[Analysis of correlated data: problems and examples in industrial medicine]

In occupational health we are frequently faced with data that are not independent, but to recognize the lack of such an independence is not yet a common practice. To help researchers in the field to treat correlated data in a proper way this paper has two aims: to highlight practical situations in which the data are not independent and to show the main differences between a statistical analysis which does consider the correlation appropriately and one which doesn't. As to the first aim, four typical examples are discussed: repeated measurements in the same subject (e.g., cells in which the number of sister chromatid exchanges is counted); health effects observed in multiple organs (e.g., visual impairment in both eyes); evaluation of prevention programs (e.g., exposure assessment to styrene before and after environment remediation); longitudinal studies of health effects (e.g., changes over time of pulmonary function parameters). With respect to the second aim a practical exercise is described completely. Measurements of exposure to a toxic substance in two different departments before and after environment remediation are evaluated with statistical tools which both do and do not consider the correlation between such measurements. Differences in the results obtained, with particular reference to indices of variability (e.g., standard errors), point toward the need of analysing correlated data with appropriate statistical tools that take correlation into account.