Thoughts on the 2020 Democratic presidential primary

This post is an endorsement of a Democratic candidate for president, but it is not a typical endorsement. When I set out to write this post, I didn’t have a particular conclusion in mind. Instead, I figured out how I wanted to think about the primary, then did research on the underlying facts, and finally came to a conclusion. Perhaps you’ll disagree with my analysis. If so, my hope is that you take away something valuable about the process I used to reach my conclusion.

These thoughts will be most useful to you if:

(1) You think about things in a similar way to me (a quantitative/analytic sort of thinking).

(2) You think Trump is very bad (I argue this below because it’s important to my argument that Trump is in fact extremely bad, but I think I’m unlikely to convince you if you don’t already think this).

(3) You buy into the thesis of this essay: the world is complicated and we sometimes need to make decisions based on incomplete information, based on numbers we don’t have. In such cases, we should make our best guess about the numbers anyway and work with them, even if we realize we could do better given more information.


In this post I will make a three-pronged argument:

Part 1: If you compare the various people who might be elected president in 2020 by how good of a president they will be, the difference between Trump and the Democratic candidates is vastly larger than the difference between the various Democrats.* As a result, if one candidate is more likely to beat Trump than another — even by as little as 5% — it is preferable for that candidate to be the nominee. [Confidence: medium-high.]

*For reasons that I will discuss, I think Mike Bloomberg is a slight exception to this.

Part 2: (If you’re pressed for time, this is the part you should read. No, seriously, this is a long post and if you’re only going to read part of it, you should read Part 2.) Klobuchar (most electable) >= Biden >= Bloomberg >= Buttigieg > Sanders > Warren (least electable). (Here, “>=” means “very slightly/negligibly more electable in my estimation”.) [Confidence: low, but informed. That is, I could easily change my mind based on future evidence but I think I’ve evaluated existing evidence reasonably well.]

Part 3: In a multi-candidate race, it often makes sense to vote for someone who isn’t your first choice. For instance, in the 2000 election, if you preferred Nader to Gore to Bush, it made sense to vote for Gore because Nader didn’t stand a chance. For this reason, my actual recommendation of who to vote for may not be the most electable candidate, and may change over time. I will make a recommendation and explain my reasoning for it in this part of the post, and will update it throughout the primary, as my recommendation may changed depending on who has a good shot at the nomination.

As of March 1, I recommend voting for Joe Biden on Super Tuesday (with possible exceptions in edge cases). [Confidence: high assuming Parts 1 and 2, medium overall.]

[On January 31, I recommended caucusing for Joe Biden in the Iowa caucuses. I didn’t make recommendations in New Hampshire, Nevada, and South Carolina because I didn’t have time, but suspect I would have backed Buttigieg, Biden, and Biden, respectively.]





Thesis: If you compare the various people who might be elected president in 2020 by how good of a president they will be, the difference between Trump and the Democratic candidates is vastly larger than the difference between the various Democrats. As a result, if one candidate is more likely to beat Trump than another — even by as little as 5% — it is preferable for that candidate to be the nominee. [Confidence: medium-high.]

(If you’re already convinced of this, I encourage you to skip to Part 2.)


This argument comes in two parts. First I’ll argue that Trump is really bad. Then I’ll argue that from a practical standpoint, the difference between any two of the Democratic candidates being elected president isn’t that large.

A. Trump is really bad

I imagine I’m mostly preaching to the choir here, but this is an important part of my argument: if I want to argue that the differences between the Democrats are dwarfed by the difference between them and Trump — so much so that electability is essentially the only thing we should care about — then I need to demonstrate that Trump is much, much worse than the Democrats running.

In my view, the worst things Trump has done fall into three categories: (1) weakening environmental protections, (2) acts and words of xenophobia, and (3) undermining democracy.


(1) Weakening environmental protections

For a comprehensive summary of Trump’s actions on this front, see this article.

These rollbacks — particularly those concerning air pollution and emissions — are really bad for the world because they accelerate the pace of climate change. Climate change is potentially a catastrophic risk because it will cause sea level rise that will wipe out entire cities. A whopping 150 million people live in places that will be below the projected high tide line in 2050. At the same time, experts predict that climate change will cause global shortages of drinkable water. These effects — not to mention others — will create huge social, political, and economic instability. I’m afraid this will happen no matter what we do at this point, but Trump’s rollbacks certainly contribute to the scale of the problem. If Trump is reelected, we should expect the rollbacks to continue.

A separate issue is air pollution. Fine particulate matter pollution (the most harmful kind of air pollution) in the United States had been steadily declining through 2016. Then, in 2017, and again in 2018, fine particulate matter levels increased, reaching levels that were 5.5% higher than in 2016. One (admittedly not the only) cause of this trend is likely that the Clean Air Act has been much more weakly enforced since Trump took office. Although less serious than climate change, air pollution is an important issue: the 5% increase in fine particulate matter levels is estimated to have caused 10,000 premature deaths. A full reversal of clean air standards would likely cause hundreds of thousands of additional deaths every year, and we should expect Clean Air Act enforcement to continue to weaken if Trump is reelected.


(2) Acts and words of xenophobia

One of Trump’s first acts as president was signing a ban on travel from seven Muslim-majority countries, as well as a blanket ban on the admission of Syrian refugees. Leaving aside the question of whether the ban is effective at its stated goal of preventing terror attacks (probably not), an important effect of this policy is that it helped normalize bigotry. Trump’s campaign rhetoric — in which he called for a “total and complete shutdown of Muslims entering the United States” — demonstrates that the policy came from a place of xenophobia. Whether you view the travel ban as a policy that discriminates on the basis of nationality or on the basis of religion, it is clearly a step backward that is reminiscent of darker times in U.S. history.

More recently, the Trump administration implemented a family separation policy at the US-Mexico border. Child psychologists warn that this policy has caused lifelong trauma to the thousands of children affected. I see this policy as a needless act of cruelty. Perhaps the intention for this policy was to impose such a harsh punishment on migrants that they would be deterred by crossing the border. But even under a charitable interpretation, the policy seeks to modestly improve outcomes for American citizens at huge expense to non-citizens.

Those are two of Trump’s xenophobic policies. But let us not lose sight of Trump’s racist and xenophobic language: his insistence that Barack Obama wasn’t born in the U.S.; his 2015 comments calling Mexicans who crossed the U.S. border illegally drug dealers and rapists; his tweets telling four congresswomen (three of whom were from the U.S.) to go back to where they came from; and more.

These words and policies constitute a drift toward an America where espousing racism and xenophobia is regarded as more acceptable. There used to be a social norm that it wasn’t okay to say clearly racist things. Trump was in some sense the first big offender of this norm, which is why most high-profile Republicans condemned him following his original proposal of a Muslim ban. But Trump barrelled right through, and now when a politician says something racist, it won’t be as shocking: after all, we’re already used to the president of the United States making bigoted comments. By eroding this norm, Trump has made xenophobic policies more likely in the future, even after he stops being president. If he wins a second term, I expect this effect to become even stronger.

[Edit: I no longer stand by the previous paragraph, as it seems to me that norms against racism have strengthened over the Trump administration, perhaps as a result of backlash against Trump.]


