Most of the country saw modest shifts in vote margin from 2016 to 2020. The Texas border stands out for the intensity and breadth of the pro-Republican shift. Miami is also attention getting, but here we focus on Texas.
It isn’t that the Texas border counties gave majorities to Trump. Most didn’t. But the swing in these counties, many with large Hispanic populations, was unexpected.
Democratic gains in urban and suburban counties were partially offset by GOP gains along the border.
The Texas border counties stand out nationally, not just in the state. This chart shows 2020 vote margin by 2016 margin, with the Texas border counties highlighted in red.
And I mentioned Miami earlier. It’s the big gray circle below the diagonal near the Texas counties.
For Texas to become a competitive state would be quite a thing. To do so, Democrats must consolidate gains and stop the loses. For Republicans, the prospect of balancing urban/suburban loss with border and western gains is a possible solution for continued hold on statewide offices.
Hmmmm. I wonder what this is about? (OK: tl;dr it is about midterm elections and what to expect in 2022).
“Surge and decline” is the title of a 1960 article by Angus Campbell:
Abstract
The tides of party voting are as fascinating as the fluctuations of economic activity. Regularities in the ebb and flow of voting with the alternation of Congressional and Presidential elections challenge the analyst to find an explanation. This article seeks it in propositions rooted in survey data.
Campbell’s explanation was that presidential elections are largely driven by “short term forces” that provide a temporary advantage to one party, usually the winner of the presidency. Then in the midterm, with no presidential election, short term forces are less important and the electorate shrinks with lower turnout (a “low stimulus election” relative to presidential) and the “long-term force” of partisanship becomes relatively more dominant in the midterm. The result is a return to the partisan balance after the surge favoring the presidential winner two years earlier.
This is a beautifully elegant theory. It is rooted in a simple model of partisanship, short-term forces and the inevitable decline of turnout in midterms. I love it.
Alas, it no longer commands general acceptance as a theory of midterm seat loss. More emphasis is now given to presidential approval, economic conditions and incumbency. Those theories bring substantial empirical evidence, and are certainly sensible. But to me they lack Campbell’s beautiful simplicity.
I’m not here to argue theories of midterm loss, but rather to simply illustrate the votes side of midterm losses. That the president’s party almost always loses house seats in midterms is a fact. Here I look at the decline in votes for the president’s party from presidential to midterm.
For fun, I’m going to walk you through the puzzle and steps that end with the figure at the top of this post. Here is the first step. What IS this??
As my poor former students know, I enjoy starting a class example with a mystery. Show the chart above, invite speculation as to what it might be. I love this one because it looks like random noise with no relationship at all.
Next step: Revealing the variables.
OK here are the variables. National Democratic percentage of the 2-party House vote in the midterm by the national 2 party vote in the previous presidential election. Not much of a relationship. Some votes go up (above the diagonal) and about as many go down (below diagonal.)
Ahh, but what about control of the presidency? The midterm-loss of seats by the president’s party is well known. The national vote, here, shows the same pattern. Dems do better (above diagonal) with GOP president, and do worse with a Dem president, almost always (except 2002.)
This is the surge and decline of votes. Almost always the president’s party wins a smaller share of votes in the midterm than they did in the presidential year. Not surprisingly, fewer votes translate into fewer seats, but that’s not our topic here. For that see this post.
The pattern is clear if we fit a regression for midterms with Rep presidents (red line) and one for Dem pres (blue line). Now the upward slope is clear (midterm performance IS related to prior pres vote) and the party of president shifts the lines up (Rep pres) or down (Dem pres).
FWIW the red and blue lines are nearly parallel. A test of a pooled model finds the difference in slopes to be statistically insignificant (p=.9465). I’m using the separate regressions here, but there would be minimal difference for a pooled model.
So what does the model tell us about, say 2018? The red line estimates the Dem 2-pty vote in 2018 to be 53.1%, up from the Dem 2-pty 2016 vote of 49.5%. In fact, Dems got 54.4%, 1.3 points better than the model and 4.9 points over their 2016 performance.
