[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Cohen's Fallacy
My response to the paper quoted by Rad health below was published
in J. Radiol. Prot. 21:64-65;2001. It is reproduced below following the
paper it responds to in the same journal.
Several times Rad health has attacked my work by quoting published
papers by Field et al and by Lubin, but he has never noted my
published responses to these papers. Is that proper behavior for a
scientist? Here he has done it again; can that be accidental?
On Sun, 27 Jan 2002, Rad health wrote:
>
> J. Radiol. Prot. 20 (June 2000) 221-222
>
> LETTER TO THE EDITOR
>
> Reply to `Explaining the lung cancer versus radon exposure data for USA
> counties'
>
> Sarah Darby and Richard Doll
> Clinical Trial Service Unit, University of Oxford, Nuffield Department of
> Clinical Medicine, Harkness Building, Radcliffe Infirmary, Oxford OX2 6HE,
> UK
>
> Professor Cohen states in his letter that his analysis `encompasses all of
> the Doll suggestions'. It is, however, logically impossible for it to have
> done so using data at the level of counties. This is because the effect of
> cigarette smoking on the relationship between residential radon and
> individual lung cancer risk will be determined by the relationship between
> smoking status and lung cancer among the individuals within each county.
> Unless smoking is irrelevant to lung cancer risk (which we know to be
> untrue) or smoking status and residential radon are uncorrelated within each
> county (which seems unlikely), the relationship between residential radon
> and lung cancer at the county level will differ from that at the level of
> the individual in a way that cannot be overcome by including corrections for
> smoking habits at the county level, even if these corrections correctly
> represent the smoking habits of the individuals within each county. The
> difference in the relationship between a risk factor and a disease rate at
> the level of the individual and at an area level is the ecologic fallacy and
> is described in detail by Greenland and Robins (1994) and Morgenstern
> (1998). Lubin (1998) has also demonstrated that biases caused by the
> ecologic fallacy can be of any magnitude from minus infinity to plus
> infinity.
>
> In two recent studies (Lagarde and Pershagen 1999, Darby et al 2000),
> parallel individual and ecological analyses have been carried out of
> identical data from case-control studies of residential radon (Peshagen et
> al 1994, Darby et al 1998). These analyses have shown that, in addition to
> any bias caused by the ecological fallacy, ecological studies of residential
> radon and lung cancer are also prone to biases caused by determinants of
> lung cancer risk that vary at the level of the ecological unit concerned. In
> these two examples, the additional variables were latitude and urban/rural
> status respectively. The explanation of these variables is not yet well
> understood and they may well be, in part, surrogate measures for some
> aspects of the subjects' smoking history not accounted for by the measures
> of smoking status that have been derived from the individual questionnaire
> data and used in the analysis of the data for individuals. They had only a
> minor effect on analysis at this level but a substantial effect on the
> ecological analyses. The presence of these variables is further evidence of
> the pitfalls of ecological studies.
--The following is my published response to the above.
J. Radiol. Prot. 21 (March 2001) 64-65
LETTER TO THE EDITOR
Radon exposure and the risk of lung cancer
Dear Sir
Our study found a very strong negative correlation between lung cancer and
radon exposure in USA counties
(Cohen 1995), a discrepancy with the positive correlation predicted by
linear-no-threshold theory (LNT) of
over 20 standard deviations. Drs Darby and Doll (2000) suggested that an
explanation for this discrepancy
may lie in the ecological fallacies. My purpose here is to consider that
issue.
The classical `ecological fallacy' arises from the fact that the average
exposure does not, in general,
determine the average risk, as is obvious for situations where there is a
threshold. But we avoid this problem
by designing our study as a test of LNT - in LNT, the average exposure
does determine the average risk (e.g.
person-Sv determines the number of deaths).
Darby and Doll suggest that the ecological fallacy applied to corrections
for smoking may be important. But
following BEIR-IV, we use separate risk factors for smokers and for
non-smokers which eliminates this
problem if we assume different average radon exposures for these; this was
shown to have little effect on our
results (Cohen 1998, 1995).
However, the ecological fallacy does arise for other confounding factors
(CFs) - the average value of a CF
does not, in general, determine its confounding effects. Darby and Doll
note that Lubin (1998) demonstrated
mathematically that the error in assuming otherwise can be infinite.
For example, consider annual income as a CF that might confound the radon
versus lung cancer relationship -
maybe very poor people have lower radon and, for unrelated reasons, have
higher lung cancer rates than
others. A problem with ecological studies is that average income is not
necessarily a measure of what fraction
of the population is very poor. A case-control study does much better; in
principle, it selects cases and
controls of matched incomes.