(3) Undermining democracy

Democracy around the world rests on norms. Some basic norms — the bare minimum — include things like “don’t jail your political opponents”, “don’t rig elections”, and “don’t give yourself dictatorial powers.” America holds itself to a higher standard: its norms include “don’t use emergency powers to deal with things that aren’t emergencies”, “be honest (for the most part) about your administration’s actions”, and “don’t profit off of your elected position.” Such norms are important because they help the system of checks and balances do its job and prevent egregious abuses of power. Unfortunately, Trump has aggressively sabotaged democratic norms as president. Here are ten examples:

  1. He has (as of January 19th) lied over 16,000 times (a rate of about 15 lies a day), vastly more than any predecessor in recent memory. Now, when he lies, even about something important, it barely even makes the news.
  2. Perhaps the most damaging lie of them all: following the 2016 election, he (falsely) claimed that millions of people voted illegally. 28% of Americans believed him and an additional 35% said they weren’t sure. This has laid the groundwork for a similar — and potentially devastating — claim in 2020 if he loses the election.
  3. He withheld military aid to Ukraine in order to pressure the Ukrainian government to launch an investigation into Joe Biden.
  4. He has condoned (and maybe even encouraged?) Russia’s interference in the 2016 election, publicly declaring that he trusts Vladimir Putin over the U.S. intelligence community.
  5. He fired FBI director James Comey, likely for leading an investigation into Trump’s ties to Russia.
  6. He has demonized the mainstream media far more than his predecessors, even calling them the “enemy of the American people.” As a result, 72 percent of Republicans trust Trump more than the media, and 43 percent think Trump should have the power to shut down news organizations, statistics that are pretty scary for the future of the free press.
  7. After suffering a defeat in Congress, Trump declared a national emergency to be able to appropriate funds for the border wall, abusing a presidential power reserved for actual emergencies.
  8. He has gone after any Republican who has shown him disloyalty, thus making it impossible for any Republican who cares about their political future to criticize him in a public forum.
  9. He has profited off of the presidency by refusing to sever ties to his businesses. More importantly, this has affected U.S. foreign policy: under Trump, the United States gives special treatment to those who help Trump profit.
  10. He has undermined our alliances with democratic countries such as Germany, supported anti-democratic leaders in historically democratic countries such as Hungary, and showered praise on and granted legitimacy to authoritarian regimes such as North Korea’s.

These actions range from really bad to unprecedented in U.S. history. I see these violations as incredibly dangerous to America’s future. Trump has broken many healthy democratic norms, and it’s really hard to reestablish a norm that has been broken. Trump’s presidency may be a first step in the United States becoming an undemocratic country, and the ramifications of the world’s greatest power becoming undemocratic would be enormous.


B. The Democrats won’t be that different from each other as president

To state the obvious, all of the Democratic candidates would be far, far better than Trump with regard to environmental protection, respect for people of other countries and ethnicities, and democratic norms.1 Each of the Democrats will reinstate the regulations that were in place prior to the Trump presidency; they will not demonize foreigners and people of color; and there is no indication that any of them will violate norms on anywhere close to Trump’s scale.

But also, each of the Democrats will be able to get roughly the same things done if elected president. That’s because they are all much more liberal than whoever the median senator will be — Joe Manchin (D-WV), if Democrats are lucky. If Manchin doesn’t want Medicare for All, we won’t have Medicare for All. If he doesn’t want to ban semi-automatic weapons, we won’t ban semi-automatic weapons. I’m not taking a position on either policy, but I am asserting that the limiting factor on what passes through Congress won’t be what the president wants; it will be what someone like Joe Manchin (or Mitt Romney) wants.

A caveat here is that in some domains the president can act without Congressional oversight. The most prominent of these is foreign policy. And while there are certainly foreign policy differences between the candidates, they are reasonably small (as compared with the whole American political spectrum). All candidates want to mostly withdraw troops from Afghanistan (though there are differences as to whether a small number should remain). All candidates want to reinstate the Iran nuclear agreement. All candidates view NATO as a valuable tool for cooperation and multilateral sanctions. These positions contrast quite sharply with Trump’s words during the 2016 campaign and his subsequent actions.

In summary, I don’t think a Sanders presidency would be all that different from a Biden presidency. But both would be vastly better than a Trump presidency.


So let’s say your favorite of the major candidates (i.e. the one you’d most prefer as president) — call them Candidate A — has a 50% chance of beating Trump. And let’s say your least favorite of the major candidates (Candidate B) has a 55% chance of beating Trump. Who would you rather have as the nominee?

Or, put another way, would you prefer a 50% chance of Candidate A as president and a 50% chance of Trump, or a 55% chance of Candidate B and a 45% chance of Trump?

I think the answer is reasonably clear: Trump is so, so much worse than any of the Democratic contenders that you should prefer the latter. A fifty percent chance of a slightly better president matters less than avoiding a extra five percent chance of another four years of one of the worst presidents in United States history, one who will continue to exacerbate climate change, normalize bigotry, and lead the United States further down the path to authoritarianism.

So, to summarize this part: if one candidate is more likely to beat Trump than another — even by a small amount — it is better for America (and the world) if that candidate is the Democratic nominee.


1. A possible exception is Mike Bloomberg. I think there’s a solid argument to be made that Bloomberg is substantially worse on Democratic norms than any of the other candidates, because of his use of money to pressure individuals and groups into supporting him, e.g. when he was seeking to change the law so he could run for a third term as mayor.


So, which of the top Democratic contenders are more electable and which are less? That’s the subject of Part 2 of this essay.



Thesis: The rough order of the leading Democrats by electability is Klobuchar (most electable) >= Biden >= Bloomberg >= Buttigieg > Sanders > Warren (least electable). (Here, “>=” means “very slightly/negligibly more electable in my estimation”.) [Confidence: low, but informed. That is, I could easily change my mind based on future evidence but I think I’ve evaluated existing evidence reasonably well.]


Before I get into the weeds of the argument, I want to be very clear about what electability means here: I say that a candidate is electable if they have a high probability of beating Trump, conditional on being the Democratic nominee. For instance, if candidate A has a 10% chance of being the Democratic nominee and a 70% chance of beating Trump if they’re the nominee, and candidate B has a 50% chance of being the nominee and a 65% chance of beating Trump if they’re the nominee, then candidate A is (slightly) more electable, even though candidate B is overall more likely to be elected.

People’s intuitions about electability are notoriously bad. In 2008, the conventional wisdom was that Barack Obama was unelectable. In the 2016 cycle, people thought Hillary Clinton was very electable. (I remember thinking that Clinton was a uniquely electable candidate as late as 2015.) More generally, people’s intuitions about electability are colored by which candidate they like and whose policies they prefer.


If we can’t rely on our intuitions to judge electability, what can we do? The most obvious thing is to look at data: what lessons about electability can we learn from past elections? This brings us to the first thing that we (probably) know about electability:

(1) After adjusting for name recognition, candidates who do better in head-to-head polls against Trump are more electable.

In an earlier blog post, I argued that general election polls this far ahead of the election are predictive of the eventual outcome. I estimate that if candidate A currently does 10 points better against Trump in head-to-head polls compared to candidate B (e.g. A beats Trump by 8 points in polls, whereas B loses to Trump by 2 points), then we should expect candidate A to outperform candidate B by about 4 points in a hypothetical general election matchup against Trump. If A does 5 points better than B in polls, we’d expect that A would do 2 points better against Trump than B would, and so on.

(How did I come up with this estimate? There’s a lot to say! See Appendix A (warning: the discussion gets pretty technical).)

This is a pretty big effect: if one Democratic candidate is doing 5 points better in polls against Trump compared to another candidate — and thus is expected to do 2 points better in the general election — that perhaps amounts to a whole 20% difference in the probability that the first candidate wins (because this election is likely to be quite close)!

