Re: Propensity
- From: pashby@xxxxxxxxxxxxxxxxx (Peter Ashby)
- Date: Wed, 19 Jul 2006 13:28:57 GMT
Toby Kelsey <toby_kelsey@xxxxxxxxxxxx> wrote:
Peter Ashby wrote:
Philip <pp417@xxxxxxxxxxxxx> wrote:
I think your second example might also have something to do with the point
Peter B makes about statistical significance being arbitrary.
Up to a point. Empirically you can study disease distribution where a
common cause is known and this will give you distributions within
particular confidence issues when modelled. With that background, which
is what epidemiologists have, then modelling confounders included, you
can say that a particular cluster has no apparent common cause and even
put some numbers of the probability.
There is some information missing here. As well as determining whether a
correlation of cancer and known causes is statistically significant, any
study should also report its "power", that is what level of correlation it
is expected the statistics should be able to uncover (find statistically
significant). For example a study into the link between smoking and cancer
would find no significant correlation if it studied only 10 people for 1
year.
The study should contain a statement something like "a correlation between
increased radiation exposure and an increased cancer incidence of under
20% over a lifetime would not be detected by this study". As well as a
lower bound for detection an upper bound for non-detection is needed,
otherwise any lack of correlation reported is meaningless.
That is entirely true and is indeed a major problem with most cases of
'cancer clusters'. The reason a major study was launched into cases at
Sellafield (and of the children of ex Sellafield workers) was because
the numbers involved were such that the study was sufficiently powered
for a meaningful result.
Unfortunately this critical piece of information is rarely reported.
Again I agree absolutely, which is why I take all such reports in the
general media with a pinch of salt until I can check such things. Though
I have noticed that the media are becoming better at reporting the sizes
of studies.
>>> > If you ask people to randomly place dots on a piece of paper what
>>> > you get are a lot of rather evenly spaced dots. This is a good
>>> > example of how people misunderstand randomness. Another way we see
>>> > it is worry over cancer clusters, meaning epidemiologists are
>>> > forever having to explain to people that randomness tends to be
>>> > rather clumpy and as such the vast majority of 'clusters' are to be
>>> > expected.
The people are quite correct to be sceptical or worried if the power of the
study is not clearly explained.
Absense of evidence (or insufficient data for significance) is not evidence of
absense (or lack of causal link).
True, but if the link were of the size that the worried tend to worry
about most studies would pick it up. However many 'clusters' involve
very different tumours, and since we now recognize that cancer is not
one disease such 'clusters' are very unlikely to be real.
Peter
--
Add my middle initial to email me. It has become attached to a country
.
- References:
- Re: Propensity
- From: Toby Kelsey
- Re: Propensity
- Prev by Date: Re: Propensity
- Next by Date: Re: Propensity
- Previous by thread: Re: Propensity
- Next by thread: Re: Propensity
- Index(es):
Relevant Pages
|
