Congressional Retirements, 1789-2024

The test

Using the membership rosters of the House, 1789 to 2025, joined to the federal record of who ran and how they did. Three files are stacked, and they are not equally good. Voteview’s member file is the machine-readable form of the official membership record: one row per member per Congress — the numbered two-year session that convenes after each House election, so the 118th Congress is the one that met in 2023 and 2024 — back to the First, maintained by political scientists at UCLA and downloadable at https://voteview.com/data. The Federal Election Commission’s (the FEC’s) Federal Elections compilations, 2004–2022, report every House candidate’s vote in the primary, the election each party holds to choose its nominee, and in the general election in November, compiled from the states’ official canvasses and downloadable at https://www.fec.gov/introduction-campaign-finance/election-results-and-voting-information/. Brookings’ Vital Statistics on Congress counts retirements and defeats back to 1946, and gives the derivation something independent to be checked against.

A roster is the only instrument that can hold this question. Leaving is a membership fact: a person served in Congress N and did not serve in Congress N+1, which the roster states for 118 consecutive Congresses with no inference at all. A returns file cannot see the member who never appears on a ballot, and that member is precisely the subject. The roster chapter introduces this record, what it was made for, and why most House turnover is invisible to election returns.

The data

Put the roster beside the candidate file and each departing member falls into one of four places:

The first of those four is a residual, the category for what is left over, and residuals are where the trouble is. A graceful retiree, a member resigning under indictment, and a member quitting ahead of an expected defeat all land in it together. Deaths can be pulled out, because the roster records a death year; Senate runs can be pulled out, because those members appear on the FEC’s Senate side. What remains is not “retirement” in any sense that means chose to stop. It is did not appear on a ballot, and no file records the difference.

The joined table, exits.csv, carries one row per sitting member per election cycle. Each row names congress, state_abbrev, district_code and election_year. on_house_ballot says whether the member appears anywhere in that cycle’s candidate file. denied and ge_win hold the primary and November verdicts where a verdict exists, and outcome is the classification assembled from the other three. Here is one member, re-elected in one cycle and absent from the next:

Read the two rows. In the first, Mike Rogers of Michigan holds his seat in the 112th Congress; the 2012 candidate file contains him (on_house_ballot is TRUE), his party did not refuse him renomination (denied is FALSE), and he won in November (ge_win is TRUE), so the outcome is re-elected. In the second he holds the same seat in the 113th Congress, and the 2014 candidate file does not contain him at all: on_house_ballot is FALSE, so denied and ge_win have nothing to record. The roster gives him no death year, and the FEC’s Senate file does not list him either, so he falls into the residual: did not seek re-election. That is everything the two files can say. Whether he chose to go, or read the district and decided not to try, is in neither.

Three cleaning decisions had to be made, and each would have produced a confident wrong answer if skipped. First, Voteview issues a member a new identifier when they switch parties mid-Congress, so a switcher reads as a departure and an arrival that never happened. Members are identified here instead by their number in the Biographical Directory of the United States Congress, Congress’s own permanent register of everyone who has served, which survives a switch; that removes 16 duplicate rows. Second, the two files share no key at all, so incumbents — the members already holding a seat — are matched once by name and thereafter by the FEC’s permanent candidate number. Of 3,936 incumbent candidacies, 3931 of 3938 (99.82%) matched, and only 18 needed anything looser than an exact surname.

Third, the FEC’s own incumbent marker is sometimes wrong and sometimes missing, so every marked incumbent must also appear in the roster for that Congress. That check dropped 7 impostor rows and recovered 6 unmarked incumbents by identifier.

One column in the roster quietly stops. occupancy, the only field recording mid-term replacements, is filled through the 114th Congress and empty after it, with no footnote and no warning:

That cliff decides the denominator, the bottom number every rate here is divided by. The natural unit would be the 435-member class a general election seats, which is exactly what occupancy would isolate and cannot after 2015. So the denominator throughout is everyone who held a House seat during a Congress: 445 people in the Congress that met in 2023, not 435 seats. Brookings counts the other way, which is one reason the two series run close without ever quite meeting.

Before you read on, commit to a number. Between 2004 and 2022, 790 members left the House. How many of them do you expect were removed by voters — beaten in November by the other party, or refused renomination by their own? Write down a count, or a share of 790, before you look at the figures.

What it says about democracy

The long view first, because the modern House is the strange one.

Figure 1. The share of the House that did not come back, every Congress from 1789 to 2023. An area on a time axis, because the quantity is a share of one chamber measured 118 times at even intervals. The finding is the shape of the whole series: a chamber that once emptied itself, and stopped. The peak is 1841, when 76 percent of the House did not return; the floor since 1900 is 1987, at 8.2 percent.

The fall is the professionalization of Congress. Polsby (1968) described the House acquiring boundaries, an internal career structure and seniority rules — an institution rather than a way station.² Members stopped going home because there came to be something worth staying for. The question this brief asks starts where that story finishes: once turnover has fallen this far, who does the removing?

Those are the 4,443 member-cycles — one sitting member facing one election — from 2004 to 2022. Four out of five members came back. Of the 790 who left, voters removed 283: 225 beaten in November, 58 denied renomination by their own party. The other 507, 64 percent of all departures, faced no voter at all. The modal way to leave the House of Representatives is to stop being a candidate for it.

Figure 2. Every House departure from 2004 to 2022, split at the line. To the left of each year, members who left without standing on a ballot they could lose; to the right, members somebody defeated. The left side totals 507 and the right side 283. Most readers place the primary share far too high: a sitting member is roughly 3.9 times more likely to be beaten in November than to be refused renomination.

The wrong lesson would be that the House is safe and nothing changed. What changed is who does the removing. Compare the earliest and latest quarter centuries of the Brookings series:

The two lines swapped. In the first postwar quarter century, November removed about 31 members a cycle and 28 left on their own; in the most recent, November removes 17 and 38 leave. That is the same fact the house-competition brief reaches from the other side: where the November result is a foregone conclusion, the seat empties only when its occupant decides to go.

If departures are decisions, they should respond to circumstances, and they do — in patterns worth telling. The first is the redistricting cycle: after each ten-year census every state redraws its House district lines, and the first election on the new map is the one held in a year ending in 2. Those elections against all the others:

Retirements barely move, and primary defeats more than double. The folk theory says a dismantled district produces a retirement; the record says a new map more often puts two members into one district, and two members in one district produce a primary. The clustering is real, and it is in the other column. The second pattern is the wave:

Figure 3. Retirements against seats won or lost, one point per party per election, 1946 to 2022. Retirement counts from Brookings; each party’s seat change computed from the roster. The horizontal bars are the mean seat change within each quarter of the retirement range: 2.9, 2.5, 0.4 and -7.6 seats.

A wave of retirements is a political signal, but only when it is a wave. Below about 21 retirements, a party’s count says nothing about how the election will go; above it, the average party loses 7.6 seats. Across all 78 points the correlation is -0.22, far too loose for the forecasting tool it is treated as on television. And the figure cannot say which way the arrow runs: members reading the polls, or parties recruiting badly in the same bad years, fit it equally well.³

What this brief cannot tell you

It never observes a decision. “Did not seek re-election” is a category of ignorance. It holds the member who wanted to leave and the member who ran out of options, and telling them apart would need a record of intent no agency keeps. Every sentence above about departure being chosen is a sentence about a category that also holds the pushed.

It cannot reach 2024 from the returns. The FEC had published no 2024 compilation when this was written, so the four-way sorting stops at 2022 while the membership series runs to 2023. The Brookings column covers 2024 and is not built from returns.

Its death rule is an approximation. The roster records a death year, not a date, so a member who left in January and died in November cannot be told from one who died in office. The rule used is conservative, and it undercounts.

And two centuries are not one series. A departure in 1841 and one in 2021 are counted identically, and they are not the same event. Multi-member districts, dissolving parties, and House careers that ran through state legislatures all sit inside the steep part of Figure 1.

What you should have learned

• Between 2004 and 2022, 790 members left the House, and 64 percent of them never stood on a ballot they could have lost; voters removed only 283.
• Of the 283 voter removals, 225 came in November and 58 in primaries, so general defeat outnumbers primary defeat by about 3.9 to one.
• Turnover collapsed long before anyone alive voted: from 42.7 percent of the chamber per Congress in the first four decades to 15.0 percent since 1980.
• Redistricting years barely move retirements and more than double primary defeats, because a new map merges members rather than erasing them.
• A retirement wave signals a bad year for the party only in the top quarter of the range; below it, the count carries no signal at all.

Extensions

None of these is answered above, and every one can be answered from the tables the Sources section hands over.

• exits_by_year.csv splits every departure into reelected, lost_general, denied, senate_run, died and not_running. Compute each one’s share of members by year. Does the chosen-versus-imposed balance hold in every cycle, or do wave years break it?
• departures.csv runs back to the first Congress with pct_left. Follow it and decide when serving in the House stopped being something people did briefly. Is the fall steady, or in steps?
• vsoc.csv carries retired, lost_primary and lost_general for every election since 1946. Group by whether year ends in 2 and check the redistricting pattern for yourself.
• compare.csv shows three sources counting the same primary defeats, with gap_brookings and gap_giroux. Find a year where they disagree and decide which source you would trust, and why.
• Stretch. When the FEC publishes its 2024 Federal Elections workbook, extend the sorting two more years and see whether the 2022 pattern held.
• Stretch. Pick one high-retirement cycle and find each retiring member’s announcement date in a news archive. Do the announcements cluster before or after the polls turned against the party?

