Every candy in the dataset
Scatter plot of 85 candies: sugar percentile on the horizontal axis, win rate on the vertical axis.
The contestants
Each dot is one of 85 Halloween candies. FiveThirtyEight showed people random pairs — Twix or Skittles? Snickers or candy corn? — about 269,000 times, and recorded which one they'd rather find in their trick-or-treat bag. A candy's height on this chart is its win rate: the share of matchups it won. Left to right is how sugary it is compared to the others.
The obvious theory
If sugar drove candy love, this trend line would climb steeply. It doesn't. The correlation between sugar and win rate is r = 0.23 — sugar explains about five percent of why one candy beats another. Going from the least sugary candy in the set to the most sugary buys you roughly ten points of win rate, while the gap between the best and worst candies is over sixty.
Exhibit A
Meet Reese's Miniatures: the second most-loved candy of all 85 — and the third least sugary. It sits in the 3rd percentile for sugar and wins 82% of its matchups. If sweetness were the point, this candy could not exist.
The real divide
Now color the dots by one ingredient. Chocolate candies (brown) average a 60.9% win rate; everything else averages 42.1%. That's an 18.8-point gap from a single yes/no attribute — nearly double sugar's effect across its entire range. And every one of the top ten candies contains chocolate.
The payoff
Chocolate is worth +18.8 points of win rate. It beats every other attribute — crisped rice, peanuts, bar form — and it holds up in a regression controlling for all of them (+19.7). Fruity and hard candies actually lose points. Sugar isn't what we're choosing. Chocolate is.
What this data can't tell us
- "Sugar" here is a percentile, not grams. The dataset ranks candies against each other; it can't say anything about absolute sweetness, and all 85 candies are already sweet. This is "more vs less sugary among candy," not "sweet vs unsweet."
- Preference in a survey ≠ purchases. People voted on which they'd rather receive, online, in 2017, in a US-centric sample. Sales data could tell a different story.
- n = 85, and attributes overlap. Chocolate candies are also more likely to be bars with peanuts, so the smaller regression coefficients are noisy. I only lean on the chocolate effect, which is consistent across raw gaps, correlation, and regression.
- Two of the "candies" are coins. FiveThirtyEight included One Dime and One Quarter as controls. I kept them because voters really saw them; dropping them changes no conclusion.
- Correlation, not causation. Chocolate predicts winning; this data can't prove it causes it. Maybe chocolate brands just market better.