Click to get your own widget

Monday, June 23, 2008

HOW TO GET BIG RESULTS FROM SMALL FIGURES

Scotland on Sunday did a news item on how researchers had issued a report about how smoking causes all sorts of previously unnoted diseases.

The new study, which has been published in the journal Annals of Oncology, was carried out by a team led by experts at Glasgow University and was based on data from 17,363 male civil servants based in London.

Information about their health and habits has been collated since the 1960s in an effort to gain information about health trends and find links between lifestyle and illness. The original link between smoking and lung cancer was found through similar analysis of medical data.

The study found:

• A 43% increase in the chances of dying from cancer of the colon if the person smokes.

• A 40% higher likelihood of dying from rectal cancer.

• An increase of 23% in the chances of losing one's life to prostate cancer.

• A 53% rise in mortality from lymphatic leukaemia among smokers.

The study concluded: "Cigarette smoking appears to be a risk factor for several malignancies of previously unclear association with tobacco use."

Dr David Batty, of the Medical Research Council Social and Public Health Sciences Unit, based at the University of Glasgow, said: "What this study shows is that smoking is linked to more kinds of cancer than previously thought.


I put up a reply saying in detail why this is a complete abuse of statistics & thus science. I am pleasantly surprised to see that all most all other commentators understood this too. I am reprinting it here because such "studies" are very common space fillers in the media & the same basic reasons why they are fraudulent apply to most of them.

There is tremendous pressure on researchers to find new & PC causes of death & it is always possible to find blips in statistics since they necessarily have random variations particularly where the numbers are small.

In general good scientists don't declare a result even worth publishing if it doesn't show a change in risk of 100% (known as a relative risk of 1) where the sample is small.

Lets look at the figures.

The study was on "data from 17,363 male civil servants based in London"

& found

" A 43% increase in the chances of dying from cancer of the colon if the person smokes"

How many of those would be smokers - say 7,000

What proportion of people die of cancer of the colon - I'm guessing but say 1 in 2,000.

So that would mean 3.5 dying.

A 43% increase would be 5 people.

Since it is rare for only .5 of a person to die we are talking about 1 death more than average which is obviously well within statistical variability.

The other figures are similarly nonsense.

It is mathematically certain that there will be similar statistically pointless connection to a less than normal death rate from other diseases.

This would not have even been published in a medical journal let alone got its authors newspaper publicity if they had found a similar "connection" between voting Green & dying from falling off fridges (another small but measurable cause of death).

Comments:
Every beginning course in Statistics will cover Poisson's analysis of deaths by horse-kicks in the Prussian artillery regiments (or some more boring equivalent). It follows that people who publish rubbish of the sort you allude to are well aware of what they are doing.
 
Post a Comment

<< Home

This page is powered by Blogger. Isn't yours?

British Blogs.