Re: Wing: Descriptive Epidemiology by Any Other Name...

["John Moulder" <jmoulder@post.its.mcw.edu>]

> From: "Sandy Perle" <sandyfl@ix.netcom.com> 
>  
> John Moulder stated: 
>  
> > January copy of the EHP has not yet reached subscribers.  It has 
> > only been in the last 24-48 hours that a copy has appeared on the 
> > NIEHS website. 
>  
> The news media printed the details of the allegations from Wing 
> and colleagues... Again, WHY haven't there been 
> any comments from HPS, or other organizations? 

What was in the media was nothing more (and often less) than what was in the 
press release.  As a scientist, I cannot critique a study when I haven't seen 
it.

To critque a study on the basis of a press release is as irresponsible as 
publishing by press release in the first place.

I had actually meant to append the following to the previous message, but it 
appears that I did not.

---From the NIEHS website-------
http://ehpnet1.niehs.nih.gov/docs/1997/105(1)/wing.html

Steve Wing, David Richardson, Donna Armstrong, and Douglas Crawford- Brown: A 
Reevaluation of Cancer Incidence Near the Three Mile Island Nuclear Plant: The 
Collision of Evidence and Assumptions.  Environ Health Perspect 105:52-57 
(1997)

Author Abstract
Previous studies concluded that there was no evidence that the 1979 nuclear 
accident at Three Mile Island (TMI) affected cancer incidence in the 
surrounding area; however, there were logical and methodological problems in 
earlier reports that led us to reconsider data previously collected. A 10-mile 
area around TMI was divided into 69 study tracts, which were assigned 
radiation dose estimates based on radiation readings and models of atmospheric 
dispersion. Incident cancers from 1975 to 1985 were ascertained from hospital 
records and assigned to study tracts. Associations between accident doses and 
incidence rates of leukemia, lung cancer, and all cancer were assessed using 
relative dose estimates calculated by the earlier investigators. Adjustments 
were made for age, sex, socioeconomic characteristics, and preaccident 
variation in incidence. Considering a 2-year latency, the estimated percent 
increase per dose unit +/- standard error was 0.020 +/- 0.012 for all cancer, 
0.082 +/- 0.032 for lung cancer, and 0.116 +/- 0.067 for leukemia. Adjustment 
for socioeconomic variables increased the estimates to 0.034 +/- 0.013, 0.103 
+/- 0.035, and 0.139 +/- 0.073 for all cancer, lung cancer, and leukemia, 
respectively. Associations were generally larger considering a 5-year latency, 
but were based on smaller numbers of cases. Results support the hypothesis 
that radiation doses are related to increased cancer incidence around TMI. The 
analysis avoids medical detection bias, but suffers from inaccurate dose 
classification; therefore, results may underestimate the magnitude of the 
association between radiation and cancer incidence. These associations would 
not be expected, based on previous estimates of near-background levels of 
radiation exposure following the accident. Key words: dose-response 
relationships, ecologic studies, environmental epidemiology, ionizing 
radiation, methodology, neoplasms, nuclear power.  

---------

Some comments based on my first reading of the article.

1)  This is standard ecological-design risk-detection epidemiology.  It looks 
for associations without any real consideration of plausibility or any serious 
consideration of what else is known about the agent or the disease.  The fact 
that it is "standard" does not make it good science; but stuff like this 
appears in the epidemiological literature all the time (e.g., "hot dogs cause 
childhood leukemia", which appeared in 1994 in a major epidemiological 
journal)

2)  This is re-analysis of the Hatch et al data (Amer J Epidem 132:397-412).  
No new information appears to have been added.

3)  The authors use the dose estimates previously made by Hatch et al (Amer J 
Epidem 132:397-412), but assume that while the Hatch et al data is relatively 
correct, it is absolutely wrong.  That is, they assume that the pattern is 
correct but that the absolute dose estimates are wrong.  The authors' grounds 
for disputing the absolute dose estimates are largely anecdotal (e.g., 
testimony to the NRC about acute radiation sickness).

4)  The authors use "relative dose estimates".  How these units correspond to 
"dose" is not stated.

5) The results assume a two- or five-year latency.  

6)  Only cancers appearing between 1981 and 1985 are considered.  The 
explanation appears to be that this is the only data available (i.e., it's 
okay to use inappropriate data if that's all that is handy).  Since the latent 
period for radiation induction of lung cancer is generally believed to be 20+ 
years, this time period makes no sense.

7)  Errors are in "standard errors", multiply these by 1.96 to get 95% 
confidence intervals and to assess statistical significance.

8)  Cases are assigned "doses" based on their place of residence at the time 
of diagnosis.  It is not known where these people lived at the time of their 
"exposure".

9)  If anyone moved post-accident, or was treated in an out-of-area hospital, 
the case would be lost.

9)  Cancer rates are adjusted for age, sex and socioeconomic status, but not 
for smoking history.  

10)  Two different "models" are used to calculate "expected" cancer rates; one 
uses socioeconomic status as a variable, the other does not.

11)  The authors make the point that their relative risks are highest for 
leukemia, without mentioning that these relative risks are also not 
significantly different from one.

12)  While exposure-response data is presented, no trend analysis appears to 
have been done, and the data presentation is such that it would be difficult 
for someone else to do.  Any attempt to seriously look at the trend analysis 
would also have to deal with the fact that less cancer was observed than was 
expected at the lower "doses".

Quick Hill criteria analysis:

- Strength of association:  Weak, relative risks are generally below 2 for the 
groups with the highest exposures.

- Consistency:  None, results are inconsistent with all other epidemiological 
studies of radiation induced cancer.

- Specificity:  Little relevance, as we know that radiation can induce a 
number of different types of cancer.

- Temporality:  Okay, the reported increase was after exposure.

- Biological gradient:  Unknown, no analysis of exposure-response done.

- Plausibility:  None.  There is no other epidemiological or experimental 
evidence to support relative risks this high at this dose.  With the possible 
exception of leukemia, there is also no data to support latency periods this 
short.

- Coherence:  None.  Interpretation of the results is incompatible with 
essentially everything else that is known about radiation-induced cancer.

- Experiment:  None.  The experimental data on radiation-induced leukemia and 
lung cancer does not support effects at this dose, or in the cancer of lung 
cancer, a latency period this short.

- Analogy.  None that I can think of.  That is, I know of other cases where a 
low level environmental exposure has produced an effect that is incompatible 
with the results of higher dose exposures and incompatible with substantial 
experimental evidence.

- Overall:  The observed associations, but themselves, add little or nothing 
to our knowledge of the effects of low level radiation exposure.





John Moulder (jmoulder@its.mcw.edu)