New year, new election.
This time around, May's election will have the Mayoral race at the top of the ballot, and City Council races. Odd-year primaries also bring Philadelphia’s lesser-followed judicial elections. These elections are problematic: we vote for multiple candidates each election, upwards of 7, and even the most educated voter really doesn't know anything about any of them. In these low-information elections, the most important factor in getting elected is whether a candidate ends up in the first column of the ballot. Ballot position is decided by a lottery. Philadelphia elects its judges at random.
I’ve looked into this before, measuring the impact of ballot position on the Court of Common Pleas, and then that impact by neighborhood. We even ran an experiment to see if we could improve the impact of the Philadelphia Bar's Recommendations. But first, let’s revisit the story, update it with 2017 results, and establish the gruesome baseline.
The Court of Common Pleas
The most egregious race is for the city’s Court of Common Pleas. That court is responsible for major civil and criminal trials, juvenile and domestic relations, and orphans, so the bulk of our judicial cases. Common Pleas judges are elected to 10-year terms.
The field for Common Pleas is huge; we don’t know how many openings there will be this year, but in the five elections since 2009, there was an average of 31 candidates competing for 9 openings. Even the most informed voter doesn’t know the name of most of these candidates when she enters the voting booth, and ends up either relying on others' recommendations or voting at random. This is exactly the type of election where other factors--flyers handed out in front of a polling place, the order of names on the ballot--will dominate.
Because of this low attention, some of the judges that do end up winning come with serious allegations against them. In 2015, Scott Diclaudio won; months later he would be censured for misconduct and negligence, and then get caught having given illegal donations in the DA Seth Williams probe. The 2015 race also elected Judge Lyris Younge, who has since been removed from Family Court for violating family rights, and just in October made headlines by evicting a full building of residents with less than a week’s notice. In 2016, Mark Cohen was voted out as state rep after reports on his excessive use of per-diem expenses. In 2017, he was elected to the Court of Common Pleas.
Every one of those candidates—Diclaudio, Younge, and Cohen—was in the first column of the ballot.
Random luck is more important than Democratic Endorsement or Quality
The order of names on the ballot is entirely random, so this gives us a rare chance to causally estimate the importance of order, compared to other factors.
First, what do I mean by ballot order? Here's a picture of the 2017 Democratic ballot for the Court of Common Pleas.
There are so many candidates that they need to be arranged in a rectangle. In 2017, candidates were organized into 11 rows and 3 columns. Six of the nine winners came from the first column. (The empty spaces are candidates who dropped out of the race after seeing their lottery pick.)
The shape of the ballot changes wildly over years. Sometimes it’s short and wide, in 2017 it was tall and thin. But in every year, candidates in the first column fare better than others. The results year over year show that the first column consistently receives more votes:
As the above makes clear, candidates do win from later columns every year. But first-column candidates win more than twice as often as if all columns won equally. Below, I collapse the columns for each year, and calculate the number of actual winners from each column, compared to the expected winners if winners were completely at random. The 2011 election marked the high-water mark, when more than three times as many winners came from the first column than should have.
This effect means that Philadelphia ends up with unqualified judges. Let's use recommendations by the Philadelphia Bar Association as a measure of candidate quality. The Bar Association's Judicial Commission does an in-depth review of candidates each cycle, interviewing candidates and people who have worked with them. They then rate candidates Recommended or Not Recommended. A Bar Association Recommendation should be considered a low bar; on average, they Recommend two thirds of candidates and more than twice as many candidates as are able to win in a given year. While the Bar Association maintains confidentiality about why they Recommend or don't Recommend a given candidate (giving candidates wiggle room to complain about its decisions), my understanding is that there have to be serious concerns to be Not Recommended: a history of corruption, gross incompetence, etc.
In short, it's worrisome when Not Recommended candidates win. But win they do. In the last two elections, a total of six Not Recommended candidates won. They were all from the first column.
Note that, from a causal point, this analysis could be done. Ballot position is randomized, so there is no need to control for other factors. The effect we see has the benefit of combining the various ways ballot position might matter: clearly because voters are more likely to push buttons in the first column, but also perhaps if endorsers take ballot position into account, or candidates campaign harder once they see their lottery result. This analysis is uncommonly clean, and we could be done.
But clearly other factors affect who wins, too. How does the strength of ballot position compare to those others?
Let’s compare the strength of ballot position to three other explanatory features: endorsement by the Democratic City Committee, recommendation by the Philadelphia Bar Association, and endorsement by the Philadelphia Inquirer. (There are two other obvious ones, Ward-level endorsements and money spent. I don't currently have data on that, but maybe I'll get around to it!)
I regress the log of each candidate's total votes on their ballot position (being in the first, second, or third column, versus later ones, and being in the first or second row), endorsements from the Democratic City Committee and the Inquirer, and Recommendation by the Philadelphia Bar, using year fixed effects.
Being in the first column nearly triples your votes versus being in the fourth or later. The second largest effect is the Democratic City Committee endorsement, which doubles your votes versus not having it. (Notice that the DCC correlation isn't plausibly causal, the Committee certainly takes other factors into its consideration. Perhaps this correlation is absorbing candidate's abilities to raise money or campaign. And it almost certainly is absorbing some of the first-column effect: the DCC is known to take ballot position into consideration!)
The Philadelphia Bar Association's correlation is sadly indistinguishable from none at all, but with a *big* caveat: the Philadelphia Inquirer has begun simply adopting the Bar's recommendations as its own, so that Inquirer correlation represents entirely Recommended candidates.
