The Court from Left to Right
33 justices ranked two ways — by how often they voted liberal, and by how far that put them from the colleagues in the room. The two lists are not the same list.
Data through the Supreme Court’s 2024 term, from the Supreme Court Database, 2025 Release 01. Shares are percentages of a justice’s ideologically coded votes; differences from colleagues are in percentage points.
By raw share of liberal votes it is William O. Douglas at 78.8%, with William Rehnquist last at 30.0%. Measured against the justices who decided the very same cases, the answers change: Thurgood Marshall comes first and the second Justice Harlan comes last. Douglas sat on a liberal Court and Harlan sat on one too — which is exactly why the raw number, on its own, partly ranks the eras rather than the judges.
Two rankings, not one
The first column is the one people mean: the share of a justice’s ideologically coded votes that the Supreme Court Database records as liberal. It is the honest answer to the question as usually asked, and it has a real flaw. The Court chooses its own docket, and that docket changes: a bench hearing criminal-procedure appeals in 1962 and one hearing administrative-law disputes in 2012 will produce different percentages from identical judges.
The second column narrows that problem sharply, though it does not remove it. It is the gap — the justice’s liberal share minus the share for the other justices deciding the same cases. Same docket, same day, same participating colleagues. A positive number means voting liberal more often than the people in the room; a negative one, less often. Zero is not neutrality. It is agreement with the bench of the day.
Neither is the truth on its own. Read them together, and where they disagree, give the second more weight about the justice and the first more weight about the era.
And read both with their error bars, which is what the ± figures in the table are. They are not decoration. Of the 32 steps from one justice to the next in the table below, four are larger than the uncertainty in the two figures they separate — and three more sit within a tenth of the line, so even that count is not firm. The order is real at a distance and noise up close, and the section below the table says how far apart two justices have to be before the ranking can tell them apart.
The 33 Justices With a Long Enough Record, Most to Least Liberal
A justice needs 500 ideologically coded cases to appear here, so that a handful of terms cannot outrank a thirty-year tenure. Seven further justices are covered by the database but fall short of that and are named under the table. The first figure column is a share of that justice’s coded votes; the second is a difference in percentage points, positive meaning more liberal than the participating colleagues and negative less. Each carries twice its sampling error.
Ordered by raw liberal share. The last column is the rank by gap instead, with the arrow showing how far a justice moves between the two orderings.
| # | Justice | Liberal votes | Difference from colleagues (pts) | Rank by difference |
|---|---|---|---|---|
| 1 | Douglas OT1946–1975 · 3,707 cases | 78.8% ±1.4 | +21.6 ±1.4 | 2↓1 |
| 2 | Warren OT1953–1968 · 2,100 cases | 74.6% ±1.9 | +12.5 ±1.3 | 9↓7 |
| 3 | Black OT1946–1970 · 3,151 cases | 73.0% ±1.6 | +14.3 ±1.4 | 7↓4 |
| 4 | Fortas OT1965–1968 · 525 cases | 72.4% ±4.0 | +9.1 ±2.7 | 12↓8 |
| 5 | T. Marshall OT1967–1990 · 3,537 cases | 72.0% ±1.5 | +23.5 ±1.2 | 1↑4 |
| 6 | Brennan OT1956–1989 · 5,087 cases | 71.2% ±1.3 | +18.4 ±1.0 | 3↑3 |
| 7 | Sotomayor OT2009–2024 · 1,091 cases | 64.6% ±2.9 | +17.4 ±2.2 | 4↑3 |
| 8 | Kagan OT2009–2024 · 973 cases | 61.9% ±3.1 | +14.5 ±2.1 | 6↑2 |
| 9 | Stevens OT1975–2009 · 4,019 cases | 60.4% ±1.5 | +14.1 ±1.3 | 8↑1 |
| 10 | Ginsburg OT1993–2019 · 2,162 cases | 60.4% ±2.1 | +14.8 ±1.5 | 5↑5 |
| 11 | Souter OT1990–2008 · 1,690 cases | 57.2% ±2.4 | +11.5 ±1.6 | 10↑1 |
| 12 | Breyer OT1994–2021 · 2,168 cases | 57.0% ±2.1 | +11.2 ±1.5 | 11↑1 |
| 13 | Clark OT1949–1966 · 2,132 cases | 54.1% ±2.2 | −7.3 ±1.5 | 17↓4 |
| 14 | Blackmun OT1969–1993 · 3,561 cases | 51.8% ±1.7 | +2.7 ±1.3 | 13↑1 |
| 15 | Stewart OT1958–1980 · 3,429 cases | 50.7% ±1.7 | −6.6 ±1.3 | 15 |
