Skip to content

Scores vs. Ranks — Don't Confuse Ranks and Ratings

The single most important distinction in ballot design — and the one people get wrong most often.

→ See it, don't just read it: Alternate ballot styles — one voter, three ballots shows the same voter's opinion on a ranking ballot (read by RCV-IRV, STV, or Ranked Robin), a Yes/No ballot (Approval), and a 0–5 scoring ballot (Score/Range, STAR) — side by side, with what each captures and throws away.

There are two fundamentally different ways a ballot can ask for your opinion, and they are not interchangeable:

  • Ranking (ordinal): put the candidates in order — 1st, 2nd, 3rd. This captures relative preference: "I like A more than B." It says nothing about how much.
  • Rating / Scoring (cardinal): give each candidate an independent value — 0–5 stars, A–F. This captures absolute preference: "I like A this much." It carries both the order and the strength.

A rank answers "which do you prefer?" A score answers "how much do you like each one?" Those are different questions, and a ballot that asks one cannot be silently treated as if it asked the other.

The map — from one voter's opinion to a method

Read it in two steps, and keep the two steps separate: first the ballot (what the voter is asked to record), then the tabulation (what the count does with it). Nearly every taxonomy error comes from mixing those two levels.

flowchart LR
    P["<b>Voter preferences</b>"]

    P --> RATED["<b>RATED ballot</b> — cardinal<br/><i>score each candidate independently<br/>· equal scores allowed · blank = 0</i>"]
    P --> RANKED["<b>RANKED ballot</b> — ordinal<br/><i>order only, never degree<br/>· 'RCV' names THIS, not any one method</i>"]

    RATED --> ADD["<b>Single-round · additive</b>"]
    RATED --> RUN["<b>Score, then runoff</b>"]
    ADD --> SCORE["Score / Range Voting<br/><i>highest total wins</i>"]
    ADD --> APPR["Approval<br/><i>= Score on a 0–1 scale</i>"]
    ADD --> MJ["Majority Judgment<br/><i>median, not sum</i>"]
    RUN --> STAR["<b>STAR Voting</b><br/>R1 Scoring — sum, top two advance<br/>R2 <b>Automatic Runoff</b> — majority<br/>preference between those two"]

    RANKED --> COND["<b>Condorcet / pairwise</b><br/><i>beats-all winner · summable</i>"]
    RANKED --> ELIM["<b>Sequential elimination</b><br/><i>eliminate &amp; transfer · not summable</i>"]
    RANKED --> POS["<b>Positional</b><br/><i>points by rank position</i>"]
    COND --> RR["<b>Ranked Robin</b> (RCV-RR)<br/>Ranked Pairs · Schulze · Minimax"]
    ELIM --> IRV["<b>RCV-IRV</b> — single-winner<br/><b>STV</b> — multi-winner, proportional"]
    POS --> BB["Borda · Bucklin"]

Four things this diagram is careful about, because popular versions of it usually aren't:

  • STAR's second round is a runoff, not a ranking. It counts, for each ballot, which of the two finalists that voter scored higher — a majority preference, not an ordering of the field. Calling it "ranking" feeds the very confusion this page exists to clear up.
  • Both sides branch by tabulation. Whether a ballot permits equal ranks is a ballot rule, not a method family — most US RCV-IRV forbids them, Ranked Robin allows them, and the same is true of truncation limits. Those belong as footnotes on the ballot, not as branches. (What defines IRV is sequential elimination, not its equal-ranks rule.)
  • Condorcet is a family, not a method. Ranked Robin sits inside it, alongside Ranked Pairs, Schulze and Minimax — it isn't a sibling of "Condorcet."
  • A ranked ballot is not one method. IRV, Ranked Robin and STV read the same ballot and can elect different winners. The full ranked breakdown — with aliases and the summability split — is the canonical ranked-method family tree.

(Single-winner unless marked; STV is the proportional multi-winner branch. The rated side has proportional forms too — STAR-PR — and a majoritarian multi-winner one, Bloc STAR.)

