Tuesday 20 August 2013

Choropleth madness

Hat-tip to @briantimoney for sharing this too-good-to-be-true piece from Business Insider just as I'm about to set off for two weeks of carto-conversation in Germany at the International Cartographic Association conference then the UK at the Society of Cartographer's Summer School and the British Cartographic Societies Symposium. 

MAPS: A Poll Asked America Which States Were The Drunkest, The Hottest And Which Had The Silliest Accent


A poll of 1603 people were asked to answer questions about US States that weren't their own and the responses are amazing....actually, it's not the responses that are amazing, it's the fact that someone thought they could fashion any sort of sense out of the results and then map them into meaningless gross generalisations.

Here's one...



Brilliant...but merely one of 22 maps in the report.  Twenty-two...all apparently giving us an insight into the American psyche

Yes, that's 22 maps based on a sample that represents 0.000005% of the American population. Were they even representative of the population in socio-economic terms? Were they spread evenly across the 52 States (31 people per State) or was the survey stratified so that more populated States had more respondents?  Or was it just a random interweb survey of whomever could be bothered to click through a mundane survey? Were responses based on first-hand knowledge or just stereotypical perceptions?  Bad science all round in relation to the survey.

And the maps...well, each one is beautifully crafted in Web Mercator (not equal area...which of course creates visual problems) and each has a separate colour (really nice design touch). But wait...there's no legends and we are expected to imply that darker hues mean 'more'.  More of what? Totals? yes...yes...they look like totals.  Twenty-two non-normalised choropleths in a single article.  This has to be some sort of record. I can't be certain because there's no details but I would wager serious wonga on these maps all being totals which of course renders them useless when comparing results across a map.

All in all...a sham of a mockery of a sham of a mockery. Bad science. Bad survey. Bad cartography. Bad reporting. And why get so worked up about it?...because only a couple of hours after being published, this piece of utter crap has been viewed nearly 17,000 times, liked on Facebook over 1,000 times, re-tweeted over 200 times etc etc.  Viral choropleth madness! How many of those people have looked at the maps and seen that the distributions are based on a poorly framed survey and because they're not normalised they mean nothing?  I bet it's something like 0.000005% of them (oh, that's less than 1 person...but near enough I guess) 

UPDATE: About 24hrs after going live the post had got over 300,000 views, 30,000 Facebook likes, 1700 tweets etc. A week later...690,000 page views. Nice visibility for poor maps!

UPDATE 2: The author, Walter Hickey has contacted me with details of the survey.  It was stratified so at least there is some sense to the distribution of respondents by State.  I'd still claim (or whine as someone mentioned below) that 1400 people is not a particularly good sample. I had to do 1500 surveys for my PhD in a town with a population of 210,000 to get a statistically significant sample.  Not sure how 1400 in a population of 310,000,000 can in any way be representative.  Regardless, the maps are still wrong so it's a moot point.

4 comments:

  1. Hey, wish you reached out, I'd be happy to share the data, etc. with you if you wanted to personally verify its substance. Shoot me an email at whickey@businessinsider.com, I'd love to chat about the methodology.

    ReplyDelete
  2. Walter, thanks for commenting. This blog is intended as a 'reach-out'...constructive criticism. No way anyone could have reached out to you before publication. I'm not querying the substance of the data you've collected...merely what's been done with it.

    ReplyDelete
  3. I don't understand why people whine about "sample size" without actually doing the statistical analysis to back it up. 1400 people is a very good sample size for the population of the US (as the author of the piece clarified, margin of error of 3 with a 95% CI).

    ReplyDelete
  4. Awesome talk here in Dresden at #iccDD2013 this Saturday afternoon. Seems like great challenges ahead for quality work (not just only in the cartography domain).

    ReplyDelete