| 16 | B. White OT1961–1992 · 4,701 cases | 48.8% ±1.5 | −5.1 ±1.0 | 14↑2 |
| 17 | Frankfurter OT1946–1961 · 1,721 cases | 46.8% ±2.4 | −11.4 ±2.0 | 24↓7 |
| 18 | Minton OT1949–1956 · 658 cases | 46.0% ±3.9 | −8.5 ±2.7 | 19↓1 |
| 19 | Reed OT1946–1956 · 1,085 cases | 45.6% ±3.1 | −11.1 ±2.3 | 23↓4 |
| 20 | Vinson OT1946–1952 · 762 cases | 45.1% ±3.7 | −8.8 ±2.5 | 20 |
| 21 | Burton OT1946–1958 · 1,327 cases | 44.5% ±2.7 | −13.4 ±2.0 | 25↓4 |
| 22 | Whittaker OT1956–1961 · 641 cases | 43.5% ±3.9 | −18.9 ±2.9 | 31↓9 |
| 23 | Harlan II OT1954–1970 · 2,190 cases | 43.2% ±2.1 | −21.6 ±1.7 | 33↓10 |
| 24 | J. Roberts OT2005–2024 · 1,405 cases | 41.9% ±2.6 | −7.5 ±1.6 | 18↑6 |
| 25 | Kennedy OT1987–2017 · 2,740 cases | 41.8% ±1.9 | −7.2 ±1.3 | 16↑9 |
| 26 | R. Jackson OT1946–1953 · 804 cases | 39.8% ±3.5 | −14.3 ±2.8 | 26 |
| 27 | Powell OT1971–1986 · 2,383 cases | 39.3% ±2.0 | −10.5 ±1.3 | 22↑5 |
| 28 | O'Connor OT1981–2005 · 2,798 cases | 38.8% ±1.8 | −10.1 ±1.3 | 21↑7 |
| 29 | Scalia OT1986–2015 · 2,777 cases | 35.3% ±1.8 | −14.5 ±1.3 | 27↑2 |
| 30 | Alito OT2005–2024 · 1,361 cases | 34.3% ±2.6 | −16.5 ±1.9 | 28↑2 |
| 31 | Burger OT1969–1985 · 2,692 cases | 34.2% ±1.8 | −16.9 ±1.2 | 29↑2 |
| 32 | Thomas OT1991–2024 · 2,667 cases | 32.2% ±1.8 | −17.2 ±1.4 | 30↑2 |
| 33 | Rehnquist OT1971–2004 · 4,300 cases | 30.0% ±1.4 | −20.4 ±1.0 | 32↑1 |
7 more justices appear in the database with records too short to rank — Barrett, Goldberg, Gorsuch, K. Jackson, Kavanaugh, Murphy, W. Rutledge.
Where the two lists disagree
Four justices chosen not for the size of the move but for what the move shows: in each case the choice of measure changes what you would say about them. Several others shift as far or further.
The largest disagreement on the page. A 43.2% liberal share puts Harlan in the middle of the pack; measured against the justices who sat with him he is the single most conservative record in the database, 21.6 points below his own bench. He spent his career dissenting from the Warren Court, and a raw percentage cannot see that — because the Warren Court decided liberal cases liberally, and Harlan was in the room for all of them.
Nearly the same story and rarely told: 43.5% raw, 18.9 points below the Court he sat on, which moves him from 22nd to 31st. His tenure was short and unhappy, and the raw figure flatters it.
The cleanest illustration of why one number is not enough. Clark voted liberal in 54.1% of coded cases — above the halfway mark, and 13th of 33. Against the justices deciding the same cases he is 7.3 points to the conservative side. Both numbers are correct. Only the second tells you where he sat on his own Court.
The reverse case. Kennedy’s 41.8% is one of the lower raw shares here, which reads as firmly conservative; against his own colleagues he is 7.2 points under, and rises nine places to 16th. He was the swing vote of a conservative Court, and only the relative measure shows it.
How much of this order is real
A ranking of 33 people implies 32 claims: that each justice sat to the left of the next one. Each figure behind those claims is an estimate from a sample of cases, so each carries an error — the ± in the table, twice the standard error from resampling that justice’s own cases. Counting a step as real only when the difference exceeds twice the two errors combined:
- By raw share — 4 of 32 steps clear the line, 3 more within a tenth of it.
- By difference from colleagues — 4 of 32, 3 more within a tenth.
Those exact integers should not be trusted, and the borderline counts are why. The error bars come from resampling, and a handful of steps sit so close to the threshold that the count moves depending on where the resampling lands — it was measured at four different sample sizes up to thirty thousand draws and did not settle. What does not move is the shape: almost none of the steps are real.
That sounds like it demolishes the table. It does not, and the reason is worth stating plainly: 461 of the 528 possible pairings of these justices are clearly ordered. The ranking is reliable at a distance and meaningless up close. Two justices six places apart in the colleague-adjusted order are always genuinely ordered; two standing next to each other almost never are. On the raw-share order it takes eight places.
