Article Text

Download PDFPDF

Geographically based approaches can identify environmental causes of disease
  1. Manolis Kogevinas1,
  2. Neil Pearce2
  1. 1Respiratory and Environmental Health Research Unit, Municipal Institute of Medical Reseach (IMIM), Barcelona, Spain and Department of Social Medicine, Medical School, University of Crete, Herakleion, Greece
  2. 2Department of Biomedical Sciences and Human Oncology, University of Turin, Italy and Centre for Public Health Research, Massey University Wellington Campus, Wellington, New Zealand
  1. Correspondence to:
 Dr M Kogevinas
 Respiratory and Environmental Health Research Unit, Municipal Institute of Medical Reseach (IMIM), 80 Dr Aiguader Rd, Barcelona 08003, Spain;

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

There is growing evidence to suggest the potential importance of specific environmental exposures in the causation of cancer in children.

The causes of child leukaemia and, more generally of malignant neoplasms in childhood have been long investigated. Various theories have been proposed on the aetiological mechanisms and risk factors for these neoplasms, including the influence of genetic, prenatal, and postnatal exposures. With a few exceptions, particularly for in utero exposure to ionising radiation and some genetic syndromes, strong and consistent evidence on the importance of specific risk factors has not been produced. One recent exception is the association of ambient air pollution with child leukaemia, an association that is also supported by experimental evidence. For example, a recent experimental study1 showed that filtration of particulate matter in ambient air significantly reduced heritable mutation rates in laboratory mice housed outdoors near a major highway and two integrated steel mills. On the other hand, extremely low frequency electromagnetic fields (ELF-EMF) emitted by power lines have received a great deal of attention in recent years, but the evidence of an association with childhood leukaemia remains uncertain; furthermore, even if ELF-EMF are causally associated with child leukaemia they could explain only a very small fraction of all child leukaemias, perhaps around 1%.2 Exposure to pesticides, both general and specific, has also been examined extensively in relation to childhood cancers, but the evidence is far from conclusive.3

The paper by George Knox in this issue4 follows other recent work examining proximity to industrial sources of exposure to toxic atmospheric emissions and childhood cancer.5 In this paper he expands on this work by evaluating the effects of specific exposures and their combinations.

Knox applies an approach that differs from that of other studies in two respects. Firstly, the data available are extensive both regarding the number of cases and the number of sources and exposures examined. Secondly, he uses a migration index as the main source for the calculation of the relative risk.

Knox has assembled an impressive combination of records including data on hotspots derived from information on atmospheric emissions, trade directories, and map inspections. He has then used this to examine birth and death records for children dying from cancer and has classified each address with regards to its proximity to hotspots for atmospheric emissions, particularly focusing on migration in and out of hotspots.

The migration index that is used in this analysis could be seen as a type of a case-crossover study design. Children moving houses are examined regarding the distance of their houses from specific sources at birth and at the time of cancer diagnosis. The “case” address is the address the child lived in at the time of birth, and exposures at this time are assumed to be aetiologically relevant to the subsequent cancer. The “control” address is the address that (the same) child lived in at the time of death from cancer, and exposures at this time are assumed to not be aetiologically relevant. If these assumptions are correct, then the informative “case-control pairs” are those in which the “case address” at time of birth and the “control address” at time of death are different. For a particular definition of exposure (for example, living within 1.0 km of a hotspot), the ratio of the number of discordant “pairs” in which the “case address” is exposed (and the “control address” is not) to the number of discordant pairs in which the “control address” is exposed (and the “case address” is not), yields the odds ratio in a manner analogous to a “matched pair” analysis of a more orthodox case-control study. This in turn can be applied to the overall prevalence of “exposure” among all cases (not just those who moved) to yield an estimate of the population attributable risk.

There are several important assumptions underlying this comparison. Firstly, it assumes that exposures occurring at birth are more aetiologically relevant than those occurring close to the time of cancer diagnosis and death. When this assumption is invalid (for a particular exposure), the findings reported here will also be invalid. In the most extreme case, in which, for a particular exposure, only very recent exposure is aetiologically relevant, the method of Knox will produce the inverse of the true odds ratio. Nevertheless, for most exposures, and for most types of childhood cancer, this assumption is reasonable, and there is also some experimental evidence supporting the importance of early, including prenatal, exposures.6 Secondly, Knox’s method assumes that there is short term migration equilibrium among the general child population. This assumption would not be valid if, for example, there was a general tendency in the population to move away from environmental hotspots. Thirdly, it also assumes that there is no specific tendency for families to move away from environmental hotspots after the diagnosis of cancer in the child. An attempt to address these latter two assumptions is the restriction of the analysis to short migrations of less than one kilometre. Furthermore, for any such tendencies to produce significant bias they would need to apply to both identified and non-identified environmental hotspots. Many of the exposures considered (for example, hospitals) might be assumed to fall into the latter category. Finally, one must also consider the problems related to exposure misclassification attributable to lack of information on duration of residence and in the quantification of environmental exposures, although the resulting misclassification is likely to be non-differential and therefore to produce a bias towards the null.7

Notwithstanding these potential problems, the approach followed by Knox is innovative, and the findings are of considerable interest. He shows a number of interesting associations of environmental exposures with childhood tumours. He identifies several specific chemicals as being closely associated with child tumours, particularly butadiene and exposures from internal combustion engines. More than these specific associations, which may or may not be supported by further research, what makes this paper unique is its capacity to combine in a meaningful way a series of routinely collected data on births, deaths, residences, sources of exposures, and specific exposures.

How valid can all this be? More generally, how valid can geographical comparisons be for the identification of the causes of diseases and particularly child tumours? An extensive application of small area statistics was first proposed as a proactive measure after the Seascale leukaemia cluster (Black report). Collection and systematic analysis of such data are undoubtedly important for surveillance within the context of proactive research. However, how efficient has this approach been in identifying new (or old) causes of disease? Or at least what new hypotheses have been generated through the application of small area statistics type approaches? Probably few, at least when small area statistics are applied using comparatively crude exposure scenarios such as homocentric circles around a single point source, as has frequently been the case. There are several reasons why few, if any, risk factors have been identified by applying such approaches. Firstly, many of the environmental exposures examined in relation to cancers and reproductive outcomes such as congenital malformations, are probably not associated with high risks at common environmental levels of exposure. Secondly, such studies entail a comparatively high misclassification of the risk factor(s) of interest and of potential confounders. Finally, many of the hypotheses examined are frequently based on weak a priori evidence and are unlikely to be correct. This latter issue may not be perceived necessarily as a significant problem as small area statistical data are readily available and their use could be seen as a screening tool to test new hypotheses and/or generate them. However, a weak risk factor with a great deal of exposure misclassification is bound to produce weak results, which may not be reproduced, even if the exposure is common and therefore has a comparatively high population attributable risk. Nevertheless, small area based approaches are becoming more attractive as methods evolve, both with regard to the availability of records and methods of statistical analysis. Studies such as those by Knox clearly indicate the potential importance of specific environmental exposures in the causation of cancer in children. It will be important to further investigate these findings, and to further test their validity with other studies that have different methodologies and assumptions.

There is growing evidence to suggest the potential importance of specific environmental exposures in the causation of cancer in children.



  • Funding: Neil Pearce’s work on this paper was supported by the Health Research Council of New Zealand, and the Progetto Lagrange, Fondazione CRT/ISI. M Kogevinas was partially supported by ISCIII (Red de Centros RCESP C03/09), Madrid.

  • Competing interests: none.

Linked Articles