Thursday 17 March 2005

Lies, damn lies, and... ach, you know.

John B is discussing the infamous Lancet report:

If you don't accept that the 100,000 number from the Lancet study on Iraq war causalties represents a probable lower bound (given its exclusion of Falluja, where we appear to have killed everyone) on the number of Iraqis who died in the 18 months following the war and otherwise wouldn't have died in the 18 months following the war, and you do not have a PhD in a statistical discipline, then you are an ignorant bigot

I don't like to boast about my qualifications (mainly because I don't think that qualified equals intelligent), but I'll mention them this once, since it's in context. I have a B.Sc. in Mathematics and Logic & Philosophy of Science. No, it's not a Ph.D., but I still feel qualified to say what I'm about to say.

Statistics, when applied to situations containing significant elements dependent on consciousness, is a load of guesswork and outright bollocks dressed up in fancy and misleading terminology. If you're talking about the movement of molecules in a vacuum, statistics is accurate. If you're studying the spread of disease, it's useful. If you're studying the flow of calls through telephone exchanges, it's extremely useful, but, as any call centre manager will tell you, quite fallible. If you're studying elections, it tells you that Labour will win the 1992 election. If you're studying crime, it tells you that the Washington sniper is white and working alone — something the stats experts that the FBI call "profilers" insisted was true right up till he got caught. In my considered opinion, the number of deaths caused in a war is a subject inherently unsuited to this type of study. I'm aware that professional statisticians believe that everything in the world is a fit subject for their discipline, but they're wrong, and I'm far from being the only scientist who thinks so.

The way you find out how many people have died in a war is by counting the bodies and accounting for the ones you can't find, by talking to every family affected, not just some of them. It takes ages, so won't satisfy the impatience of the anti-war crowd, but it is at least accurate. Pretending that death by gunfire is sufficiently similar in nature to death by bubonic plague that its prevalence can be studied in the same way is not accurate.

No, I haven't read the Lancet study, and have no intention of doing so. I don't need to read it to accept its finding that there is a 95% probability that the number of deaths falls between 8000 and 198000, which seems reasonable, although I'd want to add the caveat that, in this context, "probability" is dependent on all sorts of parameters distinct to this study and is not the same thing as "probability" when applied to dice-rolling. I'm aware of the statistical laws which say that the real figure is more likely to be around the mid-point of the range than its extremes, and believe those laws to be misapplied in any case which involves, as war does, such a large degree of intention and direction.

If the Iraqi Army and the terrorists and the civilian population had all been moving randomly while the Coalition forces fired randomly, then it'd be time to call the statisticians. Back in reality, get counting.


Yes, I know, of course a confidence interval isn't a probability, strictly speaking, and I hereby slap my own wrist for using the terms interchangeably. However, for the purposes of my argument, it is the same type of thing as a probability, in that its meaning and predictive power get all shot to hell when you bring human volition into the equation.

No comments: