Dr. Lubin's response to Cohen's statements

[FIELDRW@aol.com]

I feel that I must comment on the exchange by Cohen and Klugh.  Dr. Cohen's
suggestion that because I made no attempt to claim his reward, I must have
accepted the adequacy of his responses to articles by me and others is 
fatuous.
It is not that I don't believe Cohen's results -- below 300 Bq/m3, county lung
cancer rates are negatively correlated with "average" county radon
concentrations --  but that his results have no relevance regarding risk to
individuals.  They neither support nor contradict current risk estimates.  
Thus,
continuing the exchange was pointless.  Upon re-reading his responses to the
exchange in Health Physics and on this listserve, I can only conclude that Dr.
Cohen has entirely missed the point of my comments and those of others.  Let's
try one more time.

1.  It is clear that Cohen's ecological analysis is markedly discrepant from 
the
analytic studies of residential radon exposure and of radon-exposed miners, 
from animal studies and from current dosimetric/biophysical models.  As I 
made clear in my last letter in Health Physics, Cohen's linear-quadric model 
for lung
cancer mortality rates and residential radon is not supported by the
case-control studies or the miner studies.  His model in fact predicts a 
strong
protective effect of living in houses with concentrations under about 300 
Bq/m3
- this pattern just does not occur in the analytic data.   

The relevant question then becomes -- how are we to interpret Cohen's results 
in
light of the findings from the analytic studies?  More importantly, when
considering whether and to what extent residential radon increases (or in 
light
of Cohen's result, decreases) the risk of lung cancer, is it even necessary 
that
Cohen's results should be "explained"?  The main point of my articles, as well
as articles by Greenland and Robins, Samet et al, Piantadosi, Muirhead et al,
and others, is that the answer is an emphatic no to the latter question.  
These
articles show that the methodology of ecological analysis itself has such
limitations that no meaningful inference from the relationship observed at the
grouped level can be made to the dose-response relationship for individuals.
Thus, Cohen's regression results cannot serve as an evaluator of radon effects
for individuals.  This is why epidemiologists use ecological analyses only for
hypothesis generating and not hypothesis testing.  In the real world of
epidemiology, in order to have any credence, an association observed at the
ecological level must first be demonstrated using individual-level studies.  
The
converse is never the case.  One would never consider that results from 
analytic
studies needed to be "validated" by ecological analysis; the very notion is
absurd.  Thus, there is little point in trying to identify any specific reason
why Cohen's regression results in a negative trend, since the analytic 
evidence
is so strong.  The methodology itself is intrinsically flawed.  Given diverse
results from ecological analyses and from analytic studies, analytic studies a
priori have the greater claim to validity.

What is the flaw in ecological studies?  Again, many articles have 
demonstrated
that the principal source of the problem is that correlations among risk 
factors
can occur within counties.  Moreover, the correlations of those risk factors 
may
vary among the counties.  Therefore, one cannot simply define a model for the
within county correlations, apply it across counties and claim adjustment of
these effects.  This is the central rationale for the proposal by Sheppard and
Prentice to "correct" ecological studies by randomly sampling populations 
within
county in order to estimate the joint distribution of risk factors.  Cohen's
repeated assertions that the "ecological fallacy" does not apply to him, does
not make it so.  He cannot create models that account for both the risk factor
correlations within county and their variations across counties.  My original
paper in Health Physics shows that the amount of correlation between radon 
level
and another risk factor can be small indeed, on the order of 0.05.

2.  Cohen suggests that his analysis is not confounded because the negative
trend of lung cancer rates and county "mean" radon level is unchanged after
adjusting for hundreds of different factors.  That is not the point at all, 
and
indeed is irrelevant to the control of within county confounding.  One can 
never
adjust for factors that operate within county by endlessly adding more and 
more
county-level information.  His ecological analysis has no more claim to 
validity
if he were to add thousands of additional county-level variables.  
County-level
factors are not the source of the limitations.

3.  While the within county confounding issue is really paramount, it is also
the case that data used in many ecological analyses are often of limited 
quality
and difficult to interpret.  Cohen's data are no exception.  Lung cancer
mortality from death certificates are often misspecified and covariate
information is of poor quality.  Of particular concern, Cohen uses mortality
rates where the disease-relevant exposure period is at least a decade or more
prior to available radon data.  The county "average" radon concentration may
bear little relationship to the "average" county radon concentration 
prevailing
up to 30 years and more prior to the mortality data.  The population
composition, and risk factors, as well as lung cancer mortality rates, have
changed considerably in recent decades.  (See for example the NCI cancer atlas
for changes in lung cancer mortality, ).  This problem is exacerbated by the
further limitation that average radon level (and other risk factors) for a
county does not translate into lung cancer-relevant dose, and that the
correlations among risk factors may vary over time.

Residential radon studies in general do not suffer from the flaw that radon
measurements post­date the disease occurrence by 10-30 years.  While radon
measurements are contemporary, in residential studies, interviews provide
detailed information on residential history, so investigators know precisely 
how
long subjects were resident in the various houses, and on house modifications,
heating systems, etc.  In the Iowa study, enrollment in the study required 
that
the subject has lived in the current house for at least the preceding 20 
years.

4.  It must also be pointed out that not all ecological studies agree with
Cohen's results, and big does not equate with valid.  Many ecological studies
have been published; some show decreasing trend, no trend, or increasing 
trend.
In an attempt to circumvent the limitation of mortality data and poor exposure
data, several investigators have used population registry files to create lung
cancer incidence rates for counties.  Analyzing these incidence rates together
with county radon measurements and other information from population-based
case-control studies, investigators have found patterns of associations that
show increasing risks with radon.  None show decreasing risks.

This exchange has been going on for nearly a decade.  To paraphrase 
Piantadosi,
given results from analytic studies and ecological studies, the former must
carry the greater weight in our considerations.  Thus, the suggestion that
Cohen's results must be "explained" is misplaced -- the results are 
irrelevant.
It is time to move on and address the many important questions that remain: 
what
are the consequences of exposure measurement error on risk estimates; what is 
a
realistic level of uncertainty on the risk associated with long-term residence
in houses in the range of 75-200 Bq/m3; are there specific genetic 
polymorphisms
that define susceptible sub-populations?

Jay Lubin 

Jay Lubin, PhD
National Cancer Institute
Biostatistics Branch, EPS/8042
6120 Executive Blvd
Bethesda, MD 20892-7244
Tel: 301-496-3357
Fax: 301-402-0081
Email: lubinj@exchange.nih.gov
************************************************************************
The RADSAFE Frequently Asked Questions list, archives and subscription
information can be accessed at http://www.ehs.uiuc.edu/~rad/radsafe.html