Read the ends as findings and the middle as a cluster. That the gap column needs a shorter distance than the raw one is not a coincidence: resampling a case moves a justice and his colleagues together, so differencing tends to cancel the noise they have in common. Its error bar is the narrower one for all but one of these justices. It is the same reason that column is the better measure.
The Other Order
The same 33 justices ranked only by distance from their own bench — the five furthest to each side.
- T. Marshall+23.5 · 72.0% raw
- Douglas+21.6 · 78.8% raw
- Brennan+18.4 · 71.2% raw
- Sotomayor+17.4 · 64.6% raw
- Ginsburg+14.8 · 60.4% raw
Method, and its limits
Built from the Supreme Court Database (2025 Release 01, justice-centered), covering the 1946 through 2024 terms. One vote per justice per case. Cases the database declines to code ideologically are excluded. A justice needs 500 coded cases to be ranked at all, which is why 7 of the 40 covered here do not appear in the table.
These are the 1946-onward figures, not whole careers. The database also publishes a legacy file running back to 1791, and eight of the justices covered here were already sitting before 1946 — Hugo Black and Stanley Reed from 1937, Felix Frankfurter and William O. Douglas from 1938. Their full records are on their own profile pages and differ, sometimes sharply: Frankfurter voted liberal in 53.1% of coded cases across his whole tenure and 46.8% of them from 1946. A ranking of one period has to be built from that period, so this page uses the 1946-onward record for everyone in it, and says so rather than quietly mixing the two.
“Liberal” is the database’s word, not ours, and it is a term of art rather than a political score: pro-union in labor cases, anti-business in economic cases, pro-federal-power in federalism cases, and — in the database’s own deliberate usage — pro-United-States in tax cases. Criminal procedure, civil rights, the First Amendment, due process, privacy and attorneys share a single combined definition running to seventeen clauses, so the label carries less precision in those areas than a ranking implies. The full definitions are on the codebook page.
A ranking is a stronger claim than the data supports at the margins. Adjacent places are often separated by a fraction of a point. Stevens and Ginsburg both round to 60.4%. Where the published figures tie, the ranking is settled on the exact unrounded rate (Stevens 60.4379 to Ginsburg 60.4070), which is precise but not meaningful. Treat the ends as real and the middle as a cluster.
The identification problem, which is the serious objection. Michael Bailey’s argument against reading vote-derived measures across time is that a shift in the numbers cannot be told apart from a shift in the cases: when conservative votes rise, it may be the justices who moved or it may be what the Court agreed to hear. Treating such scores as comparable across decades, he argues, requires assuming the distribution of case characteristics is constant, which it is not. His own remedy is to bridge the Court to actors outside it — Congress, the presidency — so that case content and ideology can be separated. Bailey, “Is Today’s Court the Most Conservative in Sixty Years?”, Journal of Politics 75 (2013): 821–34; see also his “Measuring Ideology on the Courts” (2017), which also reports a separate objection held by some scholars — that percent-conservative measures like the first column here are partly circular, since votes build a measure of ideology that is then used to explain votes.
Both columns inherit the first problem. The gap column narrows it, because the comparison group heard the identical docket, but it does not remove it: it controls for which cases arrived, not for what a justice was being asked in them.
Questions
Who is the most liberal Supreme Court justice?
It depends which question you are asking. By raw share of liberal votes, William O. Douglas — 78.8% across 3,707 ideologically coded cases from 1946 to 1975. Measured instead against the justices who decided the same cases, Thurgood Marshall comes first, 23.5 points more liberal than his own colleagues. Douglas sat on a more liberal Court than Marshall did, and a raw percentage cannot tell the two apart.
Who is the most conservative Supreme Court justice?
By raw share, William Rehnquist, at 30.0%. Against his own colleagues, the second Justice Harlan — 21.6 points to the conservative side of the Warren Court he served on, the widest conservative gap in the database. (The widest gap in either direction belongs to Thurgood Marshall, 23.5 points the other way.) Rehnquist sat with conservatives; Harlan largely did not.
Does a low liberal-vote share mean a justice was conservative?
Not on its own. The share depends heavily on which cases the Court chose to hear, and those change decade by decade. Anthony Kennedy and John Roberts both record low raw shares while sitting close to the centre of their own benches; Tom Clark records a share above 50% while sitting to the right of his. That is why both measures appear side by side here.
What does the Supreme Court Database mean by a "liberal" vote?
It is a term of art defined issue by issue, not a political label — pro-union in labor cases, anti-business in economic cases, pro-federal-power in federalism cases, and pro-United-States in tax disputes. Criminal procedure, civil rights, the First Amendment, due process, privacy and attorneys all share a single combined definition. The coding is Harold Spaeth’s and his successors’, and scholars contest parts of it.
Companion pages: how far each justice moved across a career, and who voted with whom. Every Court before 1946 — and the seven whose justices cannot be ranked at all — is at Every Court, Left to Right.