Egeland GM, Sweeney MH, Fingerhut MA, et al. Total serum testosterone and gonadotropins in workers exposed to dioxin. Am J Epidemiol 1994;139:272-81. 2.
Letters to the Editor
RE: "TOTAL SERUM TESTOSTERONE AND GONADOTROPINS IN WORKERS EXPOSED TO D10X1N" Egeland et al. (1) reported significantly low testosterone and high gonadotropin levels in men exposed to dioxin. Noting that this hormone profile is hypothesized to cause low offspring sex ratios (proportions male), I predicted that if dioxin causes the hormone profile associated with it, then exposed workers should produce an excess of daughters after exposure (2). Mocarelli et al. (3) recently reported a highly significant excess of daughters born to parents following such exposure in the Seveso disaster in 1976.1 infer that dioxin was indeed responsible for the hormone profile of men exposed to it as described by Egeland et al. (1). REFERENCES
1. Egeland GM, Sweeney MH, Fingerhut MA, et al. Total serum testosterone and gonadotropins in workers exposed to dioxin. Am J Epidemiol 1994;139:272-81. 2. James WH. Re: 'Total serum testosterone and gonadotropins in workers exposed to dioxin." (Letter). Am J Epidemiol 1995:141:476-7. 3. Mocarelli P, Brambilla P, Gerthoux PM, et al. Change in sex ratio with exposure to dioxin. (Letter). Lancet 1996;348:409. William H. James Department of Genetics and Biometry University College London Wolfson House 4 Stephenson Way London NW1 2HE, England Editor's note: In accordance with Journal policy, Dr. Egeland and her co-authors were given the opportunity to reply to the above letter, but they chose not to do so.
RE: "SURVEY INFERENCE FOR SUBPOPULATIONS" In a recent paper, Graubard and Kom (1) discussed the problem of making inferences for a subpopulation of interest from the entire sample, in particular a sample obtained from surveys. The authors argued that simply eliminating individuals outside the subset of interest from the survey sample may not be appropriate for the calculation of standard errors and confidence intervals, because of the complex design and unequal weighting frequently used in the surveys (1). The inadequacy of sample-to-population inference has been a concern not only for subpopulation analysis, but for survey data analysis in general (2-5). I wish to point out that while this issue is relevant and important in administrative (or policy) research, it is not relevant in scientific research. In administrative (or policy) research, the purpose is to provide the health profile, or to describe the health care utilization, or to evaluate the programs or practices in a particular community for a particular period of time, so that the planners or regulation agencies can implement adequate procedures to enhance the health status or to increase the effectiveness and efficiency of the services for the community. In this case, appropriate statistical procedures should be used in the analysis, to take the complex design and unequal weighting of the survey into consideration, so that the results can properly reflect the experiences of the target population (or finite population) of the survey. In scientific research, however, the concern is Am J Epidemiol
Vol. 145, No. 6, 1997
569
not with any particular experience per se. Such experiences are only used to leam about the relation of interest in the abstract without any place or time referent (6). For example, the relation between smoking and lung cancer has been studied in various groups of persons in the world. However, the purpose of this research was not to explore whether smoking was the cause of lung cancer in the group of subjects studied (or the "target" or "finite" population they represented), but rather whether smoking causes lung cancer in human beings. In scientific research, therefore, specification of study base is a matter of choice, without consideration of the "probability" or "representativeness" of the "target" or "finite" population. Indeed, in experimental research, which is generally considered the paradigm of observational studies (7), there is no principle in effect that, for example, the mice in the experiment should be a "probability" sample from a very large "target" or "finite" population of mice. Rather, it is well understood that the inference from any experiment with a particular group of mice is to the generic, abstract domain of mice of the same kind (or even human beings), based on scientific judgment rather than technical sample-to-population inference. In this case, actual subjects obtained from the survey should be analyzed with routine statistical techniques. Complicated statistical techniques taking into consideration complex designs and unequal probability sampling of the survey are not only unnecessary but are misleading as well. REFERENCES
1. Graubard BI, Kom EL. Survey inference for subpopulations. Am J Epidemiol 1996; 144:102-6. 2. Kish L, Frankel MR. Inference from complex surveys (with discussion). J R Stat Soc [B] 1974;36:1-37. 3. Rao JNK, Thomas DR. The analysis of cross-classified categorical data from complex surveys. Sociol Methodol 1988; 18: 213-69. 4. Thomas DR. Inference using complex data from surveys and experiments. Can Psychology 1993;34:415-31. 5. Korn EL, Graubard BI. Examples of differing weighted and unweighted estimated from a sample survey. Am Stat 1995; 49:291-5. 6. Miettinen OS. Theoretical epidemiology. New York: Delmar Publishers Inc, 1985. 7. Hill AB. Observation and experiment. N Engl J Med 1953; 248:995-1001. Shi Wu Wen
Bureau of Reproductive and Child Health LCDC Building #6, First Floor Tunney's Pasture, A-L0601E2 Ottawa, Ontario K1A 0L2 Canada THE AUTHORS REPLY Dr. Wen (1) questions the necessity and even the correctness of taking into account the clustering and differentia] weighting of the sample design when analyzing survey data for scientific investigations as opposed to descriptive studies. Although this concern obviously goes beyond the variance-estimation topic of our paper (2), it is useful to discuss this issue in the context of the analyses in our paper rather than in the abstract. Consider the logistic regression analysis of digestive cancer with smoking and drinking. This scientific study was a population-based case-control study in which the cases were