We don’t know how Dems will do in 2022, but we do know how they did in 2020 and that there is a Dem president. The vertical black line shows the actual Dem share of 2020 House vote, and the blue arrow shows the fit: a predicted 47.8% in 2022, down from 51.6% in 2020.
This is the dilemma of every presidential party: they are almost certain to lose votes in midterm elections. For closely divided congresses (looking at you 117th) this imperils majorities. 2002 was an exception, with 1998 and 1990 almost being exceptions.
How do votes translate into seats? The relationship shifted after 1994 undoing a long standing Dem advantage. The votes-to-seats model expected Dems to hold 50.6% or 220 seats in 2021 (actual post-election was 222). For 2022 the estimate is 45.0% of seats. That would be 196 seats, a loss of 26 from the post-2020 election total.
The president’s party gained house seats in 1934, 1998 and 2002, and lost share of seats in every other midterm since 1862. Reps gained seats in 1902 as the House expanded but actually lost share of seats as Dems gained more that year.
If the historical pattern applies in 2022 the Democrats are unlikely to hold control of the House. In addition there will also be the effect of redistricting. Both parties will have an incentive to gerrymander for every advantage possible where they control the process.
Of course the past pattern may change. Political skill or folly might shift the balance away from the models. The model is useful because it gives us a basis for our expectations. We can judge party performance by whether outcomes exceed or fall short of model expectations. 12/12
How votes are converted to seats in the House of Representatives and how that has changed.
In a perfectly proportional legislature the percent of seats should equal the percent of votes received by a party. Electoral systems based on proportional representation come close to ensuring this by design.
Two-party, plurality, systems are rarely if ever proportional. They tend to reward votes disproportionately, giving more seats than votes to one party and fewer seats than votes to the other. They also often award a majority of seats for less than a majority of votes.
One measure of the bias in a system is the “representation ratio,” the percent of seats divided by percent of votes. A value over 1.0 means a party gets more seats share than votes share, and values less than 1.0 means underrepresentation.
In the US from 1942 until 1994 the Democratic party was advantaged, with a representation ratio typically around 1.1 with variation across elections. After 1994 that reversed, with the Republican party enjoying an advantage, a bit smaller than Dems had.
Part of this was the “solid South”, dominated by Democrats until the 1980s coupled with very low turnout which made winning Dem vote totals smaller than in competitive elections. The transformation of parties in the South after 1980 became a GOP advantage.
Gerrymandering also plays a role in the vote-to-seats relationship, with advantages to parties that control legislatures and governorships that create the districts. Courts imposed some limits on districting beginning in the 1960s.
From 1942 to 1994 Democrats were advantaged in all but two elections. Since 1994 Republicans have been advantaged in all but one election (2008). The advantages were persistent in each era.
The RepRatio is a simple measure of advantage, but what about how votes are converted to seats across elections? This chart shows the percent of seats won by percent of national vote won in each election. 1942-94 is different from 1996-2020.
One measure of bias is the percent of votes required to win 50 percent of seats. In 1942-94, Dems needed 48.4% of votes to reach 50% of seats. Since 1994, Dems need 51.2% of votes to reach a majority of the House. There is uncertainty but these are the expected outcomes.
Another measure is the “swing ratio”, the slope of the regression lines, measuring how much seat share changes for a 1 point change in vote share. In 1942-94 Dems got 1.80 percent more seats for a 1 percentage point increase in vote share. After 1994 it has been 1.47.
Post 1994 Republicans gained an advantage in votes required for half the seats & reduced the swing ratio to lessen the effect of votes on seats. Both eras have swing ratios over 1.0 meaning seats are more responsive to votes than pure proportionality. This is common in 2 party single-member district systems.
If we shift to the relationship between national presidential vote and seats in the House we can extend the time frame back to 1900. I divide into two partisan eras, 1932-1992 for Dems, and 1900-1928 plus 1996-2020 with a GOP House advantage.