While this analysis doesn't include candidate fundraising or Ward endorsements, those certainly matter, and would probably rank high on this list if we had the data. Two years ago, I found tentative evidence that Ward endorsements may matter even more than the first column. And Ward endorsements, in turn, cost money.
What are the solutions?
There are obvious solutions to this. I personally think holding elections for positions that no-one pays attention to is an invitation for bad results. It's better to have those positions be appointed by higher profile officials; it concentrates power in fewer hands, but at least the Mayor or Council know that voters are watching.
A more plausible solution than eliminating judicial elections altogether might be to just randomize ballot positions between divisions, rather than use the same ones across the city. That would mean that all candidates would get the first-column benefit in some divisions and not in others, and would wash away its effect, allowing other factors to determine who wins.
Unfortunately, while these are easy in theory, any changes would require a change to the Pennsylvania Election Code, and may be a hard haul. But one thing everyone seems to agree on is that our current system is bad.
The turnout in November was unprecedented. Nationally, we had the highest midterm turnout since the passage of the Voting Rights Act. In Philadelphia some 53% of registered voters voted a huge increase over 4 years ago, or really any midterm election in recent memory. Turnout can be hard to wrap your mind around, with a lot of factors that affect who votes and in what elections. What drives those numbers? Is it registration deviations? Or variations in the profile of the election? And how is that structured by neighborhood?
Decomposing the Funnel
I'm conceptualizing the entire process that filters down the whole population as a funnel, with four stages at which a fraction of the population drops out.
Let’s break down the funnel using the following accounting equation:
Voters per mile =
(Population over 18 per mile) x
(Proportion of over-18 population that are citizens) x
(Proportion of citizens that are registered to vote) x
(Proportion of registered voters that voted in 2016) x
(Proportion of those who voted in 2016 that voted in 2018).
The implications are very different if a neighborhood lags in a given one of these steps. If a ward has a low proportion of citizens that are registered, registration drives make sense. If a neighborhood has very low turnout in midterms vs presidential elections, then awareness and motivation is what matters.
We are also going to see that using metrics based on Registered Voters in Philadelphia is flawed. The lack of removal of voters from the rolls—which is actually a good practice, since the alternative is to aggressively remove potential voters—means that the proportion of registered voters that vote is not a useful metric.
Let’s walk through the equation above, map by map.
Overall, 44% of Philadelphia's adults voted in 2018. That represented high turnout (measured as a proportion of adults) from the usual suspects: the ring around Center City, Chestnut Hill, Mount Airy, and the Oak Lanes.
How does the funnel decompose these difference?
First, consider something not in that map: population density. This is obviously, almost definitionally, important. Wards with more population will have more voters.
The ring around Center City, North Philly, and the lower Northeast have much higher densities than the rest of the city. [Note: See the appendix for a discussion of how I mapped Census data to Wards]
We don’t allow all of those adults to vote. I don’t have data on explicit voter eligibility, but we can at least look at citizenship.
The city as a whole is 92% citizens, and some 53 of the city's 66 wards are more than 90%. Notable exceptions include University City, and immigrant destinations in South Philly and the lower Northeast.
The nest step in the funnel: what fraction of those citizens are registered to vote? Here's where things get hairy.
Overall in the city, there are 93% as many registered voters as there are adult citizens. That's high, and a surprising 22 wards have *more registered voters than the census thinks there are adult citizens*. What's up? Philadelphia is very slow to remove voters from the rolls, and this imbalance likely represents people who have moved, but haven't been removed.
While people tend to use facts like this to suggest conspiracy theories, Philadelphia's ratio is actually quite typical for the state, across wealthy and poor, Republican and Democratic counties. [See below!]
And it's very (very) important to point out that this is good for democracy: being too aggressive in removing voters means disenfranchising people who haven't moved. And we have no evidence that anybody who has moved is actually *voting*, just that their name remains in the database.
It does, however, mean that measuring turnout as a fraction of registered voters is misleading.
I explore the registered voter question below, but for the time being let's remove the registration process from the equation above to sidestep the issue:
Voters per mile =
(Population over 18 per mile) x
(Proportion of Over 18 pop that are citizens) x
(Proportion of citizens that voted in 2016) x
(Proportion of those who voted in 2016 that voted in 2018).
Why break out 2016 voters first, then 2018? Because there are deep differences in the processes that lead people to voting in Presidential elections versus midterms, and low participation in 2016 and 2018 have different solutions. Presidential elections are obviously high-attention, high-energy affairs. If you didn't vote in 2016, didn't turn out in the highest-profile, biggest-budget election of the cycle, you either have steep barriers to voting (time constraints, bureaucratic blockers, awareness), or are seriously disengaged. If someone didn't vote in 2016, it's hard to imagine you'd get them to vote in 2018.
Compare that to people who voted in 2016 but not in 2018. These voters (a) are all registered, (b) are able to get to the polling place, and (c) know where their polling place is. This group is potentially easier to get to vote in a midterm: they're either unaware of or uninterested in the lower-profile races of the midterm, or have made the calculated decision to skip it. Whichever the reason, it seems a lot easier to get presidential voters to vote in a midterm than to get the citizens who didn't even vote in the Presidential race.
So which citizens voted in 2016?
Overall, 63% of Philadelphia's adult citizens voted. This percentage has large variation across wards, ranging from the low 50s along the river, to 81% in Grad Hospital's Ward 30 [Scroll to the bottom for a map with Ward numbers]. Wards 28 and 47 in Strawberry Mansion and North Philly have high percentages here, and also had high percentages in the prior map of registered voters, which to me suggests a combination of high registration rates in the neighborhood and a low population estimate from the ACS (see the discussion of registered voters below).
How many of these 2016 voters came out again two years later?
This map is telling. While there were fine-grained changes in the map versus 2014's midterm, the overall pattern of who votes in midterms didn't fundamentally change: it's the predominantly White neighborhoods in Center City, its neighbors, and Chestnut Hill and Mount Airy, coupled with some high-turnout predominantly Black Wards in Overbrook, Wynnefield, and West Oak Lane.
The dark spot is the predominantly Hispanic section of North Philly. It's not entirely surprising that this region would have low turnout in a midterm, but remember that this map has as its denominator *people who voted in 2016*! So there's even disproportionate fall-off among people who voted just two years ago.
So that's the funnel. We have some small variations in citizenship, confusing variations in proportions registered, steep differences in who voted in 2016, and then severely class- and race-based differences in who comes out in the midterm. Chestnut Hill and Center City have high scores on basically all of these dimensions, leading to their electoral dominance (except for Chestnut Hill's low population density and Center City's relatively high proportion of non-citizens). Upper North Philly and the Lower Northeast are handcuffed at each stage, with telling correlations between non-citizen residents and later voting patterns, even those which could have been unrelated, such as the turnout among citizens.
What's going on with the Registered Voters?
It's odd to see more registered voters than the Census claims there are adult citizens. I've claimed this is probably due to failing to remove from the the rolls people who move, so let's look at some evidence.
First, let's consider the problem of uncertainty in the population estimates. The American Community Survey is a random sample, meaning the population counts have uncertainty. I've used the 5-year estimates to minimize that uncertainty, but some still exists. The median margin of error in the population estimates among the wards is +/- 7%. This uncertainty is larger for less populous wards: the margin of error for Ward 28's population is +/-12%. So the high registration-to-population ratios may be partially driven by an unlucky sample giving a small population estimate. But the size of the uncertainty isn't enough to explain the values that are well above one, and even if it were large enough to bring the ratios just below one, the idea that nearly 100% of the population would be registered would still be implausibly high.
So instead, let's look at evidence that these high ratios might be systematic, and due to lags in removal. First, consider renters.
Wards with more renters have higher registration rates in the population, including many rates over 100%. On the one hand, this is surprising because we actually expect renters to be *less* likely to vote; they are often newer to their home and to the city, and less invested in local politics. On the other hand, they have higher turnover, so any lag in removing voters will mean more people are registered at a given address. The visible positive correlation suggests that the second is so strong as to even overcome the first.
(For reference, here's a map of Ward numbers)
There's an even stronger signal than renters, though. Consider the Wards that lost population between 2010 and 2016, according to the Census.
The group of Wards that saw the greatest declines in population also have the highest proportions of registered voters (except for 65, which is a drastic outlier). This correlation suggests that the leaving residents may not be removed from the rolls, further pointing the finger at that as the culprit.
Finally, let's look at Philadelphia versus *the rest of the state*. It turns out that Philadelphia's registered voter to adult citizen ratio is typical of the state, especially when you consider its high turnover.
First, there isn't a strong correlation between Registered Voters per adult citizen and Democratic counties.
Allegheny and Philadelphia, to two most Democratic counties are 1st and 6th, respectively, but predominantly-Republican Pike County, outside of Scranton, is second, and the rest of Philadelphia's wealthier suburbs are 3, 4, 5, and 7. (I'm not totally sure what's going on with Forest County, at the bottom of the plot, but the next scatterplot helps a little bit).
More telling is a county's fraction of the residents that have moved in since 2000: it looks like counties with higher turnover has higher ratios, which we would expect if the culprit were lagging removal.
Philadelphia here looks entirely normal: counties with more recent arrivals have higher ratios. This also sheds some light on Forest County. With a population of only 7,000, it's an outlier in terms of long-term residents, and thus has a *very* low registration-to-citizen ratio.
Apologies for the long explanation. I didn't want to just ignore this finding, but I'm terrified it'll be discovered and used by some conspiracy theorist.
Exactly where the voters fall out of the funnel matters
Philadelphia's diverse wards have a diverse array of turnout challenges. Unfortunately, the voter registration rolls are pretty unhelpful as a signal, at least in the simplistic aggregate way I considered them here. (Again: good for democracy, bad for blogging).
Which stage of the funnel matters depends on two things: their respective sizes, and how easy it is to move each. Is it more plausible to get out the citizens who didn't vote in 2016? Or is it more plausible to get the 2016 voters out again in 2018? Where will registration drives help, and where is that not the issue?
Next year's mayoral primary, a local election with an incumbent mayor, will likely be an exaggerated version of the midterm, with even more significant fall-off than we saw in November. More on that later.
Appendix: Merging Census data with Wards
I use population data from the Census Bureau's American Community Survey, an annual sample of the U.S. population with an extensive questionnaire on demographics and housing. Because it's a sample, I use the five-year aggregate collected from 2012-2016. This will minimize the sampling uncertainty in the estimates, but mean there could be some bias in regions where the population has been changing since that period.
The Census collects data in its own regions: block groups and tracts. These don't line up with Wards, so I've built a crosswalk from census geographies to political ones:
- I map the smallest geography--Census Blocks--to the political wards.