A caveat: poorly-known candidates poll systematically worse in head-to-head polls against Trump. That’s because Trump supporters mostly just say they’ll vote for Trump no matter who he’s matched up against, whereas when Democratic-leaning independents hear “Wayne Messam” they often don’t give a response, seeing as they don’t know who Wayne Messam is.

Luckily for us, in June of 2019, HarrisX asked how voters would vote in head-to-head match-ups between Trump and 22 of the Democratic presidents, including Wayne Messam, Seth Moulton, and other unknowns. I combined these with results from a YouGov poll that asked voters which candidates they had heard about enough to form an opinion on. I found that if a candidate was 10% less well-known, they on average did 1.9% worse in head-to-head polls against Trump. (This was consistent with data from polls by other firms that featured fewer candidates.)


One more caveat: this criterion of comparison is also much less useful for lesser-known candidates. We shouldn’t expect voters to have an informed opinion of Amy Klobuchar because voters don’t know much about her right now. This means that unlike for Biden, whose current popularity might be a reasonable reflection of how popular he will be in November, Klobuchar’s head-to-head numbers against Trump won’t mean much.


But polls aren’t the only thing we can look at to predict electability. Because there have been so few recent presidential elections — only 15 since World War II — we may be able to learn more by looking at recent congressional races, for which our sample size is in the thousands. While races for the House and Senate are not entirely analogous to the campaign for president, we should expect patterns in congressional races to at least somewhat apply to presidential campaigns as well. This brings us to the second thing we (probably) know about electability.

(2) Candidates whose ideologies are closer to the political center are more electable.

This finding is very robust for U.S. House races. A study by Andrew Hall found that when an ideologically extreme candidate barely wins a primary over a moderate candidate, the candidate does 9 to 13 points worse in the general election than when a moderate candidate barely beats an extreme one, resulting in a 35 to 54 percent decrease in the probability of victory.

We should probably not expect such a large effect size for the 2020 presidential election. That’s because Hall’s study looked at elections between 1980 and 2010; since then, the number of swing voters has declined substantially, so swings in vote share from election to election are much smaller. However, we should probably expect the effect to exist in some capacity; moderate Democrats did better than progressive Democrats in the 2018 midterms, for instance.


Finally, a slightly less established (but still probably correct) conclusion to draw about electability:

(3) Candidates who are popular in their home state/district, or do well in their local elections, relative to what their state/district’s partisan lean would suggest, are more electable.

FiveThirtyEight has talked about this, e.g. in this article about Amy Klobuchar. Frustratingly, I can’t seem to find data to back up this claim, so I’m mostly relying on the fact that I trust the people at FiveThirtyEight to know what they’re talking about. But also, this proposition makes a lot of sense. If you do well in your local election, that means you have something — maybe popular policies, maybe charisma, maybe authenticity — that makes people want to vote for you.

The reason I’m ranking this criterion third is that in judging this criterion worthwhile, I’m only going on my trust of FiveThirtyEight and on common sense — whereas hard data (which supports the first two criteria) would be a more reliable basis. Also, the power of a candidate’s home approval rating to predict their national popularity may be limited, since members of Congress are often popular for things like constituent service, and popularity is not likely to extend to national appeal in such cases.

In judging the Democratic candidates by this metric, I’ll use approval rating rather than historical election results when possible. A politician’s election results depend a lot on their opponent, so using them would introduce noise into our measurements.

(A caveat: state/district popularity is a lot less informative for mayors and governors than for members of Congress, because the role of a mayor and governor is substantially less political and people cross party lines way more when voting for these offices. Data from many years ago is also less informative, as things may have changed since then.)


When assessing the electability of the top candidates, I will put the most weight on Criteria 1 and 2, as the data suggests reasonably strong effects for both. I think Criterion 3 deserves about half as much weight. (In some cases I will down-weight the criteria for certain candidates for the reasons I mentioned above: Klobuchar and Buttigieg are less well-known, so their polls against Trump are likely less predictive; Buttigieg’s popularity in South Bend isn’t super important because he’s a mayor, not a senator, and Biden’s popularity as senator is less important because that was a long time ago.)

With all that said, let’s compare the leading candidates by how they score on the indicators I listed above (and any additional considerations relevant to the specific candidates, though for the most part I’ll stay neutral unless I have compelling data).


Joe Biden

(1) Polls against Trump: somewhat high. In an average of high-quality polls in January, Biden has led Trump by 4.7 points.

(2) Moderation: high. On a scale from 1 = very liberal to 5 = moderate, FiveThirtyEight gave Biden a subjective rating of 4/5 on both economic and social policy, for an average of 4/5, tied for the highest of any candidate.

(3) Home state popularity: slightly high but not very informative. Joe Biden outperformed Delaware’s partisan lean in his Senate elections, but that was a long time ago and may not be pertinent to today.

(4) Other factors: neutral. Joe Biden grew up in Pennsylvania, so he will probably have a small advantage there over other potential candidates. On the other hand, it’s reasonable to be concerned with Joe Biden’s health and ability to perform well in the general election debates. However, his mediocre debate performances so far haven’t hurt him with Democratic primary voters, and debates change minds a lot less in the general election than in the primary.

Overall electability assessment: somewhat high.


Bernie Sanders

(1) Polls against Trump: somewhat high but this probably overestimates him. In an average of high-quality polls in February, Sanders has led Trump by 4.5 points. However, this fascinating article by political scientists David Broockman and Joshua Kalla argues that we should expect general election polls to be overestimating Sanders (and only Sanders). The reason he does about as well as Biden in these polls is that what he loses relative to Biden among registered voters who are likely to vote (older, more moderate Democrats), he gains among registered voters who are less likely to vote (generally younger voters). This effect is unique to Sanders. In order to maintain his current advantage in polls of registered voters on election day, Sanders would need a vastly larger youth turnout than we’ve ever seen. This is very unlikely to happen (and we have some evidence of this already: Sanders isn’t turning out many new voters in the primary). Based on the paper by Broockman and Kalla, it seems we should expect Sanders 1-2 points worse than he’s polling, so I would say it is fairer to characterize this indicator as pointing toward Sanders’ electability being medium.

(2) Moderation: very low. On a scale from 1 = very liberal to 5 = moderate, FiveThirtyEight gave Sanders a subjective rating of 1/5 on economic policy and 1.5/5 on social policy, for an average of 1.25/5, the lowest of any candidate.

(3) Home state popularity: medium. Per this article, Sanders’ net approval (approve minus disapprove) is 7 points higher than expected for a Democratic politician from Vermont, which is pretty average among the major candidates.

(4) Other factors: neutral. Just as with Biden, it’s reasonable to have concerns about Sanders’ health. On the other hand, Nate Cohn of the New York Times, whom I respect a lot, has a theory that Sanders will be particularly effective at turning out/persuading a particular breed of moderate voters.

Overall electability assessment: somewhat low.


Mike Bloomberg

(1) Polls against Trump: high but this probably overestimates him. In an average of high-quality polls in February, Sanders has led Trump by 5.4 points, after adjusting for voter familiarity. However, Bloomberg has spent hundreds of millions of dollars advertising himself, something other Democrats have not. Bloomberg would of course continue doing this if he were the nominee; however, other Democrats would also advertise themselves, likely well past the point of diminishing returns. This means we should perhaps expect Bloomberg’s general election poll numbers to fall (once Trump starts running large numbers of ads) but shouldn’t expect this of the other Democrats (because they will also start running ads). For this reason, I would say it is fairer to characterize this indicator as pointing toward Sanders’ electability being somewhat high.

