pnorm(.5, .5255, .038)
[1] 0.75

OK, so increasing the chance of a No outcome from 70% to 75% corresponds to increasing the expected No vote from 52% to 52.55% . . . Considering how much variation there is in the vote, that’s basically nothing!

You can play around with the numbers a bit, but you’re always gonna get something like that. For example, we can quadruple the uncertainty:

> 1 - pnorm(.5, .60, .19)
[1] 0.70

And then to bring the probability from 70% to 75% you need to increase the expected No vote from 60% to 63%, which still just isn’t that much.

On the other side, what about the claim that the “1 and Done” plan was reducing the Democrats’ chances to win the replacement election from 50% to 10%, conditional on the recall happening. I’m skeptical. Again, easy for me to say this now, after the election’s over, but hear me out. Conditional on the recall winning, that means enough voters are angry at the governor to vote Yes on recall, and it makes sense that most of these Yes-on-recall voters will want to choose a Republican. So to beat this, the Democrats would need to “consolidate behind an alternative Democrat,” in Nate’s words. But any such coordination could cost the Democrats a lot more than 0.5 percentage points, or even 3 percentage points, of the recall vote: to the extent that the Democrats could actually win the replacement election, this strongly changes a voter’s decision factors on the recall. It’s also not clear that the coordination involved with choosing a single Democrat (especially given the options available on the ballot) would be easier than the coordination involved in 1 and Done, which is a pretty simple tie-me-to-the-mast strategy.

I think Nate’s reactions (“self-destructive” . . . “bad advice” . . . “a mistake” . . . and some more in ALL CAPS) were much more dependent than he realized on some off-the-cuff probability calculations. What if instead we say that the “1 and Done” plan increased the chance of No on the recall from 50% to 90% and reduced the chance of a Democrat winning the replacement race from 30% to 25%? Then the chance of ending up with a Republican governor decreases from 35% to 7.5%.

Now, I’m not saying these numbers are right either—I just constructed them by taking Nate’s numbers and flipping them around. I think his key mistake was not realizing how tiny a change from 70% to 75% really is. That assumption of such a small benefit of “1 and Done” is what drives his strong conclusions.

In retrospect all the above numbers are way off, as No won in a landslide—but, again, one reason for this outcome was that the alternative to the governor was unappealing to so many voters—and that, in turn, was one of the goals of the “1 and Done” plan.

Again, it’s easy to write all this in hindsight, and given the polls in August it’s hard to say that Nate was clearly wrong in his above post. But I think he was a bit overconfident in his punditry, and I think that even in retrospect he missed the strategic aspect of the two questions on the ballot. He’s a big poker player so I’m surprised he didn’t think this one through, but we all miss things sometimes. The important thing is to learn from our mistakes, and Nate has a good track record of doing this.

Also, I guess it’s important for political candidates not to take too seriously the strategy tips offered by outside analysts. Including myself. I didn’t make any predictions or recommendations at all for this California race, so I’m certainly not putting myself as some sort of prognosticator or strategy expert here. I’m just coming in from the outside here and offering some reflections. That can be valuable too.

All of this is an interesting lesson on the dependence of decision recommendations on assumptions. Start with Nate’s numerical assumptions and you get a strong recommendation, complete with ALL CAPS shouting at the hapless California Democratic party. Step back and think about where those numbers came from and you get a different story. We should all remember to step back and think about where our numbers come from. I continue to think that Twitter is part of the problem, as it so often seems to be a place where people have just enough space to make strong claims but no space to document, explore, or apply self-criticism to these claims, or to respond seriously (rather than dismissively) to disagreement.

P.S. It’s kinda weird. Nate is so careful to avoid overconfidence in his election forecasts (perhaps in reaction to his decision back in 2015 to assign Trump a 2% chance of receiving the Republican presidential nomination), I’d say going too far in the direction of allowing really wacky outcomes (such as the notorious map where Biden wins every state except New Jersey) . . . but then he flips around and acts so confident in his judgment of a particular political strategy set up for a particular election. I think that’s a statistical error on Nate’s part. As discussed above, any calculation of the effectiveness of the “1 and Done” strategy depends on speculative assumptions. That’s not to say that it’s impossible to make such judgments—the California Democratic party looked at the data they had and made their decision—just that it seems strange that Nate was so sure that he knew better. Maybe this is a holdover from baseball statistics, where he learned to trust the numbers and not listen to the so-called experts. But here’s the thing: there’s a big jump from being skeptical about expert claims and being sure that the experts are wrong. Maybe also Nate fell in love with the story that the goals of the governor and the interests of the state’s Democrats were in conflict. It’s a cool story, that the selfish governor is driving the party to its doom, and it’s a story worth considering, but ultimately you have to get back to the numbers.

