How to Lie With Statistics

Chris Johnson, University of Glasgow.

I recently bought three books that have changed the way I think about statistics and probability:

Darrell Huff's How to Lie with Statistics, Penguin, 1954 (ISBN 0-14-013629-0).
William Dembski's The Design Inference: Eliminating Chance Through Small Probabilities, Cambridge, 1998 (ISBN 0-521-62387-1).
Daniel Hausman's Causal Asymmetries, Cambridge Studies in Probability, Induction and Decision Theory, Cambridge, 1998 (ISBN 0-521-66289-1).

These are three very different publications. The first book is a lighthearted discussion of the ways in which statistics are abused by the popular media. The second and third deal with causation and, in particular, look at ways in which we can distinguish events that form causal sequences from events that simply occur together by chance.

In this talk, I argue that statistical methods have been abused when they have been applied to analyse the causal factors behind human behaviour. Most human computer interaction research rests on dubious correlations and not on causal connections. If time allows, I will try to explain how this recent work on the statistics of design and on causal asymmetries may provide better methods for understanding human computer interaction (and operator error).