Final thoughts on Election Night Viewing

 “Somewhere between a landslide and a nail-biter,” was 538’s final assessment, which other than being extremely wishy-washy rings true based on all the data I’ve seen – the current polling points to a landslide, but if it’s only off by a little, it then turns into a nail-biter.  And, we all remember that polling can be off!

Here are my thoughts on following along from home tonight.  I’ve provided two 1-page PDFs (Senate & Presidential) that provide some key data to track.  Fundamentally, the three things to watch (other than Electoral college estimates) to get an idea of what’s happening are:

1. Trumps’s national percentage compared to ’16 (he got 46.1% of the vote in ’16 and would need at least 49% to win the national vote this year because there aren’t strong 3^rd party candidates)
2. Each state’s Trump percentage compared to the polling averages (I’ve included the final 538 forecast in the PDF).  If Trump is doing 2-3 points better than the final forecasts, that’s problematic for Biden, and 3-5 points, it becomes a nail-biter;
3. Georgia, North Carolina, Florida – I started looking at the confluence of counting of early ballots, battleground states and robust, rapid election night reporting last week (that was the exercise that pushed me to write the viewer’s guide in the first place).  And was heartened to see the NY Times come to the same conclusion I have – Watching GA, NC and FL (closing at 7p, 7:30p and 8p, respectively) will indicate a lot about what will happen for the rest of the night.  The NY Times will have its “prediction needles” for all three states.  In the past, I have been impressed with their forecasting accuracy, and I hope they are able to model the complexities of early vote counts arriving differently from election day counts.

I’ll be tweeting my thoughts on the above 3 as frequently as I can pry my eyes off of real time results at @evangrossman.  I have a list of other useful Twitter accounts at the bottom

My detailed guide to viewing the election results, now includes cocktail pairing suggestions, and I’ve updated the forecasts and fixed typos.

Early voting & Vote Counting

This election will set a new turnout record; before the election day polls opened, 9 states have already exceeded, or come very close to, the number of ballots cast in 2016: AZ, CO, FL, GA, MI, NC, NM, NV, TX.  Another 5 states are above 80% of their 2016 vote (CA, NJ, TN, UT, VT).  The biggest challenge from a prediction standpoint is knowing where the votes are going.  There are arguably lot of folks who sat out 2016 and who now feel 2020 has even higher stakes—on both sides.  There is definitely clear polling that shows early mail balloting has favored Biden in FL and NC and early in-person favored Trump in those states, and additional polling suggests Trump voters are more likely to show up on election day.  The one thing I believe several of the articles about vote counting have gotten wrong, is the notion that early votes and/or mail-in ballots are reported as blocks completely separate from election day votes.  The bottom line is it depends by locality.  In some places, mail-in ballots are fed into machines during election day and the unofficial numbers reported at poll close include both.  In other places one set is reported before the other.  Several states which will are planning on having large numbers of uncounted mail-in ballots on election night are still planning to report the ballots they have been able to count.

 

Methodology Notes (read if you’re bored before 7pm eastern)

Several folks have asked how I do rapid election night forecasting, so this is a fairly dense post explaining the art and science I’ve developed over the past few years.

 

The first thing to figure out is what the appropriate ‘predicate’ variables you will use to forecast with – for example, in this election it seems like it should be pretty simple to use Trump’s 16 pct of total vote as the predicate.  You can then look at a given area (let’s say the bellwether county of Erie, PA [there was a great article about Trump bellwether counties by Dan Wasserman of the Cook Political report in the NYT a few weeks back) which gave Trump 48.6% of its vote in 2016, a year when Trump got 48.6 of the vote in Pennsylvania (it’s rare that bellwether’s are that spot on, but it does happen).  A very simplistic forecast would just look at Erie county, and if Trump loses, predict the win for Biden.  I prefer to add a few other counties, typically one that more strongly supported Trump in 16 and one that more strongly supported Clinton in 16, and then also try to determine if the relative vote share will change (ie, if Philadelphia, which was 12% of all votes cast in PA in ’16 has a higher incr. in turnout than the rest of state, then even if Trump still gets the 15.4% he got in ’16, his statewide pct will decrease). 