(2) Moderation: high. On a scale from 1 = very liberal to 5 = moderate, FiveThirtyEight gave Bloomberg a subjective rating of 4.5/5 on economic policy and 3.5/5 on social policy, for an average of 4/5, tied for the highest of any candidate.

(3) Home state popularity: confusing; best to disregard. Running as an Independent, Bloomberg beat his Democratic rival in the 2009 New York mayoral election by 4 points. In 2005, running as a Republican he beat the Democrat by 19 points and in 2001 by 2 points. This is impressive performance for a Republican in New York, but I have no idea what it means for your electability as a Democratic presidential candidate if you were popular in your home state back when you were a Republican. Besides, that was a long time ago.

(4) Other factors: neutral. I don’t have anything else to add, but generally speaking Bloomberg is the candidate about whose electability I’m most uncertain. That’s because I have a lot of uncertainty about factor (1) and think factor (3) doesn’t tell us anything.

Overall electability assessment: somewhat high (but high uncertainty).


Elizabeth Warren

(1) Polls against Trump: very low. In an average of high-quality polls in February, Warren has led Trump by 1.4 points after adjusting for voter familiarity, substantially lower than any other candidate.

(2) Moderation: very low. On a scale from 1 = very liberal to 5 = moderate, FiveThirtyEight gave Warren a subjective rating of 1.5/5 on both economic policy and social policy, for an average of 1.5/5.

(3) Home state popularity: very low. Per this article, Warren’s net approval (approve minus disapprove) is 20 points lower than expected for a Democratic senator from Massachusetts.

(4) Other factors: neutral. Some people might try to make an argument one way or the other about how Warren’s gender affects her electability. The argument that it hurts her chances is that sexism may have hurt Hillary Clinton’s run for president, and so our priors should be that it will also hurt Warren. The argument that it helps Warren is that women had a slightly higher win percentage in 2018 general election races, and that this should shift our priors toward women being more electable. I’m agnostic on this point and think it’s probably minor compared to factors (1) through (3).

Overall electability assessment: low.


Pete Buttigieg

(1) Polls against Trump: somewhat high. In an average of high-quality polls in February, Buttigieg has led Trump by 4.4 points, after adjusting for voter familiarity. However, 20-25% of voters still don’t have an opinion of Buttigieg, so I wouldn’t read too much into this.

(2) Moderation: somewhat high. On a scale from 1 = very liberal to 5 = moderate, FiveThirtyEight gave Buttigieg a subjective rating of 4/5 on economic policy and 3/5 on social policy, for an average of 3.5/5.

(3) Home state popularity: somewhat high but not very informative. In his two mayoral elections, Buttigieg won 71% (2011) and 80% (2015) of the vote. This is much better than the Democratic candidate in 2007 (62%) and in 2019 (63%). However, I think mayoral election results relatively uninformative, because mayor is not a federal political office. Also, Buttigieg did a little worse than one would expect in 2010 in his only statewide race (though that was a long time ago).

(4) Other factors: neutral. Buttigieg is from the Midwest; on the other hand, people say he comes off as elitist, which is probably the opposite of what you want if you want to appeal to Midwestern voters. Also, one could argue that Buttigieg will have trouble turning out the black vote. However, one could also argue that his trouble with the black vote is already baked into his poll numbers, and that he will start doing better with black voters if he is the nominee. I have no idea which argument is correct. I’ll stay neutral and avoid non-data-driven speculation.

Overall electability assessment: somewhat high.


Amy Klobuchar

(1) Polls against Trump: slightly high but not very informative. In an average of high-quality polls in February, Klobuchar has led Trump by 4.2 points, after adjusting for voter familiarity. However, about 30% of voters still don’t have an opinion of Klobuchar, so I wouldn’t read too much into this.

(2) Moderation: somewhat high. On a scale from 1 = very liberal to 5 = moderate, FiveThirtyEight gave Klobuchar a subjective rating of 4/5 on economic policy and 3.5/5 on social policy, for an average of 3.75/5.

(3) Home state popularity: high. Per this article, Klobuchar’s net approval (approve minus disapprove) is 30 points higher than expected for a Democratic politician from Minnesota. This is consistent with her impressive performance in the 2018 elections, where she won Minnesota by 24 points in a year when fellow Democratic senator Tina Smith and governor Tim Walz both won by 11 points.

(4) Other factors: positive. Klobuchar is from Minnesota, a swing state. There’s little doubt in my mind that she will win Minnesota if she is the nominee. (In an October Mason-Dixon poll, Klobuchar outpolled Trump by 17 points, compared to 12 points for Biden, 11 for Warren, and 9 for Sanders.) To whatever extent her appeal particularly extends to neighboring Wisconsin and other Midwestern states, that’s an additional plus, though currently this hypothetical advantage isn’t showing up in polls. (Also relevant to Klobuchar: see what I wrote under “Elizabeth Warren” about gender.)

Overall electability assessment: somewhat high.


Here’s a summary chart. The cells with *s are the ones I think aren’t very informative (for the reasons above).


In the “Prob(win)” column, I listed my estimate of the probability that each candidate would win in a general election against Trump. These numbers come from combining the three factors in a 2:2:1 ratio. You shouldn’t take these numbers as being particularly precise; they just represent my best guess at the electabilities — albeit an educated guess, arrived upon systematically based on the three factors — and are very much subject to change as we get new information about the candidates (and particularly new polls). If you’re interested in the details of how I arrived at these numbers, there are plenty of those; see Appendix B!

But if you’re not interested in the details, or disagree with my methods, it’s not too important if you object to the specific probabilities. I hope, though, that I have justified the order of the probabilities relative to each other. If you’re trying to figure out who to vote for, it’s the relative probabilities that matter.



Thesis: In a multi-candidate race, it often makes sense to vote for someone who isn’t your first choice. For instance, in the 2000 election, if you preferred Nader to Gore to Bush, it made sense to vote for Gore because Nader didn’t stand a chance. For this reason, my actual recommendation of who to vote for may not be the most electable candidate, and may change over time. I will make a recommendation and explain my reasoning in this part of the post, and will update it throughout the primary.


I’ve mostly avoided math in the main part of this post, so as to make my argument make sense as broadly as possible. If you’d like to see mathematical justification for some of the intuitions in this part, see Appendix C.


Recommendation as of March 1st: vote for Joe Biden on Super Tuesday (with possible exceptions in edge cases — if you want me to recommend how you should vote, based on what state and congressional district you live in, let me know!). [Confidence: high assuming Parts 1 and 2, medium overall.]

At this point, we basically have a two-way race between Biden and Sanders, with outside chances for other candidates. As of this writing, FiveThirtyEight forecasts a 61% chance that Sanders will end up with the most delegates, a 35% chance that Biden will, and a 3% chance that Bloomberg will. So your vote mostly just matters to the extent that it affects whether Biden or Sanders will be the nominee; the other candidates aren’t viable.

(Unconvinced? See some of the discussion regarding viability in the discussion of Iowa below and the math in Appendix C.)

Because Biden is likely more electable than Sanders, my general recommendation is to vote for Biden.

