Skip to content

2020 general election, polling

April 28, 2019

The power of procrastination worked its magic again – I just wasted 3 perfectly good hours looking at polling data. I’m talking about the 2020 general election. Still far too early to speculate, but let’s do it anyway. These are my opinions, and I’ll stand by them until they change 🙂


  • Dem expected to win by about 20 electoral votes.
  • deciding states: PA, FL, (AZ, maaaaybe)
  • swing but in my opinion, looking like Dem: MI, WI
  • swing but in my opinion, looking like Rep: AZ, probably
  • watch, but prob not deciding: NV, VA, OH


The election will, even more than usual, be a referendum on the incumbent. The best public data so far is the copious national Trump approval polling, but we have a couple of per-state polls.

(1) Conceptual notes

  1. I focus on polls of registered voters, rather than all-adults, or likely voters — since the RV’s give us a consistent basis to make a comparison.
  2. I assume that in an approval poll, we can split the undecided’s 50-50 to figure voting intention (this is a dubious assumption, in some states it checks out, in others it doesn’t).
    1. Example: a poll of RV’s saying 40%/50%  Trump approve/disapprove -> implies a 45% Trump voting intention among RV’s, unless we have data to say otherwise
  3. My Likely-Voter model will be super simple:
    1. voting_intention_LV = voting_intention_RV + 2%.
    2. This is a touch on the low side historically.
    3. Unfortunately, this assumption will make all the difference.
  4. Actual polls of voting intention take precedence over approval numbers.
  5. I am prepared to completely disregard national numbers in favor of individual state analysis

(2) National Trump approval polling, from FiveThirtyEight-polls-listing:

Looking at their raw data for this year, looking ONLY at RV polls, and using the weightings they suggest, I get RV approve/disapprove = 42.5/53.5  [trailing average of the past month or so]

Splitting the undecided’s, that implies Trump gets 44.5% RV or 46.5% LV. Subject to state level breakdown, this level of national support is typically a point or two shy of putting a Republican in the white house.

(3) State level Trump approval polling, from Morning Consult, per their latest (Mar’19), for selected states.

STATE : RV-Trump-approve/disapprove (implied Trump voting-intention)

NC: 49/48 (RV=50.5 , LV=52.5)
FL: 47/49 (RV=49 , LV=51)
VA: 46/50 (RV=48 , LV=50)
OH: 45/51 (RV=47 , LV=49)
AZ: 45/51 (RV=47 , LV=49)
PA: 45/52 (RV=46.5 , LV=48.5)
NV: 43/53 (RV=45 , LV=47)
IA: 43/53 (RV=45 , LV=47)
MI: 42/53 (RV=44.5 , LV=46.5)
WI: 42/54 (RV=44 , LV=46)

NB! I don’t really trust this analysis (the numbers in the parentheses above). Some widely held consensus expectations that do sound right to me, such as OH and IA expected to vote Rep. and VA expected to vote Dem, contradict it.

(4) actual state level polling for general election matchups (per FiveThirtyEight again)

FL: There essentially isn’t any. This in itself is actually a red flag.
One poll from March 6th (RV) that 538 didn’t even list in their per-state breakdown:
favorability/unfavorable/no-answer: 43/52 (implies 45.5% RV)
re-elect/not: 40/53 (implies 43.5% RV, which seems like it’s on the low side for the perennial tossup state)

PA: 538 lists 4 state polls. Most recent one: Trump approval: about 41%, Trump RV-voting-intention about 45%. In this case, the crude formula for converting approval/disapproval into voting intention seems to fit. Other interesting state note for PA: when asked which news source they trust the most (e.g. FOX,CNN,etc), the plurality,  responded “none-of-the-above”. Presumably this includes the critical undecided / decided-but-unsure-to-turn-up voters. That one could hurt, in light of the boy-cried-wolf aspect of Russiagate. Anyway, PA is shaping up to be the sine-qua-non swing state of 2020.

AZ: 538 lists Just one poll from Feb, showing Trump even with Biden, and beating everyone else (RV voting intention).

MI/WI: I think Dems will take these states without too much trouble. 2016 was a bit of a fluke, with lower turnout among african americans (WI) and latinos (MI). (source: interesting article by a gigantic PAC). Anyway, Dem organizers won’t make the same mistake again in these states.

(5) plugging it all into the electoral map (…

Exercise for the reader =).

(6) other resources:

Article on “Likely Voter Model” from 2014 (Pew). Note that it is for a non-presidential year, but serves as a useful example still.

PA turnout data (.gov)

MI turnout data (.gov)

WI turnout data (.gov)

turnout data for all states (united states election project)

turnout data for all states (

analysis of turnout (and other factors) for 2016 post-mortem (center for american progress, a gigantic PAC)

Trump approval, with state breakdowns, via correlations, but from sizeable data set. (Morning Consult)

Trump approval, aggregation of national data (FiveThirtyEight… I recommend filtering out the non-RV data, since it just confuses the picture. They massage the topline, in ways I am mostly okay with, but you can get at the actual data in the footnotes. It’s interesting to fish out series/trends for individual pollsters)

electoral map calculator (

An Evaluation of 2016 Election Polls (American Association of Public Opinion Research) – Asks why the LV models in the 2016 polls missed so badly in the critical states (a month out, Clinton was showing as +10 in many polls in PA!), for polls conducted more than 1 week prior to the election.

  • An interesting PEW “callback study”, surveying the same voters before/after the election. About 11% of all voters did not vote as they told pollsters they would prior to the election, which is typical. What was atypical was that this 11% favored Trump, resulting in a roughly 1-2% bonus in the national outcome. I.e. the “Shy Trump Vote” phenomenon.
  • The other perhaps bigger phenomenon was that about 10% of voters in the swing states simply made up their minds during the very last week, and in the states that mattered, they broke heavily for Trump, giving him another 2-3% in those states.
  • A lesser effect was that college-educated voters were believed to be over-represented in state level polling.
  • What was unusual is that all the systematic errors in 2016 mostly went in the same direction, except for one or two pro-Trump polling houses.
  • For this year, the only real conclusion I can make is to be more skeptical of LV models.

From → Uncategorized

Leave a Comment

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 )

Twitter picture

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

Facebook photo

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

Connecting to %s

%d bloggers like this: