[Analysis of correlated data in occupational medicine: examples with binary data]

In a previous paper in this Journal we presented and discussed examples of analysis of correlated data when the response variable was continuous and normally distributed (the measurement of exposure to a toxic substance was the case in point). In this paper we extend the analysis and the discussion to take into account categorical (binary) variables (described in terms of proportions or odds); to favour the comprehension of the analogies (and discrepancies) between the two contexts we have fully developed an example that mimics the situation presented in the previous paper. Marginal, conditional, random effects and transitional models for correlated data are introduced in practical terms; the meaning of the different estimates obtained are interpreted for epidemiological purposes; the disadvantages of not considering correlation in the analysis are explained and the complexities connected to this type of analysis are fully appreciated. It is concluded that correlated data are very frequently encountered in occupational settings and that an appropriate analysis is necessary. This analysis requires sophisticated computer programs and statistical expertise, particularly in the case of categorical data.