When Numbers Lie

Sources:

Causality: Models, Reasoning, and Inference

Internet Project

Visualizing Urban Data ideaLab

Can Quantum correlations be Explained Casually

Images:

Casimir Effect

Cosmic Microwave Background

Atmospheric Carbon Dioxide

Music in this video (from the YouTube Audio Library):
Fortaleza
Good Starts
Zydeco Piano Party

[emanonmax]

So numbers don't lie. But we can misunderstand them

[seanld]

Just found your channel today (December 12th, 2017), and I'm binge-watching. You're videos are exactly the kind I like. Intellectual and informative about topics I enjoy.

Thanks for your videos! You're one of my new favorite YouTubers.

[PicotryOfficial]

"There are three kinds of lies: lies, damned lies, and statistics." - Benjamin Disraeli

[TimothyChapman]

3:11 "(hopefully...)" - That is not good enough for me.

4:26 Great example. And I think that this type of paradox is far more common than what we've been led to believe (accidentally or otherwise).

[jesswma]

I really enjoy your videos. Please keep making them, I'm sure your fan base will grow a lot in the next little while.

[theo]

I'm so glad I discovered your channel! Thank you for making these videos, the animations explain well and I've learnt new and interesting things. :)

[MrJones]

Thanks for the explanation. I have read the wikipedia page multiple times but never really understood the simpsons paradox. This video cleared it all up!

[meganc]

You deserve more subscribers! Just saw your channel because of the response video you made to veratasium, cool stuff

[npip]

Hm, perhaps important to note is that the odds of getting 150 of the same gender on one half of the randomly selected group, is less than

If you select from a group purely randomly, you wont be exposed to Simpson's Paradox with any reasonable level of possibility. And the existence of the possibility of Simpson's Paradox has no effect on the data, other than perhaps giving a bimodal distribution which would give you a poor SD if you assume it's unimodal (Which, a poor SD is not a bad thing, it just means you have the opportunity to make your deductions stronger by splitting by gender and then analyzing).

The case of Berkeley was not random, of course.

[kryp]

Wait, why were the charges of discrimination dropped when they realized they were discriminating against men?

[LarlemMagic]

So, don't hate me, but doesn't simpson's paradox apply to the wage gap?

[TheViolaBuddy]

The problem with this is that how do we know what subsets (male/female in the drug example and departments in the school example) to look at? We can't possibly split every experiment's people into, "people who prefer brie over camembert" vs. "people who prefer camembert over brie, " or "people who have at least three vowels in their first name" vs. "people who have fewer than three vowels in their first name" just in case cheese preference or vowel count might have an effect on something. It seems like, in general, there's going to be a lot of guesswork to decide how to break apart the sample into groups - and we might have a bunch of bad results from experiments that do so incorrectly.

[achronicstudent]

I have to prepare a presenatation about data science and just got your work. Thanks you saved my life, I may be thief but thank you anyway

[jonathangumpangkum]

so with all that said, numbers don't lie just as much as cameras don't lie

[luchogallardoleon]

I´ve seen this happening when comparing sales goals' fulfillment rate,

[RaymondHng]

I wonder if Simpson's Paradox is the reason behind the voting results of the 2016 Eurovision Song Contest. Australia was leading in the Jury vote with the most points and Russia led in the Televote (public vote). But when the points from the Jury and Televote were added together, Ukraine won overall.
Place Televoting Points Jury Points
1 Russia 361 Australia 320
2 Ukraine 323 Ukraine 211
3 Poland 222 France 148
4 Australia 191 Malta 137
5 Bulgaria 180 Russia 130

[karrotsrkool]

There is no paradox here you just added the fractions incorrectly. For example if you wish to add 10/50 and 90/150 you don't do (10 + 90) / (50 + 150) = 100/200 = 50% you first have to make the denominator equal 10/50 + 90/150 = 30/150 + 90/150 = (30 + 90)/(150 + 150) = 120/300 = 40%. Simple fractions can slip you up but the fault in your maths is odvious as the average of 20% and 60% is not 50% and like wise for 30% and 70% not being 40%.

[LordVulcanus]

And so, when you look at the actual data instead of the totals, the gender pay gap... Disappears!

[SnoFitzroy]

me clicking this video: hell yeah spill the tea on number drama

[Tetracarbon]

How come you didn't just mention you can use a multi linear regression to control for gender using a binary variable? Like, I understand you you didn't have to go into OLS to make your point (well done, most of the audience would have run away screaming and unsubscribing) but any researcher worth salt with a first year stats course would just stick in a control variable.

Or, realistically, probably a whole tone of control variables.