- Using that mapping, I calculate the fraction of each block group's 2010 population in each Ward. I use 2010 because blocks are only released with the Decennial Census. Largely, any changes over time in block populations are swamped by across-block variations in populations that are consistent through time.
- I then apportion each block group's ACS counts--of adults, of citizens--according to those proportions.
This method is imperfect; it will be off wherever block groups are split by a ward and the block group is internally segregated, but I think the best available approach.
Before November's election, I published a prediction of the Pennsylvania General Assembly Lower House: Democrats would win on average 101.5 seats (half of the 203), and were slight favorites (53%) to win the majority. Then I found a bug, and published an updated prediction: Democrats would win 96 seats on average (95% confidence interval from 88 to 104), and had only a 13% chance of taking the house. This prediction, while still giving Republicans the majority, represented an average 14 seat pickup for Democrats from 2016.
That prediction ended up being correct: Democrats currently have the lead in 93 seats, right in the meat of the distribution.
But as I dug into the predictions, seat-by-seat, it looked like there were a number of seats that I got wildly wrong. And I've finally decided that the model had another bug; probably one that I introduced in the fix to the first one. In this post I'll outline what it was, and what I think the high-level modelling lessons are for next time.
Where the model did well, and where it didn't
The mistake ended up being in the exact opposite of the fix I implemented: candidates in races with no incumbents.
The State House has 203 seats. This year, there were 23 uncontested Republicans and 55 uncontested Democrats. I got all of them right 😎. The party imbalance in uncontested seats was actually unprecedented in at least the last two decades, they're usually about even. Among the 125 contested races, 21 had a Democratic incumbent, 76 a Republican incumbent, and 28 no incumbent. I was a little bit worried this new imbalance meant that the process of choosing which seats to contest had changed, and that past results in contested races would be different from this year. Perhaps Democrats were contesting harder seats, and would win a lower percentage of them than in the past. That didn't prove true.
The aggregate relationship between my predictions and the results look good. I got the number of incumbents that would lose, both Democrat and Republican right. In races without incumbents, I predicted an even split of 14 Democrats and 14 Republicans, and the final result of 17 R - 11 D might seem... ok? Within the range of error, as we saw in the histogram above. But the scatterplot tells a different story.
Above is a scatterplot of my predicted outcome (the X axis) versus the actual outcome (the Y axis). Perfect predictions would be along the 45 degree line. Points are colored by the party of the incumbent (if there was one). The blue dots and the red dots look fine; my model actually expected any given point to be centered around this line, plus/minus 20 points, so that distribution was exactly expected. But black dots look horribly wrong. Among those 28 races without incumbents, I predicted all but three to be close contests, and missed many of them wildly. (That top black dot, for example, was Philadelphia's 181, where I predicted Malcolm Kenyatta (D) slightly losing to Milton Street (R). That was wrong, to put it mildly. Kenyatta won 95% of the vote.)
What happened? My new model specification, done in haste to fix the first bug, imposed a faulty logic. It forced the past Presidential races to carry the same information about races without incumbents as races with, even though races with incumbents had other information. I should have allowed the model to fall back on district partisanship when it didn't have incumbent results, but the equivalent of a modelling typo didn't allow that. Instead, all of these predictions ended up at a bland, basically even race, because the model couldn't use the right information to differentiate them. My overall house prediction ended up being good only because a few 28 of the total 203 districts were affected, and getting three too many wrong didn't make the topline look too bad. But it was a bug.
I'm new to this world of publishing predictive models based on limited datasets and with severe time constraints (I can't spend the months on a model that I would in grad school or at work). What are the lessons of how to build useful models under these constraints?
Lesson 1: Go through every single prediction. I never looked at District 181. If I had seen that prediction, I would have realized something was terribly wrong. Instead, I looked at the aggregate predictions (similar to the table, and things looked okay enough). Next time, I'll force myself to go through every single prediction (or a large enough sample of predictions if there are too many). When I tried to hand-pick sanity checks based on my gut, I happened to not choose "a race with no incumbents, but which had an incumbent for decades, and which voted for Clinton at over 85%".
Lesson 2: Prefer clarity of the model's calculations over flexibility. I fell into the trap of trying to specify the full model in a single linear form. Through generous use of interactions, I thought I would allow the model flexibility for it to identify different relationships between historic presidential races and length of incumbency. This would have been correct, if I had implemented it bug-free. But I happened to leave out an important three-way interaction. If I had fit separate models for different classes of races--perhaps allowing the estimates to be correlated across models--I would have immediately noticed the differences.
Lesson 2b: I actually learned this extension to Rule 2 in the process of fitting, but the post-hoc assessment bangs it home: when you have good information, the model can be quite simple. In this case, the final predictions did well even with the bug because the aggregate result was really pretty easy to predict given three valuable pieces of information: (a) incumbency, (b) past state house results, and (c) FiveThirtyEight's Congressional predictions. The last was vital: it's a high-quality signal about the overall sentiment of this year's election, which is the biggest factor after each district's partisanship. A model that used only these three data points and then correctly estimated districts' correlated errors around those trends would have gotten this election spot on.
Predictions will be back
For better or worse, I've convinced myself that this project is actually possible to do well, and I'm going to take another stab at it in upcoming elections. First up is May's Court of Common Pleas election. These judges are easy to predict: nobody knows anything about the candidates, so you can nail it with just structural factors. More on that later!