The naming trap. Rank and rate sound almost identical and get swapped constantly — but they are opposites in what they measure. This library uses rank only for ordering and score (synonyms: rate, grade) only for independent values. If you remember one thing: equal rankings are still rankings; a score is not a ranking.

Relative vs. absolute preference

Ranks (ordinal) Scores (cardinal)
The question "Which do you prefer?" "How much do you like each?"
Captures Order only Order + strength
Utility type Ordinal utility Cardinal utility (interval scale)
Differences meaningful? No — 1st vs. 2nd ≠ 2nd vs. 3rd Yes — a 5 vs. a 3 is a real gap
Familiar from A race: 1st, 2nd, 3rd Amazon / movie five-star reviews
Methods RCV-IRV, STV, Condorcet (Ranked Robin), Borda STAR, Score, Approval (0/1)

Because differences are meaningful only for scores, you can average scores (a column sum) but it makes no sense to "average" ranks. This is exactly why STAR is summable and IRV is not. (See STAR Is Summable — Add Up Precinct Totals and How the Count Works — STAR vs RCV-IRV, Step by Step.)

Why "ranked" is about the data, not just the ballot

A common point of confusion: Ranked Robin allows equal rankings, so why isn't it a scoring method? Because the method only ever reads the order (who beats whom), never a magnitude. Allowing ties in a ranking doesn't turn order into strength. The line between ranked and scored is what information the tabulation uses — pure order (ranked) vs. degree of support (scored) — not merely how the bubbles look.

But how the bubbles look changes what voters think they're doing. That definition is about the tabulation; the ballot is about the human. A row of bare numerals — 0 1 2 3 4 5 — invites people to fill it in like a ranking: first choice, second choice, third, each number used once. Equal Vote's STAR ballot design standards exist largely to fight that instinct: label the ends Worst / Best and draw the scale as star icons, explicitly to reinforce the star scale as opposed to ranking. Stars say rate each candidate on its own, reuse values freely; bare numerals say put them in order. The tabulation is identical either way — but a ballot that reads as a ranking gets filled in as one, and that discards precisely the intensity information the method exists to capture. Ballot design isn't decoration here; it's what makes the ballot honest.

Why the distinction has real consequences

Different winners from identical voters. A candidate who wins under a ranked method can lose under a scored method on the very same electorate, because the two are measuring different things. Treating "ranked = scored" hides that.

Expressiveness — preference vs. support. A rank carries order only; a score carries order and how strongly you feel. The sharpest way to see it: ballots 1,0,1,0 and 5,4,5,4 state the same preference but opposite support, yet as ranks they're the identical A=C > B=D — the ranking literally can't tell "I tolerate them" from "I love them." That gap is its own page: Preference vs. Support. A broadly-liked compromise candidate can look weak on first-choice ranks yet clearly strong on scores — which is part of why IRV suffers center squeeze and STAR doesn't.

You can go one way but not the other. Scores → ranks is easy (just read off the order). Ranks → scores is impossible without inventing information, because the strength was never collected.

Ballot reliability. Independent scoring removes the "grid" constraints of ranking (no overvotes, no skipped-rank traps, ties allowed, blanks = 0). Reported spoilage rates run roughly 0–2% for rated ballots vs. 4–9% for ranked ballots (and 1–4% for single-mark) — but read those with care. The compilation is rangevoting.org, which advocates for rated methods, and the rated figure rests on a thin base: almost no public election has ever used a rated ballot, so there is little real-world data to average. The ranked end is the better-evidenced half — Neely & McDaniel's San Francisco work found real RCV spoilage, with racial disparities (see the scorecard). Safe version of the claim: ranked ballots demonstrably spoil more than single-mark, and rated ballots have structurally fewer ways to go wrong; the precise percentages are not settled. And research on values measurement (Maio, Roese, Seligman & Katz) found ratings showed superior validity to rankings, because forcing a full strict order captures noise — distinctions voters don't actually feel.

Learn more