A few days after the United States presidential election, Data Driven Journalism featured a great piece about the pitfalls of relying on opinion polls to predict elections: “Will data driven election reporting ever be the same?” I myself had fallen into the trap of trusting the polls that predicted a landslide Clinton win, so in the aftermath of the unexpected election results, I started thinking about alternatives to traditional polls for predicting election results. Endorsements, perhaps. Maybe if you looked at the number of endorsements by politicians in a particular state, you’d be able to predict the outcome in that state.
Luckily, I happened to have that data on hand.
In the weeks leading up to the election, I built EndorsementDB.com, a database for tracking public endorsements (and un-endorsements) of presidential candidates by notable figures. The goal was to store endorsement metadata in a structured but flexible way, allowing for arbitrary tags and thus facilitating open-ended queries. I started by manually adding endorsements as I encountered them in the news, tweaking the schema as necessary to accommodate increasingly complex types of endorsements: “LeBron James endorses Hillary”; “The Chicago Tribune endorses Gary Johnson”; “Republicans Endorse, Unendorse And Then Re-Endorse Donald Trump”.
Although I managed to import several hundred endorsements manually, I soon realised that I had severely underestimated the sheer volume of endorsements – over 8000 in total – that had accumulated throughout the campaign season. So I changed my strategy: I would add endorsements by parsing the relevant Wikipedia pages and importing them programmatically, with manual supervision as needed. That dramatically sped up the process, allowing me to import all the endorsements from the relevant Wikipedia pages. Once those were imported, I programmatically tagged what I could based on the context on the Wikipedia pages, and then manually added in some other tags that I thought could be relevant (for example, all the members of Congress, complete with their party affiliation and state).
Endorsements among members of Congress
Now, going back to the original premise: How could we have predicted the outcome of the election without using polls? Well, given that we have a database of the endorsements made by all the members of Congress, we can start there and see how these endorsements align with the popular vote in each state.
As a side note: since this analysis is being done after the election is over, the goal of this analysis is not to try and dazzle you with my clairvoyance. (The design of these models was indisputably tainted by my knowledge of the actual election results, with the risk of overfitting and confirmation bias that entails.) Rather, the aim of this post is to explore the correlation between endorsements and election results – a correlation that I’m sure many would suspect, but might find difficult to quantify. To that end, I tried to keep the models straightforward and reasonable, with clear explanations of how I produced each of them.
For simplicity, these models all treat Nebraska and Maine as winner-take-all states even though they actually use a district-based model.
Party affiliation as a baseline
We start by ignoring endorsements and only looking at party affiliation as a baseline model. For each state, we look at the party breakdown of its members of Congress, and award that state to Trump if most members are Republican, and to Clinton if most are Democrat. If there is an equal number from each party, we treat the prediction as a tie; if the winning party has only a small edge percentage-wise, we predict a smaller margin of victory.
The results are shown in the table below. Clinton victories are shown in blue, and Trump victories are shown in red, with a white background to indicate a small margin of victory (under 5%). The numbers after the first row of the “Predicted electoral votes” column indicate the difference between the two parties, with the actual number of electoral votes awarded to each state shown in the first column.
This model, which projects a landslide Trump victory of 337 to 197 electoral votes, performs decently well: of the 50 states and D.C., only 5 predictions are wrong, and they all happen to be the states that were won narrowly, with less than 5% spread (Colorado, Maine, Nevada, New Hampshire, Virginia).
Of course, the problem with looking solely at party affiliation is that it produces a very limited model: it implies that control of Congress will lead to winning the election, which we know isn’t always the case. Plus, it implies that sitting members of Congress – some of whom were elected six years ago – are, for whatever reason, an accurate representation of voting preferences today. What’s more, we know that people don’t always vote along party lines, a fact that is especially relevant for this election, as many Republicans decided not to endorse their Party’s nominee.
Using endorsements instead
Bearing that in mind, we turn our attention instead to endorsements. For this model, we award a state to the candidate with the most endorsements among members of Congress. For simplicity, we ignore declarations of support that are explicitly stated to not be endorsements (a queer phenomenon that appears to be specific to Trump; you can see a list of such endorsements at endorsementdb.com), as well as those of the form “I would support the Republican ticket if Trump were to withdraw and be replaced by Pence”. (It would be interesting to account for these pseudo-endorsements in another analysis, however.)