There are some corner cases in which I may recommend voting for someone else. California is a particularly interesting case, where Sanders is likely to win by a lot but Biden is likely to meet the 15% viability threshold, which means that perhaps the most effective tactical vote is (counterintuitively) to vote for Elizabeth Warren, so as to reduce Sanders’ delegate margin! If you’re a California voter, feel free to talk to me about this.


Recommendation as of January 31st: caucus for Joe Biden in the Iowa caucuses. [Confidence: high assuming Parts 1 and 2 are right; medium overall.]

The Iowa caucuses are wide open: as of January 31st, FiveThirtyEight estimates that Bernie Sanders has a 37% chance, Joe Biden a 35% chance, Pete Buttigieg 14%, Elizabeth Warren 10%, and Amy Klobuchar 3%.

Only 1% of pledged delegates come from Iowa, so Iowa matters not for its delegates but for the bounce in support that candidates get from outperforming expectations. In this particular race, a good approximation is that what matters in Iowa is who wins.

Here’s why: Buttigieg and Klobuchar — and to a lesser extent Warren — need to win Iowa to have a chance at being nominated. All three are far behind Biden and Sanders in the polls, and Iowa is one of Buttigieg’s and Klobuchar’s best states. On the other hand, Biden and Sanders both have high expectations, as they are the frontrunners (Biden nationally and Sanders in Iowa). The only way for them to beat expectations in Iowa is to win.

Now, suppose you decide to vote for Amy Klobuchar. How much likelier is she to win Iowa as a result?

Let’s consider a simpler question first: suppose you decide to vote for Michael Bennet. How much likelier is he to win Iowa as a result? The answer is (basically) not at all, since Bennet is nowhere close to winning Iowa.

Another question: in 2000, Al Gore ran away with the nomination. His only opponent, Bill Bradley, never stood a chance. How much of an effect would you have had by voting for Al Gore in Iowa in 2000? The answer (again) is very little, because Gore was very likely to win in any case.

Now say the Iowa caucuses are really close, like they were in 2016 between Clinton and Sanders. How much of an effect would your vote have had? The answer is quite a bit: in such a close race, the chance of your vote affecting the outcome is still small (because lots of people vote), but they are much, much larger than if the race is not close.

For this reason, the closer a candidate’s chances of winning Iowa are to 50%, the more voting for the candidate increases the chances they win.


So the answer to our question — how much of an effect will a vote for Klobuchar have in Iowa — is, not very much. I estimate that in terms of changing who wins Iowa, voting for Biden is about 8 times more effective than voting for Klobuchar. (Where did number come from? See the math in Appendix C!) Furthermore, according to the FiveThirtyEight forecast, even if Klobuchar wins Iowa, she only has a 15 or so percent chance of winning the nomination. (Biden’s chances at winning the nomination go up a similar amount if not more if he wins Iowa.)

In conclusion, although Amy Klobuchar is perhaps a little more electable than Joe Biden, this consideration is outweighed by the fact that a vote for Joe Biden is much more likely to actually make a difference in the outcome of the primary.

This could change after Iowa, if Klobuchar wins or Biden does really poorly. It could even change before Iowa, if some pollster puts out a really good poll for Klobuchar. But given the information available right now, I recommend caucusing for Joe Biden in Iowa.



Appendix A

How much can we discern about electability from head-to-head general election polls?

In an earlier blog post, I argued that a 10-point lead in polls a year ahead of an election would predict a 3 point victory on election day. Here’s a scatter plot of the GOP margin in presidential elections as predicted by the Republican candidate’s lead in head-to-head polls a year before the election (each data point is one election cycle).


(The slope of the line becomes 0.27 if you throw out the point on the left.)

Since we are now only 9 months from the election, I’d guess polls are more predictive: perhaps a 10-point lead in polls at this point predicts an eventual 4-point victory on average.

Now, this does not mean that if candidate A is doing 10 points better than candidate B in head-to-head polls against Trump (e.g. if polls show A beating Trump by 8% and B losing to Trump by 2%), then we should expect candidate A to hypothetically do 4 points better than candidate B in the general election. After all, our problem is different: we have data for how one Democratic candidate performed in each of 15 election cycles, but we have to figure out how predictive data for five Democratic candidates is for how they would hypothetically do in one (the same) general election cycle.

That said, I think that a 10-point difference in polls predicting a 4-point difference in the final result in our setting happens to be a pretty good approximation anyway. Why?

The main reason the first setting — the one with one data point per year — is different from ours is because there are non-candidate factors at play. For instance, the state of the economy has predictive power over general election outcomes: the incumbent party tends to do better in a good economy. On the other hand, in our setting (2020 head-to-head polls against Trump), the non-candidate factors are the same for all the Democratic candidates, because they’re running in the same cycle.

But it’s not clear whether that makes polls in the first setting (where non-candidate factors are relevant) more or less predictive of the eventual outcome. On the one hand, there’s more predictable variation: if each year, voters knew all the non-candidate factors perfectly one year ahead of the election, the predictive power of polls would for sure be much higher. On the other hand, voters don’t know these factors perfectly ahead of time, meaning that these factors add noise to the predictors, making them less predictive.

So, here’s a model: each election year y, the candidate who wins the primary of the non-incumbent party1 has a certain (unknown) “candidate quality” q_y (think of this as how much voters will end up liking the candidate when the general election comes around). Also, the race has a certain (unknown) “fundamentals” factor f_y (think of this as the state of the economy and other non-candidate factors at the time of the general election). Early in the year of the election, voters have a noisy signal for each of the two quantities: they know q_y + N_{q, y} and f_y + N_{f, y}, where N-terms are noise, and they respond to polls on the basis of these two quantities. When they actually vote, they vote based on q_y + f_y. Under this model, the predictiveness of early general election polls is essentially how well q_y + N_{q, y} + f_y + N_{f, y} (the predictor) predicts q_y + f_y (the responder).

The answer (under certain independence assumptions) is that a one percentage point difference in the predictor predicts a difference of

\frac{\text{var}(q) + \text{var}(f)}{\text{var}(q) + \text{var}(f) + \text{var}(N_q) + \text{var}(N_f)}

percentage points in the responder, where e.g. \text{var}(q) is the variance of candidate qualities (across years).

On the other hand, in our setting, we have candidates of different qualities (let’s denote the quality of candidate c with q_c) and some noise term for non-candidate factors, but that term is the same for different candidates. That means that a one percentage point difference between two Democratic candidates in 2020 head-to-head polls against Trump predicts a difference of

\frac{\text{var}(q)}{\text{var}(q) + \text{var}(N_q)}

percentage points in the eventual outcome of the hypothetical general election matchups.2

This means that in the first setting, the predictive power of polls is some weighted average of how much voters’ feelings toward candidates early in the year predicts how their feelings in November, and how much voters’ feelings about non-candidate factors like the economy early in the year predicts those feelings in November.

I don’t really have an opinion about which of the two is more predictive, so I’ll make the assumption that they’re about the same. Under this assumption, the predictive power of polls in the first setting is about the same as the predictive power of polls in the setting we actually care about. That’s why I’m fine assuming that a 10-point difference between how two candidates fare against Trump in early 2020 polls has the same predictive power (about 4 points) over the difference in the eventual hypothetical general election matchups, as a historical 10-point difference between two election cycles has over the eventual outcome of those races.


1. In years when no candidate is running for reelection, the quality of the candidate of the president’s party is another unnknown. On the one hand this adds more opportunity for predictive power, on the other hand more noise; I’d say it’s reasonable to ignore this consideration.

