U.S. House · 1987 to today
Where they really stand.
Talk is cheap; votes are on the record. Every House vote since 1987 reveals where each representative actually stands, and how polarized the House has become.
Voting November 3? Every incumbent has a record. See theirs. For what's on your ballot, see your state election office.
Look up a memberThe trend
Since 2011, polarization has reached its highest levels since 1987
Every Congress since 2011 has been more polarized than any Congress from 1987 to 2011.
What does the index mean?
It measures how spread out representatives are around the circle. When polarization is low, they bunch near the top, much like points on a line. As it grows, the two sides spread around the circle until the most liberal and the most conservative curve toward each other at the back.
Every House since 1987
Watch the polarization grow
One ring per Congress: 1987 at the center, today on the outside. Each tick is a representative. Over the years the rings spread wider: polarization grows, and the most liberal and the most conservative vote together more often.
Inside each party
Democrats moved closer. Republicans became more divided.
When the ends meet
The most liberal and the most conservative, against the middle
April 2024: the House renewed FISA surveillance powers, 273 to 147, with most of both parties behind it. The no votes came from the two ends: nearly all of the 10% most conservative members, and four in five of the 10% most liberal.
A straight-line model can't explain a vote like this. To fit it, the line has to treat these members as moderates. The circle keeps them where the rest of their record puts them: at the two ends.
Want the Senate next? Tell us, and we'll run the same analysis for the Senate.
Votes, not words
Positions come from floor votes. Nothing else.
Ranges show uncertainty
Members with fewer votes tend to have wider ranges.
Standings, not spots
Compare a member's standing within each Congress, not exact positions across years.
Explore a House
119th Congress
Tap a dot to see the member.
Every member, in order
Most liberal first
Find a member
Where does your representative stand?
Current and past members since 1987.
Methods & data
How it works
- Take every recorded House vote since 1987.
- Place representatives on a circle so that those who vote alike sit close together.
- Rank them around the circle, from the liberal end to the conservative end.
Why a circle?
On a line, every vote splits members into a left group and a right group. But some votes split the middle from both ends. A line can't draw that split. A circle can.
It matters for rankings. Line models placed the four members of "the Squad" between 172nd and 203rd of 430 in 2019, as if they were moderates, because they often voted with Republicans against their own party's leadership. Adding a second dimension did not fix it.1 On the circle they rank among the four most liberal members.
The idea echoes the so-called "horseshoe theory", which asserts that "the far-left and the far-right are closer to each other than they are to the political center".1 The circle only shows that the two ends vote together, not that they agree. The paper's examples from 2019:1 the defense bill (S. 1790) was opposed by Democrats Rashida Tlaib and Earl Blumenauer and by Republicans Thomas Massie and Justin Amash. A resolution opposing the boycott movement against Israel (H.Res. 246) was opposed by 16 Democrats, citing free speech and Palestinian rights, and by Massie, on libertarian grounds that Congress should not interfere in Israel's decisions or in private boycotts. Same vote, different reasons.
When there is no ends-against-the-middle voting, the circle reduces to the traditional line. Members then stay near the top of the circle, where it is almost flat, and the rankings match the usual line model. The model's prior is built so that this line-like case is the default: members start out concentrated near the top, and only the votes can pull them around.1
Keep in mind
- Compare standings, not spots. Each Congress's circle is estimated on its own, so a dot's exact spot in 1990 and in 2020 can't be compared. You can compare the polarization index, which uses the same 0 to 100 scale every Congress, and a member's standing within their Congress ("more liberal than 80% of colleagues").
- Ranges. Each rank comes with a 95% Bayesian credible interval: under the model, the rank falls in that range with 95% probability.
- The 119th Congress is in session. Votes through .
- Missed votes. Members who missed many votes have wide ranges. Under 25 votes: not estimated. Non-voting delegates (D.C., territories) aren't included.
Details for statisticians
Model. Bayesian circular factor model.1 Members and each vote's yes and no positions are angles on the unit circle; a yes is more likely when the yes position is geodesically closer. Fitted with geodesic Hamiltonian Monte Carlo (JAX, GPU), 2 to 4 chains per Congress. New Metropolis moves that mirror or shift each vote's yes and no positions, and regularly redrawn step sizes, help chains escape configurations where they get stuck. Chains that agree are pooled; where one chain's log-likelihood is more than 3 within-chain standard deviations higher, it is used alone, and each such choice was checked against voting records.