Our approach to this problem is to use a large number of CFs. For the
example under discussion, we use as
CF the fraction of the population in various income brackets, <$5000/year,
$5000-$10 000/year, ..., >$150
000/year (10 intervals in all). In addition, we consider combinations of
adjacent brackets, and other related
characteristics such as the fraction of the population that is below the
poverty line, the percentage
unemployment, etc. It can be shown that a necessary (but not sufficient)
condition for a CF to have an effect
on our discrepancy with LNT is that it have a strong correlation, at least
25%, with radon levels, and none of
the above CFs have a correlation larger than 7%. This convinces me that
income is not an important
confounder of the lung cancer versus radon relationship. It is not a
mathematical proof, so my mind is open.
If someone can devise an acceptable model in which income does have an
impact, and this is not taken care
of by our CF, I will be happy, and even relieved, to concede.
But what about Lubin's mathematical proof? It's easy to demonstrate its
validity. Our results would be
explained if those with an income that is an integral multiple of $700
have 1000 times higher lung cancer rates
than all others, in which case lung cancer rates would depend on what
fraction of the population has those
special incomes. We have no data to show that such incomes are not
strongly correlated with radon levels,
and as Lubin showed, the error in our study could be essentially infinite.
But such a model is not acceptable
for two reasons:
(a) it is not plausible;
(b)it would also not be taken care of in case-control studies as they
don't match incomes with that accuracy.
This introduces two obvious corollaries that must be attached to Lubin's
demonstration, but were not
included in his mathematical treatment. Without them, his treatment is not
applicable to the `real world'. What
is needed is a model that avoids these two limitations, and neither I nor
anyone else has been able to devise
one.
Of course annual income is not the only CF that must be worried about.
Another example is age distribution.
Case-control studies match cases and controls by age, but in our study
average age in a county does not
represent differences in age distributions. We do use age-adjusted
mortality rates which take care of the gross
aspects of that problem, but there are limitations in the age-adjustment
process. Our solution is to use as CF
the percentage of the population in each age bracket, <1 year, 1-2 years,
..., 80-84 years, >85 years (31 age
brackets in all), and to also use combinations of adjacent age brackets.
Only one of these has a correlation
with radon above 4%, and after further investigation of that case
(correlation 7.7%), we concluded that
confounding by age distribution could not explain our discrepancy.
There are few, if any, other bases on which case-control studies match
cases and controls, but in our study
we gave similar treatments to a host of other potential confounding
factors - educational attainment, urban
versus rural differences, ethnicity, occupation, housing characteristics,
medical care, family structures, etc.
We have found nothing that can explain our discrepancy.
Let me comment more generally on Lubin's mathematical proof. I have had
extensive experience as a
theoretical physicist, a field heavily characterised by mathematical
proofs. I know the game very well and can
testify that physicists do not derive something mathematically and say
`that's the way it is'. At every stage
they test results numerically with `real world applications'. In fact,
more often they work with real world data
to reach a conclusion, and then `dress it up' for publication by
developing a mathematical treatment. If a
mathematical treatment shows something questionable, it is always easy to
develop plausible practical
numerical examples to settle the question. For example, I can easily show
how any other published ecological
study can give invalid results, and why our study is different from these.
I can't prove that there isn't an unrecognised CF that can explain the
great discrepancy between our data and
LNT (of course, an unrecognised confounder can also nullify the results of
any epidemiological study). We
have certainly exerted a great effort in looking for one, without success.
Our paper has been out for five
years, and no one else has found one. Unless someone does, I will believe
that there is no such CF, and that
our discrepancy with LNT is explained by the fact that LNT fails badly in
the low dose region where there
are no other data to test it.
The other issue raised by Darby and Doll, involving parallel individual
and ecological analyses, has been
addressed in a recent paper (Cohen 2000).
Yours faithfully,
Bernard L Cohen
Department Of Environmental and Occupational Health
100 Allen Hall
University of Pittsburgh
Pittsburgh, PA 15260
USA
URL: stacks.iop.org/0952-4746/21/64
DOI: 10.1088/0952-4746/21/1/102
************************************************************************
You are currently subscribed to the Radsafe mailing list. To unsubscribe,
send an e-mail to Majordomo@list.vanderbilt.edu Put the text "unsubscribe
radsafe" (no quote marks) in the body of the e-mail, with no subject line. You can view the Radsafe archives at http://www.vanderbilt.edu/radsafe/