2. Implicit in this notation is that the variance in quality between the top candidates and the amount of noise is the same in this election as over historical elections; this seems like a fine guess.


Appendix B

Details on how I derived my electability scores

It’s natural to ask how I turned the qualitative descriptions on the left into the probabilities on the right. Here’s a summary of that process:

  1. I got the best numbers I could for each criterion-candidate pair. That was simple enough for general election polls: I took a polling average using the methodology described in Part 2. For moderation I used the FiveThirtyEight numbers I quoted in Part 2. For home popularity I unfortunately had to resort to different metrics for different candidates: approval numbers for the three current senators and election results for Biden and Buttigieg.
  2. I assigned weights to each criterion-candidate pair based on how informative that criterion is for that candidate. Most weights were 1; Biden and Buttigieg were given 0.25 weights for home state popularity (these are less informative for the reasons described in Part 2). Similarly, lesser known candidates were given lower weights for their polls against Trump.
  3. I turned the numbers in each category into Z-scores. That is, I calculated the answers to questions such as “how many standard deviations above the mean are Biden’s poll numbers?”. The mean I used was an average weighted by the weights discussed in the previous step.
  4. I then combined the Z-scores of each candidate into one score by taking the weighted average (in this average, the weight of a criterion-candidate pair was the weight discussed in step 2 times the weight of that criterion, i.e. 2 for polls and moderation and 1 for popularity).
  5. I multiplied these scores so that their spread was roughly 20. That reflects my opinion — elaborated upon below — that the most electable candidate has a roughly 20% higher chance of beating Trump than the least electable candidate. I gave Amy Klobuchar a 5% bonus corresponding to the not-yet-accounted-for positive “Other factors” assessment I gave her in Part 2 for being from a Midwestern swing state. I then shifted all the scores so that the net probability of beating Trump (weighted by the FiveThirtyEight probabilities of each candidate winning the nomination) was 48%. The 48% number is based on my overall assessment of the race (basically, Trump is very unpopular for an incumbent, but the economy is good).

The 20% number in step 5 came from the following reasoning:

  • Trump is an incumbent, and incumbents usually win reelection, especially when the economy is good. However, Trump is quite unpopular. His approval rating is (as of 2/2) 43.6%, which is five points lower than that of any post-World War II who’s been reelected on election day. (On the other hand, it’s at the high end for presidents who were not reelected.) Trump’s approval rating has been really stable so it’s unlikely to change much before election day. Taken together, a priori I’d say Trump being re-elected is a 50/50 proposition, or maybe slightly less likely than not.
  • Elizabeth Warren would be by far the most liberal nominee of the Democratic Party in recent history. She is unpopular in Massachusetts relative to a replacement Democrat and isn’t doing well in the polls against Trump. This isn’t great for her, but she’s not that unlikely to win if nominated: the election is likely to be a referendum on Trump, who’s somewhat unpopular. I’d assign her a 35-40% chance of winning.
  • Amy Klobuchar is a moderate candidate from the Midwest who is well-liked in her home state and will be running against an unpopular incumbent. A priori, Trump has an electoral college advantage, but that might be partially or completely negated by the fact that Klobuchar is from the Midwest and is very popular in her home state. I’d say she’s favored but not by a huge amount: 55-60% seems pretty fair.


Appendix C

A model for assessing the efficacy of a vote

Here’s a simple model for how your vote affects the candidates’ probabilities of winning: if you vote for candidate c, their probability of winning increases by some amount \Delta_c, and all other candidates’ probabilities of winning decrease in proportion to their current probabilities of winning. This doesn’t capture everything (in particular, helping a candidate likely hurts candidates in the same ideological lane more than ideologically dissimilar candidates), but I think it’s a reasonable model.

Let’s say candidate c has probability p_c of winning the nomination and e_c of winning the general election conditional on being nominated. Let e be the probability that a Democrat will win the general election. Then e is the average of the candidates’ electabilities weighted by their probabilities of being nominated, i.e.

e = \sum_{x \in C} e_xp_x

where C is the set of candidates. On the other hand, the probability of a Democrat winning the general election conditional on the Democrat nominated not being c (which we will denote e_{-c}) is

e_{-c} = \frac{\sum_{x \neq c} e_x p_x}{\sum_{x \neq c} p_x} = \frac{e - e_cp_c}{1 - p_c}.

Because of our model, this probability stays the same if you vote for candidate c (because the weights in the weighted average don’t change). This means that if you decide to vote for c, the probability that a Democrat is elected becomes

e_c(p_c + \Delta_c) + e_{-c}(1 - p_c - \Delta_c).

This means your effect on the probability that a Democrat wins is

(e_c - e_{-c})\Delta_c.

That is, by voting for c, you increase the likelihood of a Democratic win by (probability c wins if nominated minus probability the Democrat wins if c isn’t nominated) times how much your vote changes the probability that c is nominated. The \Delta_c term makes sense: if candidate c has no chance, you’re unlikely to change the eventual outcome by voting for c (so \Delta_c is small), and accordingly your vote doesn’t matter very much.

Now, how do we actually estimate \Delta_c? We need to model the fact that if a candidate is really unlikely to win — or almost guaranteed to win — a vote for that candidate isn’t all that likely to change the probability that they win, as compared to if the outcome is very uncertain. The best way I know of doing this is to think of a vote as changing the odds — rather than the probability — of a candidate winning by a constant factor. That is, if the odds of a candidate winning are p:1 - p, then an additional vote for the candidate changes those odds to (1 + \epsilon)p:1 - p for some small \epsilon independent of the candidate. This model is well-motivated by Bayes’ theorem.

Under this model, we get

\Delta_c = \frac{(1 + \epsilon)p_c}{(1 + \epsilon)p_c + 1 - p_c} - p_c = \frac{\epsilon p_c(1 - p_c)}{1 + \epsilon p_c} \approx \epsilon p_c(1 - p_c).

Therefore, \Delta_c is proportional to p_c(1 - p_c).

This means that it’s pretty reasonable to vote for the candidate c that maximizes p_c(1 - p_c)(e_c - e_{-c}).

How to the candidates fare by this metric? As of March 1st, using numbers from FiveThirtyEight’s forecast (assuming that the winner of the plurality of delegates will be the nominee) and normalizing so the highest number is 100, Biden scores 100, Bloomberg 8, Warren, Buttigieg, and Klobuchar near 0, and Sanders -107. A negative number here means that voting for this candidate in the primary decreases the chance of a Democrat winning in the general election. Note though that these numbers may change very quickly as the primary proceeds and candidates’ chances of being nominated change.

[As of January 31st, these numbers were Biden 100, Klobuchar 2, Buttigieg 1, Warren -29, and Sanders -73.]

The story is a bit more complicated for the first four states, because there what matters is outperforming expectations (to get a bounce in support) rather than accruing delegates. So what matters there is some combination of the probability of winning the early state and how much winning the early state changes the probability of being nominated. But at least so far — for Iowa — Joe Biden is clearly the right choice by my analysis.


7 thoughts on “Thoughts on the 2020 Democratic presidential primary

  1. I don’t ID as a Democrat or Republican nor do I have a strong background in statistics, but this is an interesting read. A few questions:

    1. Bloomberg jumped into the race late, but he’s still doing better than Klobuchar and Buttigieg nationally. How would you rate his electability?
    2. In addition, what would Hillary Clinton’s electability be at this point in 2016 under your model?
    3. Lastly, in your Part 2 analysis, shouldn’t scandals count in “Other Factors”? Like Clinton, Biden has his own scandals (Hunter, Anita Hill, segregation stances) on which Trump will apply the same ol’ tactics. His current head-to-head lead is good now, but his scandals are also currently not at the forefront, which can skew these results.


