'How to Lie with Statistics' - Book Review

It's always great to find a book that deserves its stellar reputation: How to Lie with Statistics by Darrell Huff was first published in 1954, more than sixty years ago, but chances are very good you've heard the title and it's still in print (or at least available new on Amazon).

He demonstrates a variety of ways to lie with statistics: one of the first he presents is choosing amongst the mean, median, and mode while simply referring to it as "the average" - a method that can throw a statistic into a radically different light if you don't bother to ask which "average" you were just handed. One of my personal favourites is graph-shortening, in which you present the range on a graph from say 900 to 1000 and then claim it to be wildly active because the value has risen (or dropped) by the entire shown value - whereas if you'd shown from 0 to 1000 it would have been clear there wasn't much fluctuation at all. Then of course there's post hoc conclusions: "this fluctuates at the same time as that, therefore this causes that." Actually, you usually have no proof at all of any connection: just because they both moved doesn't mean there's a direct association ("correlation does not imply causation" - I don't think he quoted that, but happily that's often referenced now). He concludes by pointing out that even with the best of intentions, statistics remains as much an art as a science and may deceive without even intending to.

The book is short, well written, and quite entertaining: apparently it was a university introductory stats text for many years. I wish my textbooks were this readable! Highly recommended.