Interestingly, the relationship of seats and votes is essentially the same for the 1900-28 and 1996-2020 eras of GOP advantage. A test of different slopes & intercepts gives p=.69 so I combine them here.
In the 1932-1992 Dem era, a Democratic presidential vote of just 36.3% was enough to expect a 50% Dem House. In the GOP eras, a Dem president needed 51.2% of the national vote to expect half of the House.
The Democratic solid South again provided a huge advantage in the 1932-92 period. In the two eras that Reps were advantaged in the House, their advantage is much smaller, requiring Dems to get 51.9% of the pres vote, 50.3 in 1900-28 & 52.7 since 1996.
The takeaway is that Republicans converted a long time disadvantage in winning House seats to a smaller but persistent advantage after 1994. Once control of the House was won in 1994, the GOP has held an advantage, despite one reversal in 2008.
The size of the current Democratic disadvantage is important, but it should be recognized that the GOP disadvantage from 1932-1994 was far greater. Changes in regional party dominance plays a big role in that and shows party advantage can be altered.
I’ve been shocked to hear several sources I respect get the midterm seat loss story wrong. So here is my effort to clarify.
The president’s party almost always loses House seats, but there have been 4* exceptions since 1862: 1902, 1934, 1998 & 2002. *HOWEVER in 1902 the House expanded so while Reps gained seats Dems gained more, thus Reps won a smaller percentage of seats that year. So the presidents party has lost strength in all but 3 midterms since 1862.
In the Senate the president’s party usually loses seats, but not as reliably as in the House. There have been 6 exceptions since 1960.
There is little difference, on average, in House seat losses in 1st vs 2nd midterms. An average -26.4 in 1st and -28.1 in 2nd. NO SIX YEAR ITCH! NO 1ST MIDTERM CURSE EITHER, for that matter.
2nd midterms HAVE been worse in the Senate: -2.3 in 1st, -6.0 in 2nd.
So PLEASE stop saying the president’s party only gains seats “once in the last 100 years”– you know who you are. The right answer is “three times in the last 100 years.”
And don’t imply the Senate is as predictable as the House. They aren’t the same.
And… 1st term vs 2nd? Nah. This is another rant as many people bring up “first midterm” (and in a 2nd term almost always talk about the “second midterm”) as if that mattered. It doesn’t, on average. It does vary across presidencies with some bigger losses in 1st and some in 2nd midterm.
And will 2022 be different? I don’t know. But we should get the history right.
Data details
These seat changes reflect the immediate outcome of the November election. Sometimes members die, change party or resign before the Congress is sworn in, and of course changes can occur during the Congress.
Brookings hosts Vital Statistics on Congress. Note they have a typo for 1998 indicating a loss rather than a gain. I use them here with that fix
Here comes a bit about survey question wording. For those just tuning in, NPORS=National Public Opinion Reference Survey (NPORS) from Pew, which released their 2021 update today (Sept 24) (thanks, Pew!)
According to my national @MULawPoll released this week 56% say “most people can be trusted” and 44% say “most people can’t be trusted”. But today Pew released their NPORS survey conducted this summer and find just 32% say most can be trusted. What’s going on??
This difference, of course, scared the bejeezus out of me. How can Pew’s National Public Opinion Reference Survey differ so much from mine, conducted at a similar time and on a question we would expect to be a stable attitude?? Question wording, my friends. Question wording.
My question is worded “Generally speaking, would you say that most people can be trusted, or most people can’t be trusted?” That was, in fact, the wording Pew used as recently as March 2020 and July 2020. In those 2 Pew got 58% and 53% most can be trusted, close to my 56%
So did the world go all “untrusty” since 2020? Pew changed the question in 2021. Now they asked “Which statement comes closer to your view even if neither is exactly right: Most people can be trusted or You can’t be too careful in dealing with people”
And the marginals flipped: With this wording 32% most can be trusted, 68% you can’t be too careful. A year ago in Pew’s July, with the previous wording: 58% most can be trusted, 39% most cannot be trusted. So which wording should we trust?