On November 6, 2018, we (i.e. all of you readers out there) collectively generated 1,341 turnout data points to the Sixty-Six Wards Live Turnout Tracker. That made my modelling effort pretty easy: we missed the true turnout of 548,000 by only 7,000 votes! (This is almost too good… More thoughts on that in a later post.)
One novel feature of this dataset is that we can look at when through the day Philadelphians voted. I did this back in the Primary, and my surprising finding was that turnout was relatively flat. While there were before- and after-work surges, they weren’t nearly as pronounced as I had expected.
That was a midterm primary. How about this time, in the record-setting general? Two weeks ago we saw an unprecedented, energized electorate, and it looks like their time-of-day voting pattern was different too.
Estimating the overall pattern requires some statistics. Each data submission contains three pieces of information: precinct, time of day, and the cumulative votes at that precinct. Here, I fit something different than my live election-day model, since I have the true final results. I assume that between hours h-1 and h, a fraction f(h) of the city’s population votes, and each precinct as a fraction that randomly deviates from that distribution.
For division d, the fraction of the final votes that have been achieved by hour h is given by
v_d(0) = 0
v_d(h) = v_d(h-1) + f(h) + f_d(h)
where f(h) is the city-wide precinct average fraction (I use an improper, flat prior), and the deviations are drawn from a normal distribution with an unknown standard deviation sigma_dev.
Because observations don’t occur only on the hour, I use linear interpolation in between, so a data submission x_i, which occurs in division d at time h + t with t between 0 and 1, is modeled as
x_i ~ student_t(1, mean = (1-t) v_d(h) + t v_d(h+1), sd = sd_error),
with sd_error unknown.
The result is that I estimate the fraction of the votes cast in a given hour, f(h), around which divisions randomly deviate. Notice that this isn’t exactly the citywide fraction of votes if those deviations from the mean are correlated with overall turnout (if, for example, high-turnout wards also are more likely to vote before 9am), but the coverage of submissions is heavily skewed towards high-turnout wards, so you should read these results as mainly pertaining to high-turnout wards anyway (see below for more). I also don't account for the fact that needing to add up to 100% of the vote induces correlations in the estimates but... um... this will be fine.
When Philadelphia Votes
The results: In this election, we saw strong before- and after-work volume, with 26.9% of votes coming before 9am [95% CI: (25.8, 28.1)] and 27.4% of votes coming between 4 and 7pm (24.7 , 30.1).
The overall excitement of the election appears to have largely manifested in the morning, when reports of lines (in a midterm!) came from across the city. There is another surge from 4-7, as people leave work, but once again voting from 7-8 was quiet (without the thunderstorm, this time).
Do neighborhoods vote differently?
It seems likely that residents of different neighborhoods would vote at different times of the day. Unfortunately, the (albeit humbling) participation in the Turnout Tracker came disproportionately from Whiter, wealthier wards, and I don’t have great data to identify differences between groups. But here’s a tentative stab at it.
I’ve manually divided wards into six groups based on their overall turnout rates, their correlated voting patterns across the last 16 elections, and racial demographics. (Any categorization is sure to upset people, so I’m preemptively sorry. But these wards appear to vote similarly, and I've chosen clear names over euphemistic ones.)
Below is a plot of the raw data submissions, divided by the eventual final votes, by ward group. The difference between groups is not especially pronounced, but it appears that Center City and Ring did vote disproportionately before 9am (along with Manayunk, which behaves like them). [Notice that in the raw submissions, some of the fractions are over 1, meaning that a submitted voter number was higher than the final vote count.]
To estimate the fraction of the vote at exactly 9am, I filter all of the data to between 8 and 10am, drop the obvious outliers, and fit a simple linear model of fraction on time for each group. The estimated fractions at 9am are below.
Nearly 28% of the Center City and Ring Wards' votes came before 9am, which was statistically significantly higher than all other groups except for the Universities (which had large uncertainty). The other wards hovered around 24% by 9am, except for the very low-turnout North Philly and Lower Northeast wards which only had 17% of their final turnout by that hour.
Energy means early voting?
It certainly appears that in this highly energetic election, a disproportionate amount of votes came in the morning. This could mean that the population that gets energized is more likely to vote in the morning (perhaps workers, as opposed to retirees), or that in excited elections, voters want to ensure their vote in the morning.
In May 2019, the Live Election Tracker will be back. And at that point, we'll have three whole datasets to be able to evaluate time of day data! Then we'll really be able to say something. Before then, I've got a number of posts in the pipeline. Next up: evaluating my various models, some better than others. Stay tuned!
The vote count on Tuesday morning had me worried. When the Turnout Tracker crawled above 198,000 voters before 10am--half of 2014 turnout through just three hours of voting--I scrambled to my computer to double, triple check the calculations. But my code wasn't off, it was the voters that had completely changed. By 8pm that night, over 547,000 Philadelphians had voted, a 43% increase over four years ago.
I'll dig into the tracker in another post, but let's do a quick hit of the actual final turnout and some maps.
* Quick Note: In everything below, I use the votes cast for President/Governor, rather than actual turnout. This will be short by however many people leave that race blank, which appears to be < 1% of voters in past years. These results are also preliminary, representing only 99.5% of precincts.
Some 537,231 Philadelphians cast votes for Governor. This is a 157 thousand increase over the 379,046 of 2014. The absolute size of this increase is unprecedented since at least 2002, and would have been the largest proportional increase in that span if not for the *82%* increase to elect Krasner in 2017. As the plot makes jarringly clear, in the aftermath of 2016, something is different.
