The results show a possible association between exposure to air pollution from the paper, pulp and board industry and excess risk of NHL mortality, regardless of which model is used. Analysing the information contained in the EPER for 2001 shows that almost all the paper, pulp and boardindustries reported emissions of the following compounds, above the threshold established for their inclusion in the registry: CO; CO2; NO2; sulphur dioxide; organochlorinated compound mixtures; and organic carbon. Taken individually, some of these industries also reported emissions of metals (chrome, copper, nickel, lead and zinc), as well as phosphorous, nitrogen and PM10 particulate matter.
In the literature, there are few studies that link NHL to environmental exposure to chemical substances. Some occupational studies suggest a positive association with exposure to organic solvents, such as benzene [27
], trichloroethylene (TCE), tetrachloroethylene (PCE) and styrene [29
]. Other occupational studies associate exposure to pesticides with an elevated risk of NHLs [31
]. Lastly, different studies addressing the relationship between NHLs and exposure to dioxins furnish contradictory results [33
]. In one study on a large cohort of paper industry workers, mortality from non-Hodgkin's lymphoma and leukaemia was higher among workers with elevated SO(2) exposure, and a dose-response relationship with cumulative SO(2) exposure was suggested for non-Hodgkin's lymphoma. The cohort included 57,613 workers who had been employed for a minimum of 1 year in the pulp and paper industry in 12 countries [35
]. Aside from environmental exposures, there is evidence to indicate that situations associated with chronic antigenic stimulation or immunosupression favour the appearance of these tumours [6
Assessment of exposure to environmental agents that are noxious to human health is a very complex process. At present, there is a great variety of exposure-measurement strategies, depending on the timeliness and availability of resources, which include the use of remote sensors, biomarkers, or estimates of pollutant dispersion using theoretical or statistical models [36
]. With respect to this last avenue of research, there are a number of studies in the literature that seek to estimate the risk associated with proximity to hazardous sites (focused clustering) [2
]. In these and other studies, the authors have explored the idea of estimating risk according to distance [2
At present, the real availability of data from remote sensors or biomarkers is negligible. Hence, in the absence of such information, many studies have used distance as an exposure marker. This approach has been further refined, by endeavouring to model pollutant dispersion using anisotropic models that take data, such as wind direction or geographical relief [17
], into account. These models could not be applied to this study, however, for lack of information of this type.
With respect to our study, using the distance from the industry to the municipal centroid means that, as the study radius is reduced, the number of towns deemed to be exposed falls drastically. This situation leads to the elimination of exposure variables and the impossibility of studying variation in risk according to a more stringent definition of exposure for most of the emissions considered. Based on the results for the two industrial groups analysed at the three distances (production and processing of metals and mineral industries), no conclusion can be reached as to variation in risk with variation in distance to the emission source.
In ecological spatial correlation studies, Poisson regression is one of the basic tools applied to analysing the association between risk of mortality and the various potential risk factors [15
]. This type of regression forms part of so-called generalised linear models and assumes independence between observations or counts, an assumption that could be violated when working with data that have a spatial structure [15
]. Nevertheless, the use of Poisson regression may help obtain an initial assessment of the presence or absence of this association. Indeed, a number of authors have used this method to evaluate the relationship between risk factors and excess incidence or mortality in the study of non-communicable diseases in a spatial context [37
]. The second model used -the mixed model- is included as an intermediate step between a model that assumes total independence and a model that assumes autocorrelation among observations, and has the advantage of circumventing the problems of extra-Poisson dispersion, lending robustness to the estimators and using the provincial level to approach autocorrelation, which amounts to a form of stratification in the comparisons. Lastly, the third model -the BYM model- assumes that each observation is conditionally independent of the others, i.e., that observations are spatially correlated amongst themselves, with the aim of modelling the spatial effect of the risk [24
]. In none of the models, multiple comparison adjustment was considered. The probability of one spurious test result was 0.33. Due to this low probability and the number of comparisons, we decided to asses the adjustment for multiple testing by the consistency of the associations showed by the results of the different models.
In our results for almost all the industrial sectors considered, the related risks were observed to increase as the random effects covered by the spatial structure of the data were included. The relative risks yielded by the mixed model are, in general, higher than those yielded by the Poisson regression, while those yielded by the BYM model are the highest for most of the variables. The inclusion of random spatial effects terms in risk estimation, not only improves the study of the associations between environmental exposures and mortality, but also reduces proneness to "ecological bias" as a result of working on a larger scale and adjusting for unknown confounders which have a spatial distribution different to that of mortality [21
]. However, bearing the similarity of results in mind, the decision to apply the spatial model in exploratory studies of this magnitude must be carefully evaluated, due to the excessive time of computation. The ever increasing availability of health and exposure data calls for the definition of a fast and easy methodology of analysis that would optimise available resources within research groups when it came to embarking upon exploratory studies [23
None of the socio-demographic variables considered in our study appeared to act as a potential confounder, inasmuch as their elimination in the various models led to no substantial changes in the effect estimators of the distance to the industrial foci studied (data not shown). Furthermore, these possible confounding variables, defined a priori, displayed no important direct effect on risk of NHL mortality, registering RRs close to unity.
As stated above, little is known about the possible role of environmental exposures in NHL aetiology, which may be due to the fact most of the studies undertaken to date focused on small towns and poor-quality exposure measures. This implies a limited statistical power that hinders the estimate of modest RRs [8
]. This paper presents a first approach to the exploration of the influence of exposures to industrial air pollution and risk of NHL mortality vis-à-vis the entire population of a country, something that is an advantage in terms of the sheer size of the exposed population but is a drawback in terms of possible misclassification of exposure or the uniqueness of each of the installations.
Other possible limitation is the use of ICD9, that classification has not different code for each type of lymphoma included in the LNH; as a result we can not know the spatial patterns of each individual type. Moreover, mortality data only includes the more aggressive type of lymphoma. Less aggressive lymphomas have a low mortality rate and, consequently, they are not included in this study.
It should also be pointed out that the data referring to environmental industrial exposures were drawn from the first edition of the EPER. The quality of this information may conceivably improve with the new European Pollutant Release and Transfer Register (E-PRTR), which will completely replace the EPER in 2009, thereby allowing for the validity of a study of this type to be enhanced, with the possibility of evaluating the effect of specific pollutants. Moreover, though the "near versus far" analysis conducted in this study assumes all the industries of a single sector to be equal, it must nevertheless be borne in mind that each industrial source has its own characteristics, and subsequent studies will therefore have to address these on a case-by-case basis.
Finally, we should not forget that the use of aggregated data implies important assumptions. We assume that the whole population within a municipality lives in its centroid; even more, we assume that they have always been living there. Also, we do not consider the daily movement of the people to go to work or study, for instance. Hence, we are assuming that everybody within an area is exposed to the same type and amount of pollutant substances.