Can you see a correlation in a scatterplot?
Yes, within limits, and this correlation perception test measures yours. In a 2014 study of the same two-scatterplot task, the typical participant needed the correlations to differ by about 0.08 near r = 0.5 to pick the stronger one three times out of four. The test then asks a second question, from Rensink and Baldridge’s 2010 study: how strong a correlation looks to you. You set that by eye first; how your settings compare with theirs and with the real values comes on the result sheet.
This test measures that with the method researchers use. Two cakes each carry exactly 100 chocolate pearls in a scatterplot, at correlations between 0.3 and 0.8; you tap the one whose pearls line up more. Ten rows follow, each pair about 16% closer than the row above, and row 5 is set at the typical person’s 75% point from Harrison and colleagues’ data. The first miss gets a second pair on the same row; the second miss ends the chart. Your finest row is converted by simulation to the same 75% threshold, so you can see where you stand among the 179 people in that study. Three quick slides then ask you to set a cake halfway between two others, the bisection method from Rensink and Baldridge, and draw how strong correlations look to you.
It is not a test of intelligence, eyesight or statistics, only of how your eye reads a cloud of dots. One short chart of about nine pairs can land a row or two either side of where you would settle over many runs, and the pearls are bigger than the 2-pixel dots in the study. Screen size and viewing distance matter too, so a few runs tell you more than one.
What the test measures
One skill, measured like an eye chart: the smallest difference in correlation your eye can still tell apart, plus how strong correlations look to you.
The task
- Two scatterplots of exactly 100 points, drawn as chocolate pearls on cakes: pick the more correlated one, the task from Harrison, Yang, Franconeri and Chang (2014)
- Ten rows, each pair’s difference 2^(−1/4) of the row above, at base correlations from 0.3 to 0.8
- Two misses: the first gives a second pair on the same row, the second ends the chart
- Three halfway slides, Rensink and Baldridge’s bisection: set a cake halfway between no correlation and a perfect line, then halve each half
Your result
- Your row (1–10) and the difference in r you can tell apart near r = 0.5
- Where you stand among 179 online participants from Harrison et al.’s open data: the share of people you are sharper than
- Your own perception curve next to the typical one from Rensink and Baldridge’s 20 observers (a reference, not a ranking)
- Real correlations to compare with, from Meyer et al. (2001): adults’ height and weight correlate at r = 0.44
- A share card and a challenge link; the link carries your row and your halfway setting, in the part after the #
Before you start
- About three minutes: about nine pairs, then three slides
- A bigger screen helps; small cakes are harder to read
- Free, no account, no sound needed; your browser needs 3D (WebGL) switched on
Correlation perception FAQ
Is this test free? Do I need to sign up to see my result?
Yes, it’s free, and there’s nothing to sign up for: no account, no email, no payment. Your result appears on this page the moment you finish, and you can share it or play again straight away.
How good are people at judging correlation in scatterplots?
Better with strong correlations than weak ones. In Harrison et al.’s 2014 data, the typical person needed a difference of about 0.14 to tell correlations apart near r = 0.3, about 0.08 near r = 0.5 and about 0.04 near r = 0.8. How strong a correlation looks is a separate question: Rensink and Baldridge put it to 20 observers in 2010 with the same halfway slides this test ends on, and the result sheet sets your curve beside theirs once you have made your own settings.
How is the percentile in the correlation test worked out?
Your finest row is converted by simulation to the 75%-correct threshold Harrison et al. measured, then placed among the 179 participants in their open data, everyone who finished included. The comparison is on the same task with 100-point plots, though your run is one short chart while theirs were four long staircases each, so treat the percentile as approximate.
What does a correlation of 0.44 look like?
Adults’ height and weight correlate at about r = 0.44 in US data summarised by Meyer et al. (2001), and the result sheet lets you see it: tap that pair and a cake takes on exactly that correlation, with how strong it looks to most people beside how strong it looks by your own slides. That comparison waits until after the slides, so your halfway settings are your eye’s alone.
Why are strong correlations easier to tell apart?
Because the eye judges how tightly the dots hug a line, and a tight cloud changes more visibly than a loose one. Harrison et al. described this with Weber’s law; Kay and Heer’s 2016 re-analysis of the same data recommended a log-linear model instead of a straight Weber line, and that is the version this test uses.
Why did I get a different row the second time?
One short chart of about nine pairs is noisy and can land a row or two either side of your usual level. Screen size, viewing distance and the size of the cakes on your screen also move the result. Several runs give a fairer picture than one.
Does the correlation test measure statistics skill or intelligence?
No. The correlation perception test measures only how your eye reads a cloud of dots. It says nothing about intelligence, eyesight or knowledge of statistics.
Where the numbers come from
Percentiles come from Harrison et al.’s open per-participant data: a log-linear fit to the scatterplot staircases gives the typical JND at each correlation, and each of the 179 participants’ offset from it forms the comparison group. The halfway curve uses Rensink and Baldridge’s mean scale from 20 observers and is never ranked. The site does not publish its own players’ data.
- Harrison, Yang, Franconeri & Chang (2014), IEEE Transactions on Visualization and Computer Graphics — Ranking Visualizations of Correlation Using Weber’s Law
- Harrison et al. open data (TuftsVALT/ranking-correlation), re-analysed by Kay & Heer (2016), Beyond Weber’s Law
- Kay & Heer (2016), IEEE Transactions on Visualization and Computer Graphics — Beyond Weber’s Law: A Second Look at Ranking Visualizations of Correlation
- Rensink & Baldridge (2010), Computer Graphics Forum — The Perception of Correlation in Scatterplots
- Meyer et al. (2001), American Psychologist — Psychological testing and psychological assessment: A review of evidence and issues (Table 1)
How it is scored
Just-noticeable difference (JND)
The difference in correlation you pick correctly 75% of the time, the criterion Harrison et al. used. Your finest passed row is converted to it by simulation.
Rows on a log scale
Each row’s difference is a multiple of the typical person’s JND at that correlation: 2× on row 1, falling by a quarter-octave a row to about 0.42× on row 10. Row 5 is exactly the typical 75% point.
Percentile
Each of Harrison’s 179 participants has an offset from the typical JND, the same at every correlation; your offset is placed among theirs.
Bisection
Setting a plot halfway between two others, three times, traces how strong each correlation looks to you. It is drawn, not scored.