And Philadelphia somewhat outpaced the rest of the state, where votes grew by 41%. So not only did turnout surge, but Philadelphia eked out some more statewide clout, too.
Where the votes came from
The turnout boom was not evenly distributed. As we saw in 2017, it was driven by the wealthier, predominantly white wards, and especially those that have gentrified over the last twenty years.
Immediately in the morning, the Tracker made one thing clear: Ward 27 saw the biggest turnout growth. It ruined the color scale on the map.
The 27th, in University City, saw an increase in votes of 135%. That manages to dwarf even the 95% growth of Wards 31 and 18 (in Kensington/Fishtown), in second and third.
This is worth digging into more. How did a ward possibly see this much growth? It's all Penn. The division right along the river, containing most undergrad dorms, went from 116 votes in 2014 to 585 in 2018. That manages to make the second and third place divisions look a weaker green, even though they quadrupled(!!) their votes, from 88 to 363 and from 82 to 322, respectively.
This is mostly a problem of denominators, though. Turnout at Penn was tiny in 2014. It's not as if Ward 27 all of a sudden has the most votes in the city. Instead, the student-heavy ward just this year performed like the average center-city-ringing neighborhood.
As evidence, consider instead 2018 votes as a fraction of 2016. I like this comparison, because the Presidential election of 2016 probably represents the highest plausible attainable turnout for a midterm; you're just never going to get someone who doesn't vote for President to vote for Governor.
)By this comparison, Wards 9 and 22 (Chestnut Hill and Mount Airy) look great, but typically so, with 88% and 87% of the 2016 voters coming out again in 2018. In third place was Ward 31 in Kensington, and Ward 18 below it was sixth, with 87% and 85% of 2016 turnout, respectively. Notice that these two also saw the second and third most *growth* since 2014, and have quickly ascended the ranks of top voting wards in the city.
Meanwhile, many of the Black wards that always come out in midterms continued to do so, including in Overbrook and Wynnfield in West Philly, and in Cedar Brook and West Oak Lane in the Northwest. These wards didn't see their turnout grow a ton, mostly because they always, always vote (at least relative to the rest of the city).
Many of the divisions in Ward 31 to the North saw triple the 2014 turnout, whereas divisions in Ward 18 below it merely doubled.
[Edited: My map was incorrectly handling new division 18-18, so I’ve removed it until I have a fix].
The growth in these districts aren't always what a candidate will care about. Growth in a division doesn't matter all that much if it started at a very low point, or if nobody lives there. For candidates, what might be most useful is just density of votes, which will be a function of population density and turnout. In this metric, Fairmount and West Center City glow, along with Ward 46 in University City. These are all divisions with high turnout and a ton of people.
Finally, here's turnout as a function of the population over 18. This has some benefits over the typical reported turnout of voters divided by registered voters because it (a) doesn't rely on efficiently taking people off the books, which is notoriously slow in Philadelphia, and (b) bakes into the calculation any systematic differences in getting registered to vote, which I claim should be considered part of the voting process. It's main (large) downside is that it includes in the denominator non-citizens or other residents not eligible to register, which will make immigrant communities look particularly bad.
Grad Hospital, Fairmount, and Chestnut Hill shine by this metric, while North Philly, the lower Northeast, and even Penn despite all its growth have low percentages.
Sadly, one demographic has been under-represented in every single map in today's post: the Hispanic communities in North Philly. For example, Ward 7 is the darkest ward in the map of 2018 vs 2016, meaning that despite its 63% vote growth over 2014 (11th best growth in the city!), it still had the worst turnout relative to 2016: only 52% of the votes from 2016 came out for 2018. This community already has the lowest turnout in Presidential races, but it shrinks even farther in non-Presidential elections.
Two straight elections of something different
Philadelphia's turnout since 2016 has been astounding. Across the city there were 43% more votes than four years ago, and every single Ward's turnout grew.
While voters always come out in Center City, Chestnut Hill, Overbrook, and Cedar Brook, we saw unprecedented booms in Fishtown, University City, the rest of the neighborhoods ringing Center City.
These changes have now stuck around for two straight elections--2017 and 2018--and could presage a fundamental change in our city's political calculus for years to come.
Want a spreadsheet of your own to follow the results live on election night? Here's a cheat sheet of all 203 races, with my predictions and 2016 results.
(Let me know if you find any problems!)
The Live Turnout Tracker is back. And it's got some new bells and whistles.
I've projected that Philadelphia is on pace for a record turnout year, estimating that 460,000 Philadelphians will vote on November 6th. This would beat every midterm election at least since 2002. Can we really do it? Let's watch in real time!
What it is
The Turnout Tracker is a tool for all of us voters to share our information to generate real-time vote counts on Election Day. When you vote in Philadelphia, just record your Voter Number at your precinct (you'll be able to see it when you sign in). Then share it here: bit.ly/sixtysixturnout
I'll aggregate the submissions, do a little bit of math, and estimate the current number of voters across the city. You can follow it here on election day: jtannen.github.io/election_tracker.html
The model is largely the same as the primary. You can read about what it does here: announcing-the-live-election-tracker.html
I've changed a few things, given the Primary's post-mortem. In May, I used too much smoothing, and the result was that I missed the collapse in voting between 7-8pm due to the ill-timed thunderstorm. Worse, so many people participated that I could have gotten away with a lot less smoothing, so the error was avoidable.