Line as a limit. As the concentration of the members' von Mises prior and the vote-specific link precisions grow, the circular likelihood and prior approach those of the one-dimensional Euclidean probit model on the tangent line at the top of the circle; the prior (mean concentration 10, so about 95% prior mass on the upper half) favors this regime (paper Sections 4.3–4.4).1
Polarization index. 100 × χ, where χ = 1 − R̄ is the circular variance of the posterior member positions and R̄ is the length of their mean direction. χ lies between 0 and 1 on the circle, so it is comparable across Congresses (unlike the variance on a line, which depends on an arbitrary scale). Within-party indices use one party's members. Raw χ values are in the downloads. Comparisons such as "about 5 times" compare index values; for small spreads the index grows roughly with the square of the spread (χ ≈ σ²/2), so 4× the index means about 2× the spread.
Ranks. Posterior median and 95% interval of each member's rank around the circle, starting from the liberal end, among members fitted directly.
Members with many missed votes (more than 40% missing) are handled in two stages. First the model is fitted to everyone else. Then each of these members is placed with a cut posterior:23 for every posterior draw of the vote parameters, their position is drawn exactly from its conditional distribution, evaluated on a 4,096-point grid over the circle. Their votes inform their own position but do not feed back into the vote parameters. Because the draw is exact for every posterior draw, this avoids the naive cut sampler, which does not converge to a well-defined limiting distribution.2 On when cuts are appropriate and how they behave, see 456. Cut credible regions can under-cover,6 so we checked coverage in a simulation study. These members are then ranked among the fitted members.
Party switchers. Listed under the party they cast most of their votes with in that Congress. For within-party indices, members who left a party to become independent count with the party they left (Justin Amash, 116th), or with the party they joined in the same Congress (Virgil Goode, 107th). Bernie Sanders (independent) is left out of both.
One uncertain case. Pat Schroeder (D-CO) in the 103rd Congress fits two places almost equally well: around rank 80, or near the liberal end. The site shows the better fit; treat her position as uncertain.
Data, code and paper
- Download the estimates (CC BY 4.0)
- Code on GitHub
- Votes: Voteview (Lewis et al.). Names: congress-legislators.
- Yu, X. and Rodríguez, A. (2021). Spatial voting models in circular spaces: A case study of the U.S. House of Representatives. The Annals of Applied Statistics 15(4), 1897–1922. Read on Project Euclid. The Squad example and the two-dimensional model: Section 2 and Table 1. Horseshoe theory: Section 1. S. 1790 and H.Res. 246: Section 6.1. Relation to the line model: Sections 4.3–4.4.
- Plummer, M. (2015). Cuts in Bayesian graphical models. Statistics and Computing 25(1), 37–43. doi:10.1007/s11222-014-9503-z
- Lunn, D., Best, N., Spiegelhalter, D., Graham, G. and Neuenschwander, B. (2009). Combining MCMC with "sequential" PKPD modelling. Journal of Pharmacokinetics and Pharmacodynamics 36(1), 19–38. doi:10.1007/s10928-008-9109-1
- Jacob, P. E., Murray, L. M., Holmes, C. C. and Robert, C. P. (2017). Better together? Statistical learning in models made of modules. arXiv:1708.08719
- Carmona, C. U. and Nicholls, G. K. (2020). Semi-modular inference: enhanced learning in multi-modular models by tempering the influence of components. AISTATS 2020, PMLR 108, 4226–4235. arXiv:2003.06804
- Pompe, E. and Jacob, P. E. (2021). Asymptotics of cut distributions and robust modular inference using Posterior Bootstrap. arXiv:2110.11149
About
Full Circle is nonpartisan and informational. It is not affiliated with any party, campaign or Voteview, endorses no one, and uses no cookies.
It is built by Xingchen Yu, who developed the circular model with Prof. Abel Rodríguez during a PhD in Statistical Science at UC Santa Cruz (published in 2021). This site reruns the model on every House since 1987 with a new GPU implementation, new sampler moves that free stuck chains, and chain checks against members' voting records, and adds plain-language tools for voters. LinkedIn