    1. Good questions!

      1. My personal assessment is that Bloomberg doesn’t have much of a chance to win the nomination. But to answer your question: I think Bloomberg is electable. He does about as well as Biden in the polls, and that’s before what ends up being about a 2-point adjustment in his favor because he’s less well-known. He is also moderate. If it starts looking like he might do well on Super Tuesday, I will strongly consider him.
      2. I looked at a few polls around this time. In an average of those polls, Sanders did about 2.5 points better than Clinton — and probably a little better than that after adjusting for the fact that Sanders was less well-known at the time. On the other hand, Clinton was more moderate (maybe about as moderate as Buttigieg?) so that would be a factor in her favor. I’d guess that this model would have assessed them to be about equally electable, or perhaps that Sanders would be a little more electable. I don’t really have an opinion on whether “Bernie would have won” in 2016, so I don’t have more to say about whether my model does well on 2016.
      3. Maybe. But Clinton *didn’t* really have scandals before the 2016 election. In particular, the emails thing is objectively not a big deal (lots of Trump administration officials have committed the exact same offense); it only became a big deal in the election because Trump found it politically useful to focus on it. I would imagine that Trump would employ the same tactic against anyone who ends up being the nominee. (For Biden, Burisma; for Warren, the whole Pocahontas thing; I’m not sure about Buttigieg and Klobuchar but I’m sure he’ll find something.) With regard to the other specific things you mentioned for Biden (Anita Hill and segregation), I think it’ll look awfully hypocritical for Trump to go after Biden on these things given his own record, and that he’ll opt for the Burisma thing instead.


  2. I read part 2, I thought it was interesting.

    my question is, do you think adjusting positively for poor name recognition is correct?

    my intuition is that at this stage poor name recognition is probably a bad sign. on which note, I think you’re overvaluing Klobuchar – I’d never heard of her, and had to look up the candidates to confirm she was running. I’d be curious to see pure name recognition numbers by candidate and party affiliation. anyway, on that note, why did you decide to profile Klobuchar over the other minor runners?

    (for reference, I think I’m registered democrat, but typically vote third party. I’ve never lived in a swing state but my intuition is that if I did I would still often vote third party.)


    1. I think that name recognition doesn’t let you draw any conclusions about electability, because whoever the Democratic nominee is *will* be known by nearly 100% of Americans by November (and lots of candidates that are likely very electable, at least according to conventional wisdom — such as Montana governor Steve Bullock — failed to gain traction; I don’t really see why to hold that against his electability in the general election).

      It’s true that Biden, Sanders, Warren, and Buttigieg are all substantially more likely to be the nominee than Klobuchar, but I’d say Klobuchar is fifth. She has an outside chance of winning the Iowa caucuses, and if she does then there’s not a bad chance that she’ll be the nominee. At the time I started writing this post, her chances looked better; perhaps if I started writing this post today I wouldn’t have included her.


  3. This was a really interesting read – thanks for the work you put into it! That being said, I have some disagreements that are potentially serious enough to undermine your conclusion. Here they are:

    1) In part 1b, you argue that the sort of policy we would get under a Democratic president can be approximated by what the median senator wants, with the caveat being issues where the president can act independent of Congress, like foreign policy. I think this model is potentially very bad, to the extent that its falseness might undermine your conclusion. Here are some reasons why:

    a) The president’s policies can have a strong causal effect on the policies that the median senator supports (if they are in the same party). An important feature of modern politics is that the president becomes the de facto leader of their party; if they push for a policy, then their party leaders and whips in Congress can effectively coordinate with the median senator, even if it’s not a policy that the median senator would ordinarily support. I don’t think this is a negligible effect — under the “median senator decides” model, I wouldn’t expect the Affordable Care Act to have been passed in 2010, and policy on the scale of ACA is something that a voter might care about *a lot*.

    b) I think you understate how much power the president has to act independent of Congress. Foreign policy isn’t the only thing is relevant here. It seems like all of your complaints about Trump’s environmental protection rollbacks are not things the median senator had a say in; Democratic nominees might differ quite a bit in how aggressively they reinstate those. More generally, by leading the federal bureaucracy, the president has quite a large say over how various laws and policies actually get enforced. Also, the president appoints positions like Supreme Court justices and the chairmen of the Fed, and I can imagine the people appointed by Biden for these positions differing quite a bit from those appointed by Bernie.

    c) The president affects political norms and trends. Based on your section about how Trump has set norms, this is clearly an important issue for you. Someone worried about the growing power of the executive might be worried about a president Sanders or Warren. That being said, the norms Trump has set seem far worse than anything the Democratic candidates would do, so I’ll mostly concede this point.

    d) The president moves the Overton window. One way that Trump has been bad is by bringing far-right ideas more into the mainstream of politics. However, I think it’s reasonable to think that a president Sanders would bring far-left ideas into the mainstream *to an even greater extent*. This connects to a worry that some people have that American politics has antibodies to nationalism, but maybe not to socialism. And if you have strong opinions about socialism (either positive or negative), this could be an issue more important to you than anything.

    e) The president affects the national political conversation in various fuzzy other ways that matter to people a lot. For example: for me, the most important issue is having a president that is not divisive. Clearly Trump is bad for this, but I could be persuaded that a president Warren might be worse. On the other hand, I could imagine someone thinking that having a female president is so important that it’s worth a 5% decrease in electability.

    2) I have various meta-disagreements that all feel related to me, so I’ll write them all here.

    a) While it’s definitely true that going through this analysis with made-up numbers provides positive information, I think that the sheer quantity of made-up numbers and the sensitivity of your model to small changes in those numbers might make the information gain very close to 0. This seems like the sort of model that would be very sensitive to errors in the numbers that you put in.

    b) I’m especially concerned about the relative weightings of the electability factors. How sensitive is your model to changing the 2:2:1 ratio? I’ve heard arguments that “although candidate X has good nationwide head-to-head numbers against Trump, that’s just because X is really popular in California, which doesn’t matter.” So if I think nationwide head-to-head polling isn’t predictive and I change your ratios to 1:2:2, will I get a completely different answer? What about 1:1:3?

    [As an aside, I would be **really** interested in seeing what your model’s sensitivity to these weights actually is! If it ends up that your conclusion *is* robust, then I will believe it much, much more. Is that something that’s easy for you to calculate?]

    c) I’m generally worried that there’s a “streetlight effect” going on here. [The streetlight effect is when you lose your keys on the street and start looking for them under a streetlight. Your friend asks “is that where you lost your keys?” and you reply “no, but this is where the light is best.”] You picked a few factors to measure electability by, most of which were quantitative and easy to measure. Then you gave the quantitative/measurable data higher weights and proceeded with your analysis. But what if these factors are actually less predictive than the fuzzy difficult-to-measure factors that people have in their head?

    Liked by 1 person

    1. Thanks for the really thoughtful critique! Here are my thoughts in response (not necessarily counterarguments because in some cases I at least partially concede your point).

