The Star Wars Map: A Whole Survey in One Chart

In 2014, FiveThirtyEight surveyed 1,186 people about Star Wars. Buried in it is a beautiful little problem: they asked each respondent how they felt about 14 characters — Han, Luke, Leia, Vader, Yoda, Jar Jar, and the rest — on a favorable / unfavorable / unfamiliar scale.

Cross-tabulate character against rating and you get an 84-cell wall of numbers. Run the classic test on it and you learn exactly one thing:

• χ² = 4,900.7 (df = 65), p < 0.0001 — the ratings are not independent of the character.
• Cramér's V = 0.29 — a solid, real association on a 0–1 scale.

Great — but a significance test won't tell you which characters are alike, who the outliers are, or the shape of the disagreement. For that you stop reading the table and start mapping it. The tool is Correspondence Analysis (CA) — the categorical cousin of PCA — and it's exactly what that wall of numbers was hiding.

Step 1 — Load it, reshape it, run it

Survey exports have a quirk: the header is two rows deep. Row 1 holds each question once — “how do you feel about these characters?” spans 14 columns — and row 2 says which character each column is. Drag the file onto Stratum and it spots the second row on its own, keeping it as a column identifier above each column instead of mistaking it for a respondent named Han Solo. Rename that identifier Character in the Navigator.

CA wants two categorical columns, not 14 side-by-side ones. Give the ratings their order once (Column Definition ▸ Order, applied to all 14 columns, Very favorably down to Very unfavorably), then the one bit of wrangling: select the 14 character columns and use Data ▸ Stack Columns with Eliminate missing rows ticked. The character identifier comes along as its own column, so fourteen wide columns become one tall, tidy pair — Character and Favorability, 11,529 ratings. Then Analyze ▸ Correspondence Analysis, Rows = Character, Columns = Favorability. No knobs to turn.

The report breaks all that χ² into dimensions, ordered by how much of the association each explains:

Two dimensions capture 95% of everything in the 84-cell table. That's the whole promise of CA: a wall of numbers collapses to a picture you can read.

Step 2 — Read the map

Right-click the report → Drill Down to Correspondence Map. Two clouds of points — the characters and the rating levels — where proximity means association (a character near a rating got that rating more than chance predicts).

Dimension 1 (76%) asks one question: adored, or not? Very favorably sits at one end; at the other sits everything short of it — Very unfavorably, but also Neither and Unfamiliar.

• The adored cluster: Han, Luke, Leia, Obi-Wan, Yoda and R2-D2 bunch tightly together beside Very favorably — 66–74% of the people who rated each gave them the top mark. C-3PO is with them but a step toward the middle (57%): well liked, not quite adored.
• The far side: Palpatine, Boba Fett, Lando and Padmé — far from adored, and mostly not because they are hated. Plenty of people simply shrug.

Dimension 2 (19%) pulls that far side apart. At one end, Unfamiliar and Neither: Lando, Padmé and Boba Fett drift there, with roughly one respondent in five not knowing them at all — CA recovered a fame axis nobody told it about. At the other end, Very unfavorably, and two characters:

• Darth Vader is divisive, not hated. He sits near the middle of Dimension 1 because 38% gave him the top mark — and 18% the bottom one. He contributes more to Dimension 2 than any other character.
• Jar Jar Binks is the one the survey truly dislikes: 25% chose Very unfavorably, more than for anyone else, and he sits farther from the center of the map than any other character. The math had one job, “find what's most unusual,” and it pointed straight at Jar Jar.

That is the whole pitch — it hands you the structure you didn't know to look for.

Step 3 — The same trick finds your segments

This isn't a Star Wars trick — it's how you surface structure in any survey: customers, employees, voters, patients. So point it at the pettiest question in the file: "Who shot first — Han or Greedo?" against a single column, "Do you consider yourself a fan?"

Cramér's V = 0.40 — the strongest association anywhere in the survey. Same three options, two populations living in different universes, and one chart draws the border between them. (And the film that actually won? Not the one with the marketing — The Empire Strikes Back took 289 first-place votes, the clear favorite. The data has taste.)

When to reach for Correspondence Analysis

The rule of thumb is simple:

When your data is categories, you can't run PCA — and most people give up and stare at a contingency table. CA is the tool you were missing: it takes the χ² you already trust and turns it into a map you can actually read. One reshape, four clicks, and an 84-cell table becomes a single chart a whole room understands.

That's the difference between a chart and an analysis.