In 1993 Landtblom et al (1) reported a case-referent study on multiple sclerosis and exposure to solvents, ionizing radiation, and animals. Data were obtained from 91 cases and 348 referents. Multiple logistic regression was used in the analysis of the data, and 14 predictors were included in the final model. The standard errors appear to be large, as reflected in the wide confidence intervals (CI). For example, the estimated odds ratio for occupational exposure to cats or dogs among men was 18 (95% CI 1.3--265). Despite the wide confidence interval the authors concluded that "The men had significantly elevated risks, determined from logistic odds ratios, for . . . occupational contact with dog or cats, . . . [p 399]." The odds ratio for X-ray treatment was reported to be for the women, 0 for the men, and 0 when both men and women were included in the model. Obviously there is a problem with the model, but the authors interpret the results in the following way "This study indicates that exposure to ionizing radiation might have an increased risk for multiple sclerosis, as observed both for patients treated with X rays and for radiological personnel [p 402]." In my opinion this example illustrates the danger of an uncritical use of the logistic regression analysis. The computer nicely calculates regression coefficients even if the number of empty cells is large and provides an error message only when the maximum number of iterations is reached. The presence of zero cells should be recognized before the multiple logistic regression is carried out. By collapsing categories and excluding predictors with low prevalence the problem with empty cells can be diminished.