Pew’s original wording produced pretty consistent results (with slight differences in the stem to the question but not to response options): Nov 2018 52-47, March 2020 53-46, July 2020 58-39. So quite a change to 32-68 with the “new” wording.
But (as they say) the “new” wording is actually the one Pew generally used before the 2018-2020 polls cited above. They had generally used the “you can’t be too careful” as the alternative. And it makes a big difference.
Here are Pew studies with “can’t be too careful”: Apr 2017: 42 (trusted)-57 (can’t be too careful); Apr 2017 42-58; Feb 2016 43-56; Aug 2014 52-48(a); Aug 2014 47-51(b); Apr 2012 37-59. ( (a)Web, (b)Phone, same field dates)
This isn’t a “house” issue with Pew. The GSS has asked the “can’t be too careful” version for a while: GSS-NORC 2018 32-63; GSS-NORC 2016 31-64; GSS-NORC 2014 30-65; GSS-NORC 2012 32-64. The stability we’d expect on this item over time and close to Pew’s current 32-68.
So… both wordings appear stable and across survey houses (my 56-44, Pew’s 58-39, 53-46, 52-47) but also GSS and Pew’s flipped 32-63, 31-64, 30-65, 32-64 and 32-68.
Which wording we should use is less clear. The “most can’t be trusted” is clear and direct, “can’t be too careful” touches on suspicion. A much deeper analysis is needed of this issue. But this is a great example of seemingly similar items producing big differences.
I think there is a lot to be said for consistency, so I don’t expect to change my wording. Also this isn’t a complaint about Pew. The variation in wording they used actually allows us to understand the effect of question wording. A big help.
The Pew NPORS is a major service to the survey research world. But question wording matters and we need to take it into account, especially with a “reference survey” that influences all of us. Also the trust item was not included in the NPORS for 2020, so surprised me.
There are other issues to consider where question wording and item construction differs in the NPORS (looking at you, party ID and leaners!) so let’s all take advantage of this great resource. But as someone said: “Trust, but verify.”
Sixteen years ago this week a hurricane hit New Orleans and I launched PoliticalArithmetik, my first blog. This week a hurricane hit New Orleans and I’m (re)launching a website, PollsAndVotes.com.
After a year of PoliticalArithmetik, Mark Blumenthal (@mysterypollster) and I launched Pollster.com (with the support of Doug Rivers) and spent several years explaining polling and providing tracking of races, presidential approval and other topics in public opinion. In 2010 HuffPost bought Pollster and Mark had a good run with that. I departed and started PollsAndVotes.com in 2011, but have not maintained the site for a while. This is the relaunch of PollsAndVotes.com.
For some while now I’ve primarily posted analysis of polling on Twitter at @PollsAndVotes. As much as I like Twitter (most of the time) I think it is time to again have a PollsAndVotes website that allows longer posts, in one place, that can be easily found and searched for older posts, like from last week or last month. Having an editor to fix typos is also welcome.
I’ll be building out this site at a somewhat deliberate pace. I’ve decided not to import the old posts from the previous PollsAndVotes.com let alone from PoliticalArithmetik. I’ll update some of those, such as partisanship trends, but start fresh with the current data.
There will be a mix of topics here, but I’ll not be trying to replicate what Pollster.com did and what FiveThirtyEight.com and RealClearPolitics.com do well already. Most of the analysis here will be deeper dives into the national and state polling data that goes beyond trends. I also hope that my fellow academics will find graphics that may be useful in teaching.
The menu topics at the top of the page will (eventually!) provide a quick guide to analysis of “Polls” and “Votes” but also Wisconsin politics, party id, voter turnout, roll call votes and the US Supreme Court. Those first two will be something of a catch-all category. <;-)
Sixteen years ago I spent Labor Day weekend at home instead of the American Political Science Association annual meeting, keeping up with news of Katrina and launching PoliticalArithmetik. What started that weekend changed my life. I’ve still got a few days until this Labor Day weekend, and am not attending APSA, though I’ve been following the news on Ida. I hope you find the site interesting and useful.