This time, only three predictions were wrong: Michigan (a tie), Virginia (small Trump margin), and Wisconsin (a tie). The result is still a Trump victory – even ignoring Wisconsin and Michigan – but with a smaller margin: 292 to 220.
This model is more reasonable in its implications – namely, that the feelings of members of Congress towards the presidential candidates can influence, or otherwise reflect, the feelings of their constituents – and the predictions are a definite improvement over the baseline model. However, the presence of ties is suboptimal, since we wouldn’t expect any ties in the actual vote counts. Furthermore, this model treats members of the Senate and the House equally, even though each state has only two members of the Senate (who are, typically, elected every six years) but up to 53 members of the House (who are elected every two years). If we want better precision, we’ll have to find a reasonable way of breaking those ties, taking this difference into account.
To that end, I came up with three ways of breaking ties. All of these models made exactly one false prediction, and correctly predicted the popular vote spread (whether it’s less than or greater than 5%) for at least 41 of the 51 states/districts. Two of the models incorrectly predict Virginia for Trump; the other model predicts Michigan for Clinton (due to both sitting Senators being Democrat).
Note that all three models predict a comfortable Trump victory.
If I had done this analysis before the election, I certainly would have been less quick to believe that Clinton would have won in a landslide, as many polls predicted.
Of course, I say this with the privilege of hindsight. If she had won, would I still have performed this analysis afterward? Would I have discarded the results when they indicated a Trump victory? Maybe; I don’t know. At the same time, it may be that these endorsements form an important bellwether for the election results, meaning that Clinton wouldn’t have had a chance unless more Republicans in Congress had decided to endorse her over third-party candidates like Evan McMullin and Gary Johnson or a Republican other than the nominee.
Incidentally, I also learned that newspaper endorsements were not a great predictor of the outcome of this election. This would be obvious to anyone who is aware that the vast majority of newspaper endorsements were for Clinton, with only a handful for Trump. If you focus on the number of local newspapers endorsements by state, you’ll find that Clinton is projected to win every single state except Alaska and Kansas, which is pretty hard to imagine. Even if we change the model to only award a state to Clinton if Trump gets no newspaper endorsements – which, to be honest, is a pretty contrived rule – we still get 20 mispredictions. Clearly, newspaper endorsements were not a great predictor of how voters felt during this election. Why that was the case is anyone’s guess.
While I believe these methods would be a useful part of a pollster’s arsenal for future elections, I do acknowledge some caveats. Given that we have only one event - the 2016 election - to draw from, there is a real risk of overfitting in these results. All I’ve shown is that for the 2016 election, there is a correlation between presidential endorsements by members of Congress and the popular vote in each state, which is slightly stronger than that between their party affiliation and the popular vote. It’s possible that is only true because of the unique circumstances of this election year: Congress was overwhelmingly under Republican control, and the Republican nominee was a uniquely polarising figure who, throughout the course of his campaign, lost the endorsement of many Republican members of Congress. It would be interesting to apply these models to previous elections to see how they hold up, but unfortunately historical endorsement data is more challenging to obtain, especially pre-2008.
It’s also unclear why, exactly, this correlation exists. Do members of Congress make their decisions based on who they think their electorate would support? Do voters change their mind based on who their representatives support? Is there something else – something geographical, or cultural – that just happens to align these two populations most of the time? Is it some combination of all three? On this, I can only speculate.
Still, the fact that even the most basic endorsement-based models did decent job of predicting the election speaks to the importance of endorsements. I personally think that such endorsements should be tracked, for accountability purposes (especially when it comes to public officials) and because of their potential as a barometer for public opinion (partly because these people often have influence, and partly because it can indicate the sentiment among their audience). Endorsements are often overlooked in statistical analysis because of the difficulty of obtaining the necessary data, which is why I will continue to work on EndorsementDB.com with the goal of making it a central source of endorsement data for this and future elections.
You can view the entirety of the tables included in this post at http://endorsementdb.com/stats/predictions, or browse the 8000+ endorsements at EndorsementDB.com. The source is on GitHub (along with the data, which I really shouldn’t be storing in Git given the size of the files) at https://github.com/endorsementdb/endorsementdb.com.
If you find this interesting and want to get involved in any way, please reach out by email or Twitter! I’d love to hear your thoughts or help you with a research question that involves endorsement data. I have some more research questions of my own – like the effects of presidential endorsements on Senate and House races – and I’ll be writing up a post if I find anything interesting, so stay tuned.
This post was featured on Data Driven Journalism as “EndorsementDB: Predicting the U.S. election without polls”.