      1. a) I think this is a good point. You’re right that the Democratic Party as a whole will end up further left by 2024 under a Sanders administration than under a Biden administration. I have some skepticism that the position of the *median* senator — even if they’re a Democrat — will change be all that different with Biden vs. Sanders as president. That’s because everyone from Biden to Sanders is way to the left of the best-case median senator (someone like Manchin) and there’s only so far Manchin can comfortably shift left. Will he shift more under Sanders than under Biden? Maybe, but that’s not obviously true to me. I can see it going both ways: he could be under more pressure to shift left if his party shifts left; or, he might break off from the Democratic establishment in reaction to Sanders becoming president, whereas he would be okay meeting Biden halfway. I don’t know.

      Another thing to consider is that some candidates (Warren, Buttigieg) want to end the filibuster whereas others (Biden) don’t and others (Sanders, Klobuchar) aren’t sure. I see arguments going both ways on whether we should end the filibuster (my guess is we shouldn’t) but I could see that as having big short-term and long-term consequences.

      b) That’s fair — perhaps I unfairly downplayed the independent power of the executive branch. My point there was that on such issues, Democrats are pretty ideologically similar. That is perhaps not true when it comes to trade (and monetary policy? I have no idea). But still, I’d say that overall they’re much more similar than different (compared to Trump).

      c) I agree with this. I also agree that the Democratic candidates differ in how they treat norms, but are all pretty similar when compared with Trump.

      d) Interesting — I haven’t heard the antibodies thing you’re saying. Could you elaborate or share a source? I’m curious where the US would have gotten antibodies for nationalism. (World War II? But then wouldn’t it have gotten anti-socialism antibodies during the Cold War?)

      e) This is fair. For what it’s worth, divisiveness already (anti)-correlates with electability, because if you alienate a lot of people, you’re likely not very electable. (Of course, this isn’t anywhere near a perfect measure: maybe there’s a candidate who conservatives really hate but independents really like, and those people are more electable than someone who is slightly disliked by both conservatives and independents.)

      I have some intuition on which Democratic candidates are divisive, but I don’t know to what extent I can trust it. My intuition on this also correlates a lot with my electability scores above, so the intuition may have come out of those scores.

      2. a) This is the only point you make that I disagree with. I think my conclusions are pretty robust to changing my made-up numbers around. Let’s take a look:

      – First, the 2:2:1 ratio. This is the thing that impacts my conclusions the most. See part b) below for a more detailed analysis, but the short version is: take a look at the chart and you’ll find that how the candidates do on the three metrics is very correlated (within a candidate, across metrics). Indeed, there are only two exceptions: Klobuchar not polling well against Trump (but we have some reason to discount that — she’s not well-known) and Sanders being very non-moderate. But overall things are stable.
      – Second, how much to discount measures that I’ve argued aren’t super relevant for particular candidates (like Klobuchar/electability and Buttigieg/home popularity). For the same reason — that the metrics are pretty correlated within a given candidate — I don’t think this matters that much.
      – Third, Klobuchar’s 5% increase for having a home state/Midwest advantage. Let’s make up some more numbers to justify this. There’s perhaps a ~50% chance that a non-Klobuchar nominee would lose to Trump. Conditioned on this, it’s basically guaranteed that Trump would win Wisconsin and perhaps Michigan and Pennsylvania. Conditioned on this, I think there’s a ~40% chance that he’d win these states just barely (within 2 percentage points, say). I think there’s a ~30% chance that Klobuchar’s Midwestern advantage would cover that gap. (A typical home-state advantage used to be 7 points of margin, though I’d expect that to have decreased in recent years, and I imagine a neighboring-state advantage is substantially lower; so maybe I’d expect her swing state advantage to be on the order of 1 or 2 points.) So that works out to a 7% increase in her probability of winning? Not too gar from my guess of 5%. I’m not sure if I should trust this number or the 5% more but I guess at this point Klobuchar doesn’t really have a chance so it doesn’t matter too much. (I tried to come up with these numbers without fudging them to match my 5% but I’m not sure I was successful.)
      – Fourth, the guess in Appendix A that the two ratios are the same. I don’t directly use that — instead it contributed to my decision on the 2:2:1 ratio — so I’d say this is subsumed by however much doubt you have about the 2:2:1 ratio.
      – Fifth, the “anchors” to my electability scores (i.e. that Trump is ~45% to win and the most electable candidate is ~20% more electable than the least electable one). The most important point here is that this doesn’t matter very much, because it doesn’t affect candidates’ [i]relative[/i] electabilities (i.e. regardless of what the numbers are, this gives a Klobuchar-Biden-Buttigieg-Sanders-Warren order). Trump being ~45% to win is in the ballpark of the consensus: the Nate Silvers of the world have generally said that this election is a tossup; PredictIt gives Republicans a 55% chance of winning which is somewhat higher than mine, but if you shift my numbers 10% that doesn’t affect my conclusions. As for the 20% number, that could be important because if you think the true number is like 5% then that hurts my argument that you should only care about electability. Here’s a further justification: a reasonable approximate prior is that the margin of this election will be uniform in [-5%, 5%]. I’d guess that it’s pretty reasonable to expect roughly a 2% difference in margin between the most and least electable candidates? Historically it would have been much higher, but we’re pretty polarized now.

      b) Great question! Here are probabilities for various weights, keeping everything else constant. (I’ve added a national poll since then so I’ll restate my 2:2:1 ratio electability percentages as well.)

      Polls weight/Moderation weight/Popularity weight: Klobuchar/Biden/Buttigieg/Sanders/Warren
      2:2:1: 64/62/56/47/42
      1:2:2: 66/61/57/48/41
      2:1:2: 64/60/55/49/40
      1:1:3: 67/59/56/50/42
      3:1:1: 61/61/54/49/41
      1:3:1: 66/62/58/46/42
      1:1:1: 65/61/56/48/41
      Infinity:1:1: 53/61/51/50/41
      1:Infinity:1: 66/63/60/43/45
      1:1:Infinity: 69/56/58/54/44

      c) There’s a lot to say here. For what it’s worth, what you call the streetlight effect is a totally reasonable way to search for your keys, right? So I think that analogy argues in favor of my approach?

      With regard to: “But what if these factors are actually less predictive than the fuzzy difficult-to-measure factors that people have in their head?” Maybe. So if there were a conventional wisdom about who is more electable, perhaps that should be given some credence. But as far as I know, there isn’t? Everyone seems to just think that the candidates they like the most are the most electable. And maybe you have some fuzzy intuitions about who is electable, but maybe your friend has different intuitions, and why should you trust your intuitions over your friend’s? If you come up with some way to turn those intuitions into something objective, I’d love to hear it, but I’d strongly hesitate before giving credence to a subjective measure that everyone disagrees on.

      To preempt a possible suggestion that I wouldn’t endorse: you could look at betting markets to find conditional probabilities of candidates winning (i.e. Pr[win election]/Pr[nominated]). I think this is bad first because betting markets are often wildly mispriced (I sold Yang at 16% at some point in 2019) and second because I’ve looked at these conditional probabilities on occasion and they’ve jumped around *a ton* day to day. (It’s also not a sensitive instrument. Biden is 12% to win the nomination on PredictIt and 5% to be elected president. That’s 42%. If someone decides to Biden for 6%, that’ll go up to 50%; if the price drops to 4% that’ll be 33%. The fact that PredictIt trades in cents, not finer units, limits our ability to get a precise measurement.)

      Anyway, thanks a bunch for this comment, because these are exactly the sorts of comments that actually help us come to the right answer: specific, clearly stated objections to specific points made in the spirit of truth-seeking. Would love to continue this back-and-forth! (Though I have a paper deadline on the 12th, so I might wait until after then to respond.)


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s