This time around, I've built an adaptive smoother based on the number of datapoints. In connection, I'm going to bootstrap the uncertainty for every prediction, instead of using a parametric estimate. Non-parametric smoothers (the algorithm that gives me my curvy line) are particularly noisy around the edges. Unfortunately, the edge is exactly what we care about most: the final turnout is the estimate right at 8pm, the edge of our window. Making the model more responsive to short-term changes in trajectory also means it could overfit data points right at 7:59, and really mess up the curve. Bootstrapping addresses this by resampling the observed datapoints at random, giving a range of estimates that aren't all dependent on the same points. The result is that it should correctly identify the true amount of noise right at 8pm, and give wide enough uncertainty accordingly. This will make my estimates less certain--and the plausible range larger--but hopefully allow the model to be robust to drastic end-of-day bends. And the more people that participate, the more the model will adapt to small signals, and the better result we'll get.
What I need from you
Vote! Get your friends to vote! Then have everyone share their voter number at bit.ly/sixtysixturnout. Y'all came up huge last time around, and I'm hoping we can do it again.
See you on November 6th!
Ok, I've gotten a ton of great engagement, and some ideas from you all for questions to ask of the model.
In doing so, I found that it was treating a certain candidate very weirdly: incumbent candidates who were uncontested in the last race. Usually it's bad form to change a model after you see the results, but this feels more like a bug than post-hoc analysis. So I've updated my model, and the results have changed fairly significantly.
What was off?
I thought I'd write up a post about what my model thought about incumbency, and pretty quickly found an issue. I rely very heavily on prior elections' results, and when a candidate hadn't been contested before, I was using a combination of prior Presidential results and congressional results. It turns out this was a bad idea: I was predicting that about 30% of these candidates would lose. Instead, in the last 8 elections that number has been about 5%.
How did I miss this? The model was performing well on past years, and this imbalance didn't have any impact there. But this year, we have a fundamental change in the number of contested races, and the balance between Democrats and Republicans. In 2016, there were 43 uncontested Democrats and 50 uncontested Republicans. In 2014, there were 56 and 57. In 2012, 45 and 49. This year, there are 55 uncontested Democrats and *23* uncontested Republicans. That huge gap multiplied this error, and gave Democrats about 5.5 too many seats.
The New Model
Since I think this is a bug, rather than a modelling decision, I decided to commit the cardinal sin of refitting the model after seeing the results. I basically changed it to treat incumbents who were uncontested the last year as a completely separate category, and the results are that (a) they do way better than I was predicting, and (b) the model is much more confident in the results. That combination of Democrats being expected to win 5.5 fewer seats, and the uncertainty shrinking, means that the probability of Dems getting over the 101.5 seat threshold is much (much) smaller.
Average Result: R 107, D 96
90% Range of Predictions: R 115, D 88 -- R 99, D 104
Probability of winning the House: R 87%, D 13%
How did seats change?
Two seats were big clues to what was going wrong with the model: 127 and 177.
Reading's seat 127 has been represented by Thomas Caltagirone for the last 41 years. He's facing his first challenger since 2002. And yet my model gave him essentially even odds. Why? Because his district was basically even in the 2016 US Congressional race. I was relying too heavily on the last results from other races, pretending those were state house results. The new model? Gives him an 88% chance of keeping his seat.
Seat 177 is familiar to Philadelphians as Rep. John Taylor's former seat, being contested by Patty Kozlowski (R) and Joe Hohenstein (D). It only gave Hohenstein a 34% chance of winning despite the fact that the longtime incumbent had stepped down, and Clinton won the district handily. This one was a weirder bank shot: I wasn't giving John Taylor his due as a candidate, so I was instead scoring the district as more Republican than it really was, and when Hohenstein lost the district in 2016, scoring him weaker than I should. The new model realizes Taylor was a strong candidate in a quite Blue district, and gives Hohenstein now a 66% chance of winning.
Overall, it is able to better differentiate candidates, and separate them from the 50% threshold. This means that there's a lot less noise in the model, which gives Democrats less chance to surge over 101.5.
[See the seat-by-seat results here]
Ok, that's it. Sorry for the thrash. Time to go build the Turnout Tracker.
Data comes from the amazing Open Elections Project.
I also leaned heavily on Ballotpedia to complement and extend the data.
GIS data is from the US Census.
 These numbers don't include cases where candidates run as e.g. D/R, which always happens in a few uncontested races (5 in 2016, 2 in 2014, 6 in 2012). For the model, I do impute their party based on other elections.
The midterm elections are only two weeks away. Finally. Thankfully.
All eyes have been focused on the national elections. But in this space I've been looking at races for the PA General Assembly's lower house. At first glance the house could appear out of reach for Democrats: the districts are famously gerrymandered, and the State Supreme Court's February decision didn't affect the state legislature boundaries. Democrats are down by 37 seats heading in to the race.
On the other hand, 19 districts that voted for Republican representatives in 2016 also voted for Clinton. And it looks like the state could have higher turnout than any midterm since at least 2002. Could the Blue Wave possibly sweep down to the state races, and change our state houses? Today I present my forecast.
Modeling the PA House
Because it's nerve-wracking to publish predictions into the world, and because I don't have nearly the extent of data to do this responsibly (9 elections since 2002 doesn't give me much to work with), I'm going to walk through some preamble before we get to the predictions. What information does the model use? When it inevitably is wrong, why will that be?
