Re: Linear Model

[dj_strom@ccmail.pnl.gov]

I was going to stay out of this, but some of these comments require a response. 
>Subject: Linear Model
>Author:  HWADE@aol.com at -SMTPlink 
>Date:    8/8/95 0:41

>1. Cohen's 1995 paper answers the previous criticisms that the "ecological 
>fallacy" marred its conclusions. HP,v68,no.2.
>Therefor, previous criticisms are no longer valid

There are several issues regarding Cohen's work in addition to the ecologic 
study: self-reported data not field validated, short-term measurements, 
measurements taken long after causal exposures may have occurred, measurements 
representing only a portion of an individual's exposure, no controls for 
*individuals* on smoking, migration, etc.  

The ecologic fallacy is "An error in interpreting associations between ecologic 
indices.  It is committed by mistakenly assuming that, because the majority of a
group has a characteristic, the characteristic is related to a health state 
common in the group" (Slome C, Brogan DR, Eyres SJ, Lednar W.  Basic 
Epidemiological Methods and Biostatistics - A Workbook. Boston: Jones and 
Bartless, 1986, Chapter 9 & p. 306).  The problem with the ecological study 
design is that it doesn't have individual doses linked to individual people.  
No amount of "correcting" can get around this.

Studies with the ecological or geographical design are *hopelessly* flawed from 
the basis of deriving causal inferences.  Counties don't get lung cancer, people
do.  Furthermore, you can't "correct" for migration.  If you want to understand 
lung cancer in Miami, then measure radon levels in New York and Havana 40 years 
ago, where the people in Miami who are dying of lung cancer today would have had
their causal exposures.  Cohen's *self-reported* measurements were made in the 
1980's and 1990's, not in the 1950s.  One properly-done study linking doses 
(before diagnosis, with an appropriate lag-time to account for latency) to 
individuals with known smoking histories is more cogent than a one hundred 
ecological studies.  See two articles debunking of ecological designs in the HPS
Newsletter:  Conrath S.  Study Design as a Determinant of Radon Epidemiologic 
Study Validity.  HPS Newsletter 18(7):1-5, July 1990.  Strom DJ. The Ecologic 
Fallacy. HPS Newsletter 19(3):13, Mar 1991.  

As I wrote to Dr. Cohen, January 3, 1990:  "The enclosed article by Alvan R. 
Feinstein entitled 'Scientific Standards in Epidemiologic Studies of the Menace 
of Daily Life' (Science 242:1256-1263, 2 DEC 88) may be of interest to you.  The
way I read it, ecologic studies fail to meet four of Feinstein's five criteria 
(ii-v)  Perhaps his work gives some insight on the unwillingness of mainstream 
epidemiologists to accept ecologic studies."  The uses of the ecologic design, 
if any, are for hypothesis generation or temporal trending, but not for 
quantitative tests of dose-response models.

>2. Cohen has now completed a case-control study that shows a negative 
>correlation between radon concentration and case-control ratio. Annual 
>Meeting abstracts, HP,supp to v68, no.6.

As Dr. Otto Raabe pointed out to Dr. Cohen in Boston, and despite the 
fact that it has been accepted for publication in *Health Physics*, this 
was not a case-control study.  In a case-control study, individual cases 
are matched on age, gender, smoking, and other variables with individual 
controls.  Dr. Cohen simply took a group of *self reported* diagnoses 
(with no attempt to control for whether the lung cancer was a primary 
tumor or a metastatic tumor, since no medical records were examined) of 
cancers other than lung and compared their present radon levels with 
those of *self reported* diagnoses of lung cancer.  This is not a 
case-control study - ask any formally-trained bona fide epidemiologist.  
Furthermore, with no histological data, no site of tumors, no medical 
records, and all self-reported diagnoses, this work would have trouble 
with peer review for an epidemiology journal.  

>3. The Japanese data are subject to the same "ecological fallacy" as other 
>studies, yet we seem to be able to accept them, and indeed use them to 
>build a case for linearity.

No, the life span study (LSS) does not have the ecological fallacy.  It is 
not a great stretch to assume that those "not in city at time of bomb" were 
not exposed; for those in the LSS cohort, each *individual* has had a dose 
assessment performed (some 75,000 are in LSS).  Thus, individual doses are 
associated with individual health outcomes (causes of death), making the 
study far, far more cogent.

>4. The paper by the IARC study group on cancer risk among nuclear industry 
>workers, Lancet, 344:1039;1994 fits the data to a linear model and therefor 
>cannot be used as an argument for the validity of the linear model. (not 
>unless you believe one can pull himself up by tugging on his bootstraps.)

If you're interested in arguments *for* the linear model, see the many 
publications of the ICRP, National Academy of Sciences, UNSCEAR (in 
particular UNSCEAR 1994, as reported at HPS Boston by Warren Sinclair and 
Burton Bennett).  Also, see the Radiological Protection Bulletin of the UK 
National Radiological Protection Board, July 1995, pp. 8-12, for a very 
cogent review supporting the linear, non-threshold dose response model.  
Also, see my upcoming letter in the HPS Newsletter.

Frankly, I don't get my risk analysis from health physicists, I get it from 
scientists and scientific organizations doing the work first-hand, as published 
in peer-reviewed literature (such as Radiation Research or American Journal of 
Epidemiology) or reports of NAS, ICRP, UNSCEAR, NRPB, etc.; and then, everything
must be examined critically and taken with a grain of salt.  

Daniel J. Strom, Ph.D., C.H.P.
Health Protection Department K3-56
Pacific Northwest Laboratory
P.O. Box 999
Richland, WA 99352-0999 USA
(509) 375-2626
(509) 375-2019 fax
dj_strom@pnl.gov