xkcd

by **josiahstevenson** » Wed Apr 06, 2011 4:16 am UTC

'So, we did the study again and found no link. It was probably a --' 'RESEARCH CONFLICTED ON GREEN JELLY BEAN/ACNE LINK; MORE STUDY RECOMMENDED'

I always was suspicious of the green ones...

by **Coffee Stain** » Wed Apr 06, 2011 4:18 am UTC

It looks like there's a mistake in panel two. It probably should read without the word 'green', or else it sort of contradicts the panel with the "Whoa." Threw me off for a little bit as to what the joke was.

by **Sizik** » Wed Apr 06, 2011 4:19 am UTC

Shouldn't the first study be "WE FOUND NO LINK BETWEEN JELLY BEANS AND ACNE", not "WE FOUND NO LINK BETWEEN GREEN JELLY BEANS AND ACNE"?

Edit: ninja'd

by **nhaines** » Wed Apr 06, 2011 4:19 am UTC

So, uhh... was panel 2 supposed to specify a color?

EDIT: niiiinnnnnjjaaaaaas!

by **rho421** » Wed Apr 06, 2011 4:19 am UTC

I love how little it seems that the general population understands statistics...

by **Daggoth** » Wed Apr 06, 2011 4:20 am UTC

but miiiiiinecraft!

by **atomfullerene** » Wed Apr 06, 2011 4:21 am UTC

Was going to post a more extensive reply...but minecraft....

by **Poohblah** » Wed Apr 06, 2011 4:22 am UTC

but miiiiiiiiinecraft!

by **josiahstevenson** » Wed Apr 06, 2011 4:22 am UTC

Actually, to be nitpicky, shouldn't it read "(P < .05)" in each case? You reject the null when the p-value is LESS than significance level you were looking for, I thought...

by **madock345** » Wed Apr 06, 2011 4:23 am UTC

I wonder if this takes the prize for most panels in an XKCD comic (I know there have been comics larger than this, but I am refering specifically to the number of panels)

Yay! Color!

by **gooley** » Wed Apr 06, 2011 4:25 am UTC

Great to see he's back to writing entertaining and professional comics.

Edit: s/threads/comics/ (wtf was I thinking?)

by **jakerman999** » Wed Apr 06, 2011 4:27 am UTC

There are also two panels with yellow(which have the same result, not the contradicting big green/little green scenario).

Quick, someone search for secret codes!

by **phlip** » Wed Apr 06, 2011 4:31 am UTC

josiahstevenson wrote:Actually, to be nitpicky, shouldn't it read "(P < .05)" in each case? You reject the null when the p-value is LESS than significance level you were looking for, I thought...

Right, so when p > 0.05, they don't reject the null hypothesis... thus "we found no link". The one panel where they do find a link does read "(p < 0.05)".

by **josiahstevenson** » Wed Apr 06, 2011 4:34 am UTC

ah, there we go.

by **kateract** » Wed Apr 06, 2011 4:36 am UTC

phlip wrote:
josiahstevenson wrote:Actually, to be nitpicky, shouldn't it read "(P < .05)" in each case? You reject the null when the p-value is LESS than significance level you were looking for, I thought...

Right, so when p > 0.05, they don't reject the null hypothesis... thus "we found no link". The one panel where they do find a link does read "(p < 0.05)".

In a (statistically?) significant number of cases, a good way to think of the null hypothesis is the "status quo" and the alternative as the "research" hypothesis. I tend to think of the p-value as "the probability that this variation occurred randomly as apposed to an unknown outside force" It's also the reason we "fail to reject" the null hypothesis instead of "accepting" the alternative hypothesis, similar to the "correlation does not imply causation" line.

by **wackojacko1138** » Wed Apr 06, 2011 4:37 am UTC

Say it with me kids:
"The p-value represents the likelihood of the evidence given the null hypothesis, not the likelihood of the null hypothesis given the evidence."
If you know, understand, and live that sentence, then you understand statistics better than most journalists and some scientists.

by **neoliminal** » Wed Apr 06, 2011 4:43 am UTC

Coffee Stain wrote:It looks like there's a mistake in panel two. It probably should read without the word 'green', or else it sort of contradicts the panel with the "Whoa." Threw me off for a little bit as to what the joke was.