As the night goes on, and you compare results from more locations against the predicate (and look at standard deviation to make sure the predicate is reasonably correlated to what you’re hoping to forecast), you can start to get a good sense of who will win the state and possible the final margin.  One could argue a better predicate for this year may be ‘16Trump+16Johnson (based on some WI polling that shows Johnson voters moving to Trump) or could be an average of Trump16 & Romney12.  I like to have a few different possible predicates at the beginning of the night, and as I learn which ones have tighter correlations (I use a simple R², but I’m sure there are more complex ways to do this) I start moving to the most predictive predicate.  Note, the predicate doesn’t have to be exact, you can always weight and/or add a few percentage points.  For example, in the recent MA Markey v Kennedy primary, I was trying to determine the winner as early as possible, in a race that some predicted could be close.  There were no obvious predicates because turnout was likely to be higher than any previous statewide primary (it ended up breaking all records) and Markey had not had a competitive statewide race in awhile.  I ended up using the combined Sanders & Warren percentages from the recent Super Tuesday primary (Markey was the clear progressive in the race).  Since Warren+Sanders got 48% of the vote statewide, I figured that if Markey performed at least 3% better on average, he would win (in the end he did 7.7% better on statewide).  The predicate wasn’t super- correlated (R²=.2) but it was much more closely correlated in the larger cities and towns, so it ended up being a useful predictor

You can take the notion of using a predicate value, and looking at deviation of early results from the predicate and use it across multiple states.  For example, the statewide polling numbers are pretty good predicates (I tend to use the 538 final prediction), and you can then compare how well Trump is doing in early reporting states to the 538 numbers to get an idea of how well he may do in other areas.  For example, if Trump is 2points above the final 538 estimate in Georgia or Florida, one could assume he may also end up 2 points above the final estimate in North Carolina (you’d likely be wrong, but at least close).  Early on (before 9pm) trying to understand how closely a prediction of a state is to the polling in that state (and in some cases the national polling) can provide guidance for how to better forecast other states.

I’d caution that even really good forecasting has  a lot of issues and inaccuracies, and in a very close election (like WI, MI & PA 4 years ago) it’s not helpful (other than to possibly predict that it’s too close to call).  There are a number of issues in any election, and in 2020 with the large number of early and mail-in votes, which in some cases may not be counted and added to unofficial totals until days after election day, any early forecast could be wrong if it can’t accurately weight the impact of the uncounted votes.  That’s one reason why I think PA and WI will be very difficult to call based on early vote #s .Even though some counties, which can start running mail-in votes through machines on Tuesday morning, will include those numbers when they first report results after the polls close, others will report them later and/or separately.  The only way we’ll be able to predict WI and PA is if nearby states, with similar polling go overwhelmingly in one direction. In other words, if the polling in MN and MI, where Biden is several points up is close, then we can assume the WI polling will also be accurate. My general, very un-scientific rule of thumb is that if a poll has a candidate winning by 4+ points, and real results from nearby states show that polling is accurate, then I think there’s a reasonable chance to think the candidate will win the state (I’d still look at other factors like turnout in big cities to see if the overall turnout map was different than past elections)

Helpful Twitter resources

I'm @evangrosmsan Here are some other folks I follow for real-time insights:

@redistrict - Dave Wasserman editor / data wonk at Cook Political

@NateSilver538 - Official feed for 538’s editor in chief

@Nate_cohn - NY TImes Upshot reporter

@joetrippi - Dem campaign consultant

@Davidaxelrod - Former Obama campaign manager, good commentary

@JacobRubashkin - Inside Elections analyst

@maristpoll - Very well respected polling org.