Sources

Data

• Lewis, J. B., Poole, K., Rosenthal, H., Boche, A., Rudkin, A. and Sonnet, L. Voteview: Congressional Roll-Call Votes Database, member file HSall_members.csv. https://voteview.com/data — retrieved 2026-08-11, 51,062 rows, Congresses 1 to 119.
• Federal Election Commission, Federal Elections, biennial compilations for 2004 through 2022. https://www.fec.gov/introduction-campaign-finance/election-results-and-voting-information/ — retrieved 2026-08-11, 10 workbooks, 22,156 House candidate rows. Compiled by the Commission from the official canvasses of the state election officials.
• Brookings Institution, Vital Statistics on Congress, Chapter 2, edition of April 2026, Table 2-7 (House incumbents retired, defeated or re-elected, 1946–2024) and Table 2-9 (House and Senate retirements by party). https://www.brookings.edu/articles/vital-statistics-on-congress/
• Giroux, G. Members of Congress Defeated For Renomination, 1966–2026. Publicly shared spreadsheet; used here only as an independent check on the primary-defeat counts, on which this build, Brookings and Giroux agree exactly in 7 of 10 elections and are never more than one apart. https://docs.google.com/spreadsheets/d/1X97S6hUby6PPHAMoEwvAxkAUo8EUwLk8xMosx4hVd8c/

The data itself

Every figure above is drawn from these tables. They are plain CSV, they open in a spreadsheet, and they are yours to take further.

compare.csv, departures.csv, dropped.csv, exits.csv, exits_by_year.csv, recovered.csv, unmatched.csv, vsoc.csv, vsoc_party.csv

One more file is not data about the subject – it is how this chapter checked its own work: checks.csv, the arithmetic verified before anything was drawn.

Academic literature

• Polsby, N. W. (1968). “The Institutionalization of the U.S. House of Representatives.” American Political Science Review 62(1): 144–168. https://www.jstor.org/stable/1953331
• Jacobson, G. C. (1989). “Strategic Politicians and the Dynamics of U.S. House Elections, 1946–86.” American Political Science Review 83(3): 773–793. https://www.jstor.org/stable/1962397
• Mayhew, D. R. (1974). Congress: The Electoral Connection. Yale University Press.
• Jacobson, G. C. and Carson, J. L. (2019). The Politics of Congressional Elections. 10th ed. Rowman & Littlefield.

I want to rebuild a public dataset from its original sources, without
any starting files. Please do the following and show your work.

THE FIRST SOURCE. Voteview's congressional member file:
https://voteview.com/static/data/out/members/HSall_members.csv
One CSV, about 6 MB. Each row is one member serving in one Congress,
1789 to the present — the standard compilation of congressional
membership. No account or key is needed.

THE SECOND SOURCE. Federal Election Commission, "Federal Elections",
the agency's biennial compilation of official results. One Excel
workbook per cycle, 2004 through 2024, each pinned by a document id,
for example https://www.fec.gov/documents/5675/federalelections2024.xlsx
It carries primary, runoff and general vote totals for every House
candidate, with an incumbent flag — the only primary-election source
compiled by a federal agency.

THE THIRD SOURCE. Brookings Institution, "Vital Statistics on
Congress", Chapter 2 workbook:
https://www.brookings.edu/wp-content/uploads/2026/04/Chpt-2.xlsx
Table 2-7 counts House incumbents retired, defeated or reelected,
1946-2024 — an independent compilation to check against.

HOW TO DOWNLOAD THE FEC WORKBOOKS. The sheet names differ from edition to
edition, and one of them ships a typo the agency itself made — the
2012 House sheet is titled "Resuts", not "Results". Find sheets by
matching their names loosely, never by assuming a fixed label.

WHAT TO BUILD.
1. From Voteview: every sitting House member in each Congress since
   2003, and whether they appear again after the next election.
2. From the FEC workbooks: for each member who left, the route out —
   retired, beaten in a primary, or beaten in November.
3. The primary-defeat count per cycle, set beside Brookings Table 2-7.

CHECK YOUR WORK. The copies this book used were downloaded in August
2026. If your rebuild matches them, you should find:
- 51,062 rows in the Voteview file, covering Congresses 1 to 119
- 22,156 House candidate rows in the FEC workbooks for 2004-2022
- 58 primary defeats over 2004-2022 by this derivation; Brookings
  counts 61, never differing from it by more than 1 in any year
Voteview regenerates its file continuously and each new Congress adds
rows; a higher row count or a top Congress past 119 means the file
grew, not that you made an error.