David S. Salsburg

The Religion of Statistics as Practiced in Medical Journals

Abstract Commentaries are informative essays dealing with viewpoints of statistical practice, statistical education, and other topics considered to be of general interest to the broad readership of The American Statistician. Commentaries are similar in spirit to Letters to the Editor, but they involve longer discussions of background, issues, and perspectives. All commentaries will be refereed for their merit and compatibility with these criteria. The way in which hypothesis tests are often used as the sole tool of statistics in medical research is satirized. This is followed by a suggestion for reform.

Point mutations at the thymidine kinase locus in L5178Y mouse lymphoma cells II. Test validation and interpretation

The L5178Y Mouse Lymphoma TK assay was studied extensively to determine if this mammalian cell assay for gene mutations at the thymidine kinase (TK) locus could provide valid, interpretable determinations of mutagenic potential, and whether this information is of value in the safety evaluation of chemicals. We first determined that test-derived TFTR mutants were phenotypically stable, possessing little or no thymidine kinase activity as measured by labeled thymidine uptake, but demonstrating 100% cross resistance to bromodeoxyuridine. Common solvent vehicles such as acetone, dimethylsulfoxide and ethanol were shown to produce little cytotoxicity and no mutagenic activity when present at 1% levels. Out of a total of 10 noncarcinogens tested, all were negative when results were analyzed by a 2-sample loge t test on control and treated mutant count means. Of the 13 putative animal carcinogens tested, 10 were positive, 2 were negative (auramine O and sodium phenobarbital), and 1 showed sporadic activity (hydrazine sulfate) in the TK assay on the basis of test-derived t statistics. 2 compounds, 1,2-epoxybutane and ICR 191, which have been described as Ames positive non-carcinogens, were also positive in the TK assay. Although this sampling of a total of 29 compounds is insufficient for precise estimations of expected false-positive or false-negative frequencies, these data indicate the TK assay can be expected to detect a majority of carcinogens as mutagens including some missed by more established point-mutation assays.

The lifetime feeding study in mice and rats--an examination of its validity as a bioassay for human carcinogens.

Currently used designs for carcinogenic bioassay (lifetime feeding studies in mice or rats using maximum tolerated doses of the test compound) are examined to see if they meet the requirements of a bioassay, using the results of 170 compounds reported on as of June, 1980, by the National Cancer Institute. It is concluded that the lifetime feeding study has never been subjected to proper validation as an assay for human carcinogens. When an attempt is made to validate it on the basis of these reported studies and those in the literature, it appears to lack acceptable specificity and sensitivity. It is suggested that a drastically different design is needed and that such redesigning of the assay will require proper validation.

A Comparison of Statistical Methods for Low Dose Extrapolation Utilizing Time-To-Tumor Data

The assessment of health risks due to low levels of exposure to potential environmental hazards based on the results of toxicological experiments necessarily involves extrapolation of results obtained at relatively high doses to the low dose region of interest. In this paper, different statistical extrapolation procedures which take into account both time-to-response and the presence of competing risks are compared using a large simulated data base. The study was designed to cover a range of plausible dose response models as well as to assess the effects of competing risks, background response, latency and experimental design on the performance of the different extrapolation procedures. It was found that point estimates of risk in the low dose region may differ from the actual risk by a factor of 1000 or more in certain situations, even when precise information on the time of occurrence of the particular lesion of interest is available. Although linearized upper confidence limits on risk can be highly conservative when the underlying dose response curve is sublinear in the low dose region, they were found not to exceed the actual risk in the low dose region by more than a factor of 10 in those cases where the underlying dose response curve was linear at low doses.

The Effects of Liffytime Feeding Studies on Patterns of Sfxile Lesions in Mice and Rats

ABSTRACTData from lifetime feeding studies in mice and rats comparing controls to three doses of Xylitol, to Sorbitol, and to sucrose are examined to illustrate the thesis that the occurrence of some tumors is often confounded with other lesions and even other tumors in lifetime feeding studies. It is suggested that it may be impossible to test whether a treatment is a “carcinogen” in the presence of such biological confounding.

Comparative pharmacology and clinical efficacy of newer agents in treatment of heart failure

The animal and human pharmacology of several new drugs (prazosin, trimazosin, pirbuterol, and carbazeran) useful in the treatment of congestive heart failure (CHF) is delineated in relation to the pharmacology of other agents employed for CHF management. Prazosin and trimazosin are selective alpha 1-blockers that cause a balanced increase in cardiac output (CO) and reduction in left ventricular filling pressure (LVFP); the reduction in diastolic blood pressure with these drugs is significantly related to increase in treadmill exercise, fall in LVFP, and increase in CO. Pirbuterol is a relatively selective beta 2-agonist with somewhat greater effects on CO than on LVFP. Early promise in CHF therapy is being shown by a novel series of cyclic adenosine monophosphate (cAMP) phosphodiesterase inhibitors with combined direct inotropic and vasodilator effects. Double-blind long-term studies demonstrate persistent efficacy of prazosin and trimazosin in CHF as measured by improvement in New York Heart Association functional class, treadmill exercise performance, and noninvasive measures of cardiac function; these data are supported by studies in which repeat cardiac catheterization has been performed after several months of therapy. Double-blind studies of other CHF drugs are in progress.

Approximate Tests and Confidence Intervals Using the Jackknife

Publisher Summary This chapter presents a discussion of relevant papers that have been significant in bringing the jackknife technique to its present status. This technique is almost universally applicable in aiding a researcher to obtain approximate tests or confidence intervals for parameters of interest. To understand its potential pitfalls, however, it is important to understand the development of some of the theory behind it. This development is far from complete as of this writing. There are two problems that are very important from a practical viewpoint that have received virtually no attention. These are the problems of deciding if ones data is adequate for the asymptotic result and what, if any, transformation should be used in conjunction with the jackknife. Only limited results give an indication of an answer to the former question. The answer to the second question is variance stabilizing transformation until a better answer comes along.

Estimating Differences in Variance when Comparing Two Methods of Assay

Morgans test for the relative precision of two assays compared in the same set, of samples is extended to the case where the underlying distribution may not be hivariate normal and where the units of the two assays may differ. The jackknife is also suggested for computing confidence intervals on the differeuces in variance or coefficients of variation.

Development of statistical analysis for single dose bronchodilators

When measurements developed for the diagnosis of patients are used to detect treatment effects in clinical trials with chronic disease, problems in definition of response and in the statistical distributions of those measurements within patients have to be resolved before the results of clinical studies can be analyzed. An example of this process is shown in the development of the analysis of single-dose bronchodilator trials.

Mathematical Modeling and the Design of Studies in Drug Research

In the most general sense, a drug can be thought of as having two dose response curves: P1(d) = Proportion of patients who “respond” at dose ≤ d. P2(d) = Proportion of patients who suffer adverse reactions at dose ≤ d. The ideal drug is one with the two curves widely separated, where the first curve is steep (so a small range of doses are adequate to most patients) and the second curve is shallow. Both pre-clinical animal studies and human clinical studies are used to estimate the form and placement of those curves. However, the curves can never be fully characterized, and there are important definitional problems that have to be resolved, such as What is a “response”? Who are the patients? What adverse reactions? Time dependent patterns of response? Etc. As a drug is developed, segments of this overall mathematical model are formulated and studies are used to estimate parameters or test different submodels. This paper will try to present a coherent picture of how the designs of traditional studies fit this process and explore alternative designs that may be more useful.