Radiat Environ Biophys (1998) 37: 71–74

© Springer-Verlag 1998

R E V I E W A RT I C L E

Freda Alexander

Clustering of childhood acute leukaemia The EUROCLUS Project

Received: 10 January 1998 / Accepted in revised form: 10 February 1998

Abstract The EUROCLUS Project is a collaborative endeavour in which incidence data for 13 351 cases of childhood leukaemia (CL) diagnosed between 1980 and 1989 in 17 countries were referenced to 26 425 small geographic areas and tested for evidence of spatial clustering. A second objective of EUROCLUS was to determine whether clustering of CL was associated with community demographic features and/or proximity to putative environmental hazards. The results show statistically significant evidence of clustering, but the magnitude is small (extra-Poisson variability = 1.65% of Poisson variability). Patterns of clustering are associated with population density and other demographic features which could indicate variations in opportunity for exposure to common infections. There is no consistent evidence that clusters are associated with proximity to nuclear facilities or other putative environmental hazards.

Introduction

Reports of clusters of childhood leukaemia (CL) have been frequently made throughout this century [1] and have been interpreted in terms of either exposure to infectious agents [2] or common environmental exposure to localised leukaemogens [3–6]. The chance proximity of cases with distinct causes remains a plausible alternative explanation. In order to determine whether this is realistic, it is essential to test large data sets for evidence of a generalised tendency to cluster. This need provided the primary motivation for the EUROCLUS project. So far, only one large data set has been analysed in this way [7], and its results have been somewhat equivocal. The second objective of EUROCLUS was to interpret clustering of CL in terms of two distinct groups of putaF. Alexander (½) Department of Public Health Sciences, University of Edinburgh, Teviot Place, Edinburgh EH8 9AG, UK

tive causative factors: infectious agents or fixed environmental hazards. Although many leukaemias in animals are caused by viruses [8] and related diseases in humans are caused by retroviruses [9] and bacteria [10], there is no clear candidate agent for CL. Thus, it is necessary to use proxies when investigating the possibility of infectious agent(s) as causes of CL and CL clusters. Both biological arguments [11] and epidemiology [12–14] focus on patterns of exposure to common infectious agents. For these, population density, community isolation and changes of these with time are appropriate proxies. We summarise here some of the pertinent results of EUROCLUS.

Methods

Geographically referenced population-based incidence data have been assembled for 12 whole countries and for defined geographical areas within a further 5 countries for the period 1980–1989. Apart from Queensland in Australia, these countries are all in Europe. Counts of cases in small census areas have been analysed using the PotthoffWhittinghill test [15, 16] to determine whether cases show evidence of spatial aggregation within small census areas, to estimate the magnitude of the extra-Poisson variation and to compare evidence for such variation between – and within – subgroups determined by age-at-diagnosis and cell type [17]. The above analyses have been repeated with adjustment for overall variation of incidence rates according to strata of population density. In addition, these strata were analysed separately to examine heterogeneity of incidence (clustering) by small area within strata; here the ratio relative to that in the whole country or region was investigated. Strata chosen in advance of data analysis were 0–150, 150–500, >500 persons/km 2. Following the initial analyses (Alexander et al., manuscript submitted), additional strata were added: 0–50, 50–100, 100–150, 150–250, 250–500, 500–750, 750–1000, