?

Log in

No account? Create an account
LiveJournal lies in their survey bar charts - Blather, Rinse, Repeat
September 21st, 2012
03:40 pm

[Link]

Previous Entry Share Next Entry
LiveJournal lies in their survey bar charts
livejournal-poll-of-the-day

A while ago, LJ was running a survey about gun control or something. The numbers reported might have been accurate, but the bar chart they displayed representing the results looked skewed to me. They're still doing it, as the image above shows.

The bar that represents 4816 votes is 160 pixels long, which dominates the chart. Your eye sure goes to that, doesn't it? And look, all of the rest of the results are tiny, not making up a noticeable fraction of the responses, right?

Well, sort of. What's the second biggest percentage? Did you say 14.5%? Whoops, no, it's actually 15.6%, which is listed fourth. For some reason, the bars aren't sorted. Which maybe isn't a big deal...

Except for the fact that the bars are the wrong size. Let's assume that somebody decided that the most room they had for a graph was 160 pixels worth of bar. That's totally fine, charts have widths, you start somewhere. The longest bar, sure, that's going to be the 4816 / 63.9% bar. The next bar should be the 14.5%, sorry, the 15.6% bar, with 1175 respondents. Doing the math... that ought to be 39 pixels wide, instead of the 9 pixels it is now. Huh.

Similarly:
1093 / 14.5% should be 36 pixels instead of 9
332 / 4.4% should be 11 pixels instead of 9
124 / 1.16% should be 4 pixels instead of 9

I'll concede that 11 pixels instead of 9 isn't a big deal, but 39 pixels instead of 9 and 4 pixels instead of 9 both seem like pretty malicious distortions of the numbers.

Now, maybe the designer wants to pipe up here and say "oh, but look, we've got these rounded corners on our bar graphics, they would look bad below 9 pixels, that's why we adjusted things". To which I reply, if your graphics are forcing you to distort the math, then change the graphics. Charts should be made as pretty as possible without distorting the data.

Looking at the amount of error in each case (or, if you want to be cranky, "lie fraction"), you see:

4816:    0%
1093:  -75%
 332:  -18%
1175:  -77%
 124:  118%


meaning that we've got bars that are about a quarter the size they should be, and a bar that's more lie than truth. Or, hey, let's display "truthiness", by subtracting the absolute value of the above percentage from 1:

4816: 100% truthy
1093:  25% truthy
 332:  82% truthy
1175:  23% truthy
 124: -18% truthy


Or, hey, because this is fun, here are some charts.

The correct values
corrected

a "truthiness" graph
truthiness

(Leave a comment)

My Website Powered by LiveJournal.com