In this issue, Karipidis and colleagues report the interesting finding of a possible association between non-Hodgkin lymphoma and occupational exposure to power-frequency magnetic fields.¹ Their results are important given that existing studies on this topic are few, with conflicting results.^2,3 As in any other community-based case–control study their subjects have been exposed to a variety of chemical and physical agents. The authors, like many others, have published associations between various occupational exposures and non-Hodgkin lymphoma in separate papers.^4–6 Every researcher ever involved in large and expensive community-based case–control studies focusing on an adverse outcome and occupational exposures will be aware of the factors increasing the likelihood of publication of the results of positive and negative exposure–response relations in multiple papers. These include a general trend towards shorter papers in journals, but an expectation that assessment methods are specifically tailored to the exposure in question and that other aspects of methodology are described in sufficient detail.

This example, however, highlights an increasingly important issue in occupational epidemiology—how to assess a series of related publications. With the use of job exposure matrices (JEMs) or experts to assign occupational exposure to cases and controls, it is possible to relate many occupational exposures to the studied health effect in the absence of clear a priori hypotheses. The risk of finding several positive associations simply by chance may thus be substantial. Where such findings are spread across several publications, there may not be full recognition of the problem of multiple comparisons or the often complex relationship between different occupational exposures included in different papers.

Moreover, later analyses may raise questions about confounding by factors that were less completely considered in papers of earlier and different focus.

The paper by Karipidis et al follows two previous publications concerning the association of lymphoma with chemical exposures like solvents, metals, organic dust and PCBs and exposure to pesticides.^5,6 A third paper was published on the protective effects of sun exposure in the same population.⁴ In the earlier papers, not all of the estimated exposures included in later analyses could be described and so adjusted for. However, when we look at the exposures considered: sunlight, pesticides, solvents, metals, organic dust and PCBs, one can safely assume that exposures to these agents did not occur independently. For instance, utility personnel clearing the undergrowth of power lines will be exposed to power-frequency magnetic fields, pesticides (herbicides) and sunlight at the same time. Also, a subset of metal workers may have been exposed to solvents and metals simultaneously. From the four papers one can derive a picture of overlapping (occupational) exposures in this population of cases and controls (see table 1).

Analyses of occupational exposures often start with looking for increased risks in certain jobs. Most case–control studies, however, are too small to come up with conclusive results when analysing at job level. The consequent use of JEMs and case-by-case expert assessment enables more powerful analyses since exposures to single (groups of) agents are combined across industries and jobs. At the same time the intrinsic and complex underlying exposure profiles are often not fully considered. With prevalence of exposure as high as those shown above and a considerable potential of strong correlations between agents in subsets of the study population, a more elaborate statistical analysis is warranted (see for instance De Roos et al⁷). Simple statistical adjustment for other exposures might often not be sufficient, given different type 1 error structures for the different exposures and possible complex correlation structures.

An analogy can be made with the field of genetics where enormous amounts of data on genetic variants are being assembled that have to be correlated with outcome parameters. The huge number of comparisons (up to thousands or even hundreds of thousands) in the genetic studies gives rise to a substantial chance that an initial “statistically significant” finding will turn out to be a false-positive finding. Therefore, several techniques have been developed to account for multiple comparisons, including false discovery rate procedures,⁸ or more Bayesian approaches like false positive report probability (FPRP), in which the prior probability of a true association of the gene (exposure) with a disease is specified a priori.⁹ Although the number of comparisons in occupational studies is very much lower than in genetics, it is our opinion that such approaches should start to find their way into case–control studies that focus on multiple occupational exposures. At a minimum, authors should make clear how many exposures were considered at that time, justifying their inclusion in the study hypotheses. Furthermore, presenting initial observations together in one paper, and applying state-of-the-art statistical analyses accounting for multiple exposures (e.g. hierarchical regression) would be extremely helpful for the readership and would better reflect the complexity in which occupational exposures occur.