17 thoughts on “Reassessing Nate Silver’s claim from last month that Democrats’ successful (in retrospect) 1-and-Done campaign against the California recall was “self-destructive . . . bad advice . . . a mistake”

  1. As I understand it the original theory of Twitter was that it would be a forum for people to share links to longer things. The people I follow tend to use it that way. You can weed the hot take people out of your timeline for the most part.

  2. The problem I have with Nate’s advice is that he is failing to take account of the effect of any other strategy other than “1 and Done” on the answer to the first question. How would the Dems have messaged a “Keep Newsom, but Vote for this Dem if you think Newsom sucks.” By putting another Dem forward you are conceding that Newsom should go at some level. How does the other Dem candidate campaign? Do you actually expect that Dem not to start firing on Newsom making the recall more likely. Do you not expect that Dem to siphon off votes on question 1?

    It would be one thing if Newsom would concede. But as long as he is not conceding, any other candidate will have to campaign against him splitting the Dem vote. It is also silly for Nate to think that the Dems can consolidate around another candidate. Who is going to make that consolidation possible. Newsom is effectively the head of the party. If some lesser leaders defected, they wouldn’t have the ability to block others. We would likely end up with many candidates, and no primary process for the Dems to decide. Dems vote would split and make recall much more likely and a Rep. victory more likely. Even if the polls in August had been correct, that “1 and Done” was the only strategy that made any sense.

    • He may be wrong about the magnitude, but he’s not “failing to take account of the effect of any other strategy on the answer to the first question”.

      That’s what the “If, for instance, this strategy increases Newsom’s chances of winning the recall from 70% to 75%” is about.

      • “this strategy” in Nate’s comment refers to the “1 and Done’. He has not taken account of other strategies effects. It is hard to imagine any other strategy that would not have made recall more likely than “1 and Done.” And, recall would have substantially increased the chance of a Rep. winning.

        • I’m not sure I understand your argument.

          > “this strategy” in Nate’s comment refers to the “1 and Done’. He has not taken account of other strategies effects.

          He compares “this strategy” with some unspecified alternative strategy.

          > It is hard to imagine any other strategy that would not have made recall more likely than “1 and Done.”

          I think he would agree and his unspecified alternative strategy makes indeed the recall more likely than the “1 and Done” strategy (30% for the former, 25% for the latter).

          Maybe your point is that a better estimate of Newsom’s change of winning with the kind of alternative strategy being discussed would not be 5% less than with “1 and Done” but 25% less or whatever. That’s a valid point, it’s “just” about the magnitude though. (But it can change the conclusion!)

  3. >”In retrospect all the above numbers are way off, as No won in a landslide”

    So you’re saying that with the benefit of knowing the outcome, a predicted 75% chance of winning was way off.

    All kinds of events (news (scandals, found documents, eyewitnesses, withheld stories, damning recordings), health, weather, etc^10) could have arisen between the 3-week-prior prediction and the election that could have changed things. Events that may not change the outcome but could turn landslides into ordinary wins, marginal wins.

    Without knowing the outcome, three weeks before the outcome was known, entertaining 4 scenarios

    25% win by a landslide
    25% win by an ordinary amount
    25% win by a bit
    25% lose by a bit

    seems reasonable and gives a 1/4 chance to the outcome that obtained. No big surprise.

    In contrast, saying a football team has a 1 in 10,000 chance of winning (an example we once talked about) is way off, but you can judge that before you know the outcome. Unforeseen events can come in and help an underdog with a high enough chance make all 1 in 10,000 soccer forecasts unreasonable).

    • Dan:

      My disagreement is not with Nate’s assessment of a 75% chance of a No vote—as I wrote in the above post, that’s not completely unreasonable conditional on that August poll whose fatal flaw was only revealed weeks later. Rather, I’m skeptical of his supposition that “this strategy increases Newsom’s chances of winning the recall from 70% to 75%, but reduces the Democrats’ chance of winning the replacement race from 50% to 10%.” Say what you want about the 75% number, but the other numbers (70%, 50%, 10%) seem to be pulled out of nowhere. One of the points in my post above was that the “from 70% to 75%” change corresponds to a tiny shift in the expected vote proportion. So Nate was pretty much assuming that “1 and Done” would not work at its main goal (increasing the No vote) and, from that, concluding that “1 and Done” was “self-destructive . . . bad advice . . . a mistake” etc. That’s basically circular reasoning, but I don’t think he realized this, because I don’t think he thought through the implications of “from 70% to 75%.”

