How to eradicate fraudulent statistical methods: statisticians must do science.

The two steps necessary for the clinical expression of a mutagenic disease, genetic damage and viability, are countervailing forces and therefore the dosage response curve for mutagens must have a maximum. To illustrate that science is common sense reduced to calculation, a new mathematical derivation of this result and supporting data are given. This example also shows that the term "context-free" is a snare and a delusion. When statistical methods are used in a scientific context where their assumptions are known to fail and where there is a reasonable presumption of intent to deceive, they are fraudulent. Estimation of low-level mutagenic risks by linear extrapolation from high-dose data is one example of such a method that is widely used by Executive Branch agencies. Other examples are given of fraudulent statistical methods that are currently used in biomedical research done by or for U.S. government agencies. In the long run, it is argued, the surest way to eradicate such fraud is for biostatisticians to do their own science.

Do Atomic Veterans have excess cancer? New results correcting for the healthy soldier bias.

Reanalysis of the National Research Council report on Mortality of Nuclear Weapons Test Participants, released June 4, 1985, shows 62% excess cancer among soldiers involved in nuclear weapons testing in 1957 codenamed PLUMBBOB, who had exposures to fallout of 300 mrem or more. Although the "healthy soldier bias" was discussed in the original report and a method of correcting for it was described, false negative results were reported because no correction was actually made. Correcting for the healthy soldier bias reveals excess digestive, respiratory, leukemia, and other cancers in PLUMBBOB participants whose reported doses were over 300 mrem.

Why proof of safety is much more difficult than proof of hazard.

On the basis of simple, generally accepted biostatistical and public health principles, it is shown that for environmental health hazards a proof of safety is much more difficult than a proof of hazard. The effective sample sizes required for proof of safety are orders of magnitude greater than what is feasible in biostatistical-epidemiological studies. Although many assurances of safety "in the name of science" have been issued by government agencies and others, few if any of these assurances are statistically valid.

Automatic encoding of clinical narrative.

An experiment in the automatic encoding of English-language medical data is described. The encoding program has two stages. First, the free-text input is parsed and the information is arranged in a tabular format by a general-purpose natural language processor developed at New York University. Then a simple code-dependent subprogram assigns numerical values to the entries on the basis of the positions the input words occupy in the information format. Results of a blind test of the encoding program using the code employed at Roswell Park Memorial Institute for earliest symptoms of head-neck cancer are presented.

Direct estimates of low-level radiation risks of lung cancer at two NRC-compliant nuclear installations: why are the new risk estimates 20 to 200 times the old official estimates?

An official report on the health hazards to nuclear submarine workers at the Portsmouth Naval Shipyard (PNS), who were exposed to low-level ionizing radiation, was based on a casual inspection of the data and not on statistical analyses of the dosage-response relationships. When these analyses are done, serious hazards from lung cancer and other causes of death are shown. As a result of the recent studies on nuclear workers, the new risk estimates have been found to be much higher than the official estimates currently used in setting NRC permissible levels. The official BEIR estimates are about one lung cancer death per year per million persons per rem[s]. The PNS data show 189 lung cancer deaths per year per million persons per rem.