Fw: Iowa Controls - Not Matched

["field" <bill-field@uiowa.edu>]

Dr. Long.



> >

> >   Howard Long MD MPH, Family Doctor and Epidemiologist

> >   363 St. Mary St., Pleasanton CA, 94566

> >   (925) 846-4411, Fax 4524, Page 787-0253 hflong@pacbell.net

> >

> > Dear Professor Field,

> >

> > I have great respect for your scientific honor and the quality of the

> > Iowa study.

> > I am happy that you write, "I AM NOT A DEFENDER OF THE LNT THEORY".

> >

> > When opposite inferences are reasonable from similar studies,

> > like your

> > Iowa study and the N Shipyard Worker Study (also case control and 10 x

> > as large as yours), I believe an experiment is needed.

> >

> > First, even if a drug company were to fund 100 clinical trials,

> > (prospective and double blind) with N in each such that some trials

> > would likely show p<.05 of chance results,but publish only  studies

> > showing benefit, would it show the medicine effective?

> > You studied the one location in 100 (Iowa women) having no negative

> > correlation of radon and lung cancer mortality, in Cohen's

> > study - 200 x

> > as large, albeit ecologic.

>

> I trust that he is not suggesting that we chose to do our study in Iowa

> because this was one of the states in the 50 or so in Cohen's study that

had

> a positive association.  The study was funded and done in Iowa because

Iowa

> is the ideal state to do such a study.  Perhaps, this later reason is why

a

> positive association has been observed in Iowa.

>

> The main argument against Cohen's studies is and always has been this:

> aggregation in an ecologic study can lead to biased estimates.  The bias

can

> be so bad as to yield estimates with the wrong sign, as Jay Lubin has

shown.

> Increasing the sample size does not reduce the bias in any way; it simply

> leads to more precise estimates.  In other words, you would be more

> precisely estimating the wrong thing.

>

> I don't understand the analogy.  It only makes sense if we purposely

picked

> Iowa because it happened to have the change association, and we reported

> Cohen's findings for Iowa.  Neither case is true.  We conducted a superior

> study independent of Cohen's design or results.  Whether our findings

agree

> with the hypotheses generated in Cohen's ecologic study is irrelevant to

me.

> I am intimately aware of how the data was collected and analyzed in both

> studies.  The results from Cohen's ecologic study are dubious because of

the

> increased risk of biases inherent in his study design.  Larger sample

sizes

> can not overcome bias (a clear case of quality versus quantity).

>

> >   .

> > Second, in Topics Under Debate, Radiation Protection

> > Dosimetry V95,1,p77

> > you write in Rebuttal, "The participants' smoking histories

> > do not need

> > to match the smoking histories of the controls since the effect of

> > smoking can be adjusted for using standard statistical methods." This

> > follows Klaus Becker's Argument that "- in the Iowa Lung

> > Cancer Study by

> > Field et al 86% of the ling cancer cases were smokers, but only 32% of

> > the controls." Ibid, p79. Our Professor of statistics at UCB PH, Bill

> > Gaffney, would often remind us,

> > "Know your assumptions!" In your study, the controls are not

> > matched. I

> > do not believe that here, "smoking can be adjusted for using standard

> > statistical methods".

>

> Multiple logistic regression was used to model the effect of residential

> radon exposure on lung cancer risk.  Included in the regression model were

> variables to adjust for the effects of smoking.  Specifically, the model

> included continuous variables for the length of time that individuals

> smoked, the number of cigarettes smoked during that time, and time since

> smoking cessation (for ex-smokers).

>

> Quoting Hosmer and Lemeshow's authoritative book, "Applied Logistic

> Regression":

>

> "One generally considers a multivariate analysis for a more comprehensive

> modeling of the data.  One goal of such an analysis is to statistically

> adjust the estimated effects of each variable in the model for differences

> in the DISTRIBUTION of and associations among the other independent

> variables.  Applying this concept to a multivariate logistic regression

> model, we may surmise that each estimated coefficient provides an estimate

> of the log odds [of lung cancer] adjusting for all other variables

[smoking]

> included in the model."

>

> I would further suggest that there is no practical way to match cases and

> controls so as to adjust for the effects of smoking on lung cancer risk.

> Smoking is so strongly and intricately associated with lung cancer that

one

> would have to match on a multitude of factors, including intensity,

> duration, and time since cessation.  Furthermore, there would undoubtedly

be

> other covariates such as age, education, second hand smoke, family

history,

> and occupational exposures that would not be matched and would have to be

> controlled for in some fashion; i.e. multiple logistic regression.

>

> >

> > I, unlike some cynical colleagues, believe that your intent,

> > like mine,

> > is to prevent lung cancer and advance science. I have in the past

> > rationalized projects (like a method of colon cancer detection) more

> > than I believe you have rationalized this power of statistics

> > to correct

> > for the difficulty in finding controls that match for smoking. I hope

> > you will come to believe, as I do, that less smoking by controls may

> > have been the reason they had less cancer, rather than the less radon.

> >

> > With continued respect and best wishes

> >

> > Howard Long

> >