      • P.S. The thing that keeps puzzling me here is now sure Nate seemed to be, back in August, in his judgment about the strategy. If the uncertainty in the election outcome is as huge as you’re implying in your above comment, then it seems weird to be soooo sure that “1 and Done” would only have trivial effects on the vote outcome. If the forecast uncertainty is so wide (with vote for No predicted to be somewhere between 0.45 and 0.75, say), then I’d think this would allow the possibility that strategies could have large effects.

      • Yes, in a world where “1 and Done” has no effect, picking some other strategy would make some sense. But, that involves a failure to think of the alternative strategies effects. Some of the alternatives clearly involve multiple Dem canidates taking shots at Newsom. Nate seems to think that we could have had some stable strategy where a bunch of Dems get behind one Dem candidate. But, that means having the Dem establishment going after Newsom. That would have had a hugh effect on the recall vote.

        • Steve:

          And, again, you’d expect that a poker player would think about these strategic issues! I think Nate just fell in love with his story and then started to gather arguments to support his position. Too bad, but it happens to all of us. I hope he can learn from this experience, first to be more thoughtful when assigning probabilities to unknown events, and second to be more careful with the punditry. Saying something loudly doesn’t make it true.

      • Andrew,
        In at least a few places you have the meaning of a ‘no’ vote backwards. For instance, you say “Conditional on the recall winning, that means enough voters are angry at the governor to vote No, and it makes sense that most of these No voters will want to choose a Republican.” But “No” means “No recall”. You’ve got it backwards: Conditional on the recall winning, that means enough voters are at the governor to vote Yes!

        In any case, I think it made sense for the Democrats to keep attractive alternatives to Newsom off the ballot, for the reasons stated. But, given that they were very successful in keeping _Democratic_ alternatives off the ballot, I think they didn’t need to go with “1 and done”: if Newsom said “Please vote No on the recall, and if you want to vote for someone on the second ballot just in case the recall succeeds, go ahead and pick whoever you think is least bad,” I don’t think that would have cost him any votes on the recall question.

        I also think it’s terrible messaging to tell people not to vote. But that’s a different issue.

        • I disagree that “1-and-Done” is telling people not to vote. The reason is that these are not two separate questions, such as voting for governor and then voting for a senator; they are intrinsically tied. (note: I’m not passing this off as a unique insight; it’s part of the point of Andrew’s post). Can we imagine a situation in which people would not vote on Q1 but have strong enough opinions on Q2 to vote just for that one? Instead, it seems (having not been involved in or even reading about the strategy development) the California Democrats wanted people to think of the two questions as a single vote. And, presumably for the reasons already mentioned about the dissonance of “keep the current person, but vote for this other person just in case”. Plus, as was mentioned, you’re asking people to scan through a long list of randos to find your person. They wanted people to vote. Making the vote easy to understand was part of that strategy.

        • Phil:

          1. Thanks for pointing out where I got things backward in the writing. I fixed—at least, I fixed all the places that I could find!

          2. If I’d been voting in California, I’d personally have chosen somebody for question 2, no matter how I’d voted on question 1. Because, sure, why not, it’s free. But when considering the election as a whole, you also need to think about coordination strategies among like-minded voters, and part of this coordination among the Democrats was to not vote on the second question.

          But, again, I’m not saying that I’m sure that “1 and Done” was the Democrats’ best message. I’m just annoyed by how sure Nate was that he knew best. It’s tricky: sometimes the political establishment does screw up, and there are principal-agent problems invovling campaign consultants and politicians. So I respect that an analyst such as Nate can offer an outside perspective, and it’s useful for the Nates of the world to second-guess campaign strategies. In this case, though, I think he got it wrong, and something can be learned from the mistake.

  4. Nate seems to be implicitly suggesting that the Democratic party should be maximizing
    Pr(Dem governor), but was rather maximizing Pr(Newsome governor).
    I disagree.
    Insofar as a political party has a single maximand, it is probably some convex combination of Pr(Dem governor) and Pr(Newsome governor).
    And there is a good argument that it should be weighted more heavily towards Pr(Newsome governor) when you consider the implications for future recall elections to avoid the Republicans taking advantage of infighting.

  5. Another facet to consider:
    Newsom had the incumbent advantage. For the Democratic party to take full advantage of this, they had to frame this vote as Newsom vs. Elder. If the public had mentally framed it as Elder vs. an “unknown” Democrat, the incumbent bonus ought to have vanished.

Leave a Reply

Your email address will not be published. Required fields are marked *