The challenge in forecasting these state races is that we don't have polling. This is a huge problem, because public sentiments in districts are all correlated with each other, and if you don't have some indication of what "type" of election it's going to be, your prediction will have to cover the options from Blue Wave to Red Wave, giving gigantic error bars. Making it useless.
To get around this, I borrow information from polls for the Congressional races, and measure how historically congressional races have correlated with their local state races. It turns out that correlation is strong. So I pull in the current predictions from FiveThirtyEight's House Model as a noisy estimate of each congressional outcome, which is highly suggestive of a strong Democratic year.
There are obviously a ton of other factors, including each district's own history, incumbent candidates, and ways that certain districts tend to move together. I ended up choosing a fairly simple model, to account for the few years I have to work with. This means that it will capture the strongest trends, and leave the rest as uncertainty. The model relies mostly on (a) a district’s last election results for state house, congress, and the president/governor, and (b) the FiveThirtyEight predictions for the current governor and congressional races. It also has election and candidate random effects, to capture correlated outcomes in an election and certain historically over-performing candidates (e.g. 177’s John Taylor).
As such, I expect the model to do a good job of capturing the overall mood of the state, as well as the historic stickiness of incumbents in specific races. Also, when we have a new, unknown candidate, it greatly increases the uncertainty in the race, since it knows nothing about them. If you have a favorite candidate that has an awesome twitter account, the model doesn’t know that. Instead, it shows the full range of outcomes for the historic variance in quality of candidates.
How to read the model
It’s scary to release a forecast into the world, especially one that’s so bullish (see below…). Here’s how to read the results: It’s a simple accounting of a few features that clearly affect elections, weighted for how those features have mattered since 2004. How do you account for gerrymandered districts? For incumbency? For the fact that 2018 is a midterm, the increase in districts that Democrats are contesting, and the overall tenor of US Congress races? This model looks at the historical size of the effects, and weights them accordingly. What it doesn’t do is account for the fact that maybe this election is Different™, and won’t look like any we’ve seen in the last 12 years.
How my model could be wrong
Given the lack of historic data, it’s possible that I haven’t captured the full range of potential types of elections there are. Given the need for lagging variables, I’ve further limited the elections I use to model to only elections from 2006-2016. If this election looks wildly different from any of those, then my forecasts could be wrong.
Particularly, if the large attention to the midterm or the nationalization of state politics has changed the correlation between competitive races and turnout, the power of incumbency, or the quality of new candidates, the model could be very wrong, largely by cases where all races swing together. Plausibly, if Democrats are running fundamentally different types of candidates (perhaps more skilled, perhaps more leftist) than in the past, the model won’t capture whether the candidates fit the district better or worse than before.
Finally, one of the largest sources of uncertainty is new candidates. I give the model no information about that, so every time it sees a new candidate, it has to cover the full range of quality, from terrible to great. This increases the uncertainty in each race, and uncertainty will shrink the prediction towards 50-50 (if the Republicans are favored to win most seats by a narrow margin, for example, added uncertainty will more often switch Republican seats to Democratic than vice versa.)
Enough hand wringing.
I predict that in two weeks the Democrats will win between 92 and 110 seats. On average in my simulations, they win 101.5 seats, which is annoyingly, bizarrely, exactly half of 203. They win the majority 53% of the time.
This is surprisingly bullish for Democrats. The average represents a 18.5 seat pickup for them, and even the low end of the confidence interval--92 seats--would be a 9 seat pickup.
My predictions are particularly optimistic on the Philadelphia area’s Democratic challengers, largely because of the sweeping victories expected in the region's Congressional races. Remember: the model doesn’t know anything about the state candidates themselves, and just uses broad indications of the district and the environment, so your knowledge of a given candidate could mean very different predictions. It gives Democrat Kristin Seale a 32% chance of unseating Quinn in Delco’s 168. It gives Hohenstein a 34% chance of winning the River Wards’ 177, now that John Taylor isn’t in the picture. And it gives Michael Doyle a 36% chance of beating Martina White in 170 in the Northeast.
Below is a widget where you can see the predictions for every single race:
[EDIT: The widget didn't embed correctly. CLICK HERE for it!]
One thing you may notice above is that even in the close races, Republicans are favored to win. That's a fascinating fact of the model, and the race in general. Republicans are actually favored to win in in 115 seats.
How is that possible? How can Republicans be favored in more seats than is my upper bound on how many they'll win? The answer is gerrymandering. Because of "cracking and packing", there are a ton of districts that are safely Republican in any typical year, but not so safe as to waste too many votes. A Blue Wave would push them just to the boundary. Any additional randomness--a good Democratic candidate, a local story--pushes them over to Democrat seats. This is also helped by the fact that Democrats are contesting more districts than ever before. And there just aren't any similarly teetering seats on the Democratic side. So while the model isn't sure which of the close seats will swing over the line, it's sure that some will. And maybe enough to win Democrats the house.
See you in two weeks!
I was stunned to predict such a close race. Democrats will pick up at least 8 seats, and are an even bet to win the house. With that aggressive prediction on the internet forever, I'm turning my attention to the Turnout Tracker. Stay tuned!
Data comes from the amazing Open Elections Project.
I also leaned heavily on Ballotpedia to complement and extend the data.
GIS data is from the US Census.
Forecast: Who will win the PA House?
The race for the Pennsylvania Senate
The race for the Pennsylvania House
Evaluating the Live Election Tracker
So you wanna be a Committeeperson