Your P-value is in another castle

February 1, 2016 6:59 AM

Guess the Correlation: The aim of the game is simple. Try to guess how correlated the two variables in a scatter plot are. The closer your guess is to the true correlation, the better.
posted by Cash4Lead (25 comments total)

What, no negative correlations?
posted by leotrotsky at 7:07 AM on February 1, 2016 [1 favorite]

Could anyone tell me how closely related this is to the dance of the p-values?
posted by clawsoon at 7:07 AM on February 1, 2016 [2 favorites]

I got 55.

If you get to 100 points you level up to Guess the Causation, which is much tougher.
posted by leotrotsky at 7:11 AM on February 1, 2016 [13 favorites]

This slightly different version is the one my Biostats professor had us play with. They show you four plots and four values and you match them up. It's a little easier than the original link (and it does include negative values, leo).
posted by Ufez Jones at 7:12 AM on February 1, 2016 [2 favorites]

I use Ufez Jones' link every semester in my Statistics class. Students develop pretty good intuition about correlation after playing with it for just a few minutes.
posted by wittgenstein at 7:19 AM on February 1, 2016

98. The high correlation ones are easy. Telling the difference between 0.6 and 0.4 is harder.

Are they assuming a 1:1 ratio? Some cases where the apparent line wasn't at 45 degrees seemed to give much worse scores than I thought they would...
posted by YAMWAK at 7:21 AM on February 1, 2016

Pretty easy game, but a great idea.
posted by demiurge at 7:39 AM on February 1, 2016

45 -- dang. I wonder how much easier it would be to draw a scatter plot of 10 dots to match a given r value.
posted by klausman at 7:50 AM on February 1, 2016

Is this like a test where if it kind of sounds like a fun idea you automatically score 75+ on the nerdiness scale?
posted by Segundus at 8:01 AM on February 1, 2016 [1 favorite]

clawsoon: this game is not actually related to p-values at all.
posted by splitpeasoup at 8:03 AM on February 1, 2016 [2 favorites]

Telling the difference between 0.6 and 0.4 is harder.

Turns out I often guess p <0.3 for values in that range. I am now wondering what opportunities for fame and profit I have stared at and dismissed in my career.
posted by Dr Dracator at 8:11 AM on February 1, 2016 [3 favorites]

I got bored and quit at 105, but I look at lots of scatterplots.

Could anyone tell me how closely related this is to the dance of the p-values?

They're mostly unconnected. The only connection I can see is that for clear relationships (correlations closer to 1 or -1) it would be hard to draw a (reasonably-sized) sample with a high p-value, while p-values for vague relationships are likely to be more widely distributed. You can play with a similar idea with the excel sheet that video links to by just setting the effect size higher.

But vague (closer to zero) correlations don't imply high or particularly variable p-values. You can have vague relationships you're quite confident about just by having larger samples. You can also play with this in the excel sheet by setting the N to maximum permitted value of 200.
posted by ROU_Xenophobe at 8:22 AM on February 1, 2016

I'm not very good at this. I can class into "high, mid, low" ok, and seem to do alright with the numerical value on the high ones (.7 and above), but I tend to underestimate for the mid and low.

I'm a theorist though, and not the sort that looks at data very frequently. Is that why I'm bad at this?
posted by nat at 8:37 AM on February 1, 2016

Kinda wish there were more misleading ones, a la Anscombe's quartet.
posted by Westringia F. at 8:39 AM on February 1, 2016 [12 favorites]

This is kind of fun. I am hopeful that a two-player version of a game like this will play a similar role for my child as GORILLAS.BAS did for me when I was ~10.
posted by mean square error at 8:51 AM on February 1, 2016 [2 favorites]

I find it basically impossible to tell any two charts with a less than 0.40 value apart from each other, but I find it very easy to get within 0.05 on charts with a greater than 0.50 value....
posted by 256 at 8:52 AM on February 1, 2016

I work in visual arts (with minimal stats training) and hit 177 before bombing out from boredom. Perhaps in the particular circumstances of this game, a knack for spatial relationships happens to be a skill that allows one to complete on the level of a person who works with these things all day.

That, or I've made a terrible career choice.
posted by oh hey at 8:54 AM on February 1, 2016 [2 favorites]

Needs more bimodals.
posted by benzenedream at 9:21 AM on February 1, 2016

203
posted by The Tensor at 11:19 AM on February 1, 2016

Helpful hint: don't add a decimal point before your estimate or you will get 0 no matter what you type after it.
posted by TedW at 11:37 AM on February 1, 2016

240. Good to know I have gotten at least one skill out of teaching intro stats.

(and thanks for the link, Ufez Jones, I'm going to email that out to my class right now!)
posted by egregious theorem at 11:39 AM on February 1, 2016

If they're recording stats on people's guesses, then they'll be able to look for correlation between the value of the correlation and the accuracy of people's guesses.
posted by RobotHero at 11:43 AM on February 1, 2016 [3 favorites]

On my third try I thought I could play forever but then I got a couple of hard ones in a row. The 0.2-0.5 range seems hardest, I was often guessing low. 311.
posted by ikalliom at 3:41 PM on February 1, 2016

Yeah the low correlations are the least intuitive: it's actually pretty easy to get a correlation of say, 0.3 that looks like complete crap with just a couple more points in the top-right corner, and pretty subtle tweaks can really determine whether you land closer to 0.2 or 0.5. Which I guess is part of the instructive value! (231 here)
posted by en forme de poire at 5:02 PM on February 1, 2016 [1 favorite]

I'm showing this to my econometrics students. I'm the cool TA!

...

no i'm not but i hope someone likes it anyway
posted by dismas at 10:48 AM on February 2, 2016 [1 favorite]