WHOA... you are using the worlds most ubiquitous coffee stain as your avatar!! It threw me off your comment a little bit.

But you're right. I wonder if he'll change it.

by **thenonsequitur** » Wed Apr 06, 2011 4:48 am UTC

I love the idea behind this comic, with the 1/20 confidence and exactly 20 color tests with one different result.

But I really hate that the panel two broke the whole thing for me by including the word "green" when it obviously shouldn't have. At first I actually had no idea what the point was because panel two threw a curve ball at me. It took me a while to realize that was just a mistake.

by **FCN** » Wed Apr 06, 2011 4:50 am UTC

So, do green jelly beans cause or cure acne? Scientists! Give me a sign.

by **Diadem** » Wed Apr 06, 2011 4:51 am UTC

Unfortunately, this is exactly how many scientists (not all, luckily, but too many) do statistics. They really do tests on dozens of variables at once, with several different statistical tests, and then publish the ones that are interesting.

For a single experiment you should always divide your allowed margin of error by the number of variables you are testing for. So if you test n different colours of jelly beans, you must require p < 0.05/n before considering a result statistically significant. But this is often ignored.

by **tremodian** » Wed Apr 06, 2011 4:53 am UTC

There is also the Familywise Error Rate (FWER) to consider given how many tests are being performed.

by **Dason** » Wed Apr 06, 2011 4:56 am UTC

Diadem wrote:For a single experiment you should always divide your allowed margin of error by the number of variables you are testing for. So if you test n different colours of jelly beans, you must require p < 0.05/n before considering a result statistically significant. But this is often ignored.

You're recommending a Bonferroni adjustment in every situation? There are lots of cases where you can get away with something a lot less conservative. Even in the most general case you could at least use Holm's method which is slightly better than Bonferroni.

Plus it's not always the case that you care about controlling the family wise error rate. You might be more interested in controlling the false discovery rate in which case you'd be looking at a different set of methods for adjusting your p-values.

by **chaucer345** » Wed Apr 06, 2011 4:59 am UTC

See, what they should have done this is run a single factor ANOVA with the colors as the levels and then performed a Tukey MCP to determine which of the levels were significantly different from each other... Or would that have helped? Maybe I'm thinking of this the wrong way.

by **Peristarkawan** » Wed Apr 06, 2011 5:06 am UTC

This demonstrates the fallacy that scientists refer to as "jelly bean picking".

by **Eternal Density** » Wed Apr 06, 2011 5:06 am UTC

I posted this to the Minecraft forum before I could continue reading at panel 2.

by **Kugala** » Wed Apr 06, 2011 5:08 am UTC

This would also work for a really good excuse to eat a ton of Jelly Beans, which is how I read it first actually.

by **Peristarkawan** » Wed Apr 06, 2011 5:12 am UTC

Diadem wrote:For a single experiment you should always divide your allowed margin of error by the number of variables you are testing for. So if you test n different colours of jelly beans, you must require p < 0.05/n before considering a result statistically significant. But this is often ignored.

Or you do your initial experiment merely to look for possible correlations that warrant further study. Then when you've identified a variable to investigate, you do an entirely new experiment, ignoring the earlier data entirely, and you only publish the results of the second experiment.

by **scarletmanuka** » Wed Apr 06, 2011 5:12 am UTC

madock345 wrote:I wonder if this takes the prize for most panels in an XKCD comic (I know there have been comics larger than this, but I am refering specifically to the number of panels)

The 2008 Christmas Special had more panels, though a lot of them were just black.
XKCD Loves the Discovery Channel had the same number of panels as this comic, and far less repetition.

by **ethereal_fire** » Wed Apr 06, 2011 5:40 am UTC

Would have helped me a lot when I took stats in second year. I COULD DO ALL THE BLOOODY MATH. but I got that stupid little p-value and couldn't remember for the life of me if it was p > 0.05 or p < 0.05 that meant the data showed a significant link. So I could get the number just didn't know what it meant. :'( BUT ALL THE STUPID QUESTIONS WERE MULTIPLE CHOICE :'( ... managed to get a 58 in th course still though. Soooo iritating cause I know how to do it all just it was just the yes or no bit that I failed to retain. ANYWAY If I'd had this I would have had no problem at all this comic could have saved my life! er... mark...

phlip wrote:
josiahstevenson wrote:Actually, to be nitpicky, shouldn't it read "(P < .05)" in each case? You reject the null when the p-value is LESS than significance level you were looking for, I thought...

Right, so when p > 0.05, they don't reject the null hypothesis... thus "we found no link". The one panel where they do find a link does read "(p < 0.05)".

SEE!!! THIS IS WHY I COULDN'T REMEMBER WHAT BLOODY WAY IT WENT. I needed some one to say < = Good, > = Bad. All this rejecting the null hypothesis explaination just became a scrambled jumble in my brain and left me very confused and not sure which way things went.

also, now I REALLY need to try minecraft.

by **iChef** » Wed Apr 06, 2011 5:41 am UTC

This is amazing. I'm going to go cut a sliver off a green jelly bean and dissolve it in a 55 gallon drum of water. Who wants to buy my new miracle acne medicine?? It is guaranteed to work 95% better than the leading acne medication by SCIENCE!!!!!!!!!

by **madock345** » Wed Apr 06, 2011 5:45 am UTC

scarletmanuka wrote:
madock345 wrote:I wonder if this takes the prize for most panels in an XKCD comic (I know there have been comics larger than this, but I am refering specifically to the number of panels)

The 2008 Christmas Special had more panels, though a lot of them were just black.
XKCD Loves the Discovery Channel had the same number of panels as this comic, and far less repetition.

Hmmm... so do black panels count? if not its tied. I say we have a vote! (or maybe Mewtwo will tell us)

by **ethereal_fire** » Wed Apr 06, 2011 5:48 am UTC

madock345, I like your sig : )

by **deskjethp** » Wed Apr 06, 2011 5:49 am UTC

jakerman999 wrote:There are also two panels with yellow(which have the same result, not the contradicting big green/little green scenario).

Quick, someone search for secret codes!

Decipherization:

1: Number of letters in the color listed in each panel = n

EDIT: Either 2a)

Take the nth letter of of the nth word in that panel wrapping around from the front of the word if necessary

EDIT: or 2b) Forget what step 2 is because you should be asleep and take the panelnumber-3th letter of the nth word [I remember taking the 20th letter of orange for the last r...at what point in the series did I screw up?]

(EDIT: The right combination of this) results in "penklndeelonkeebetwr" which anagrams to "blender week ken lepton"

Thus: Soon in the future there will be a "blender week" series featuring a guy named Ken attempting to blend leptons [possible WIB? reference?] despite the inherent nature of leptons to not be very interactive [low blendability]

by **madock345** » Wed Apr 06, 2011 5:59 am UTC

ethereal_fire wrote:madock345, I like your sig : )

I enjoy recursive signatures.

by **SteevyT** » Wed Apr 06, 2011 6:15 am UTC

ethereal_fire wrote:
also, now I REALLY need to try minecraft.

It will consume all your time.

It seems so boring, but you just can't stop.

Need....to....play.....Minecraft.....posting....so....much....work......

by **Qwerty.55** » Wed Apr 06, 2011 6:22 am UTC

madock345 wrote:
ethereal_fire wrote:madock345, I like your sig : )

I enjoy recursive signatures.

I like your new recursive sig even better than before.

by **MathUhhhSaurus** » Wed Apr 06, 2011 6:24 am UTC

My copy and pasting senses are tingling!

by **CorruptUser** » Wed Apr 06, 2011 6:38 am UTC

As an actuary, I approve this comic.

Also, my goals this weekend are to start Minecraft and Dwarf Fortress.

by **StClair** » Wed Apr 06, 2011 6:39 am UTC

but miiiiiinecraft!

by **CorruptUser** » Wed Apr 06, 2011 6:45 am UTC

Diadem wrote:Unfortunately, this is exactly how many scientists (not all, luckily, but too many) do statistics. They really do tests on dozens of variables at once, with several different statistical tests, and then publish the ones that are interesting.

For a single experiment you should always divide your allowed margin of error by the number of variables you are testing for. So if you test n different colours of jelly beans, you must require p < 0.05/n before considering a result statistically significant. But this is often ignored.

I believe you use the F-test when you have multiple variables.

Also, does anyone know if N=20 and P=1/20 is enough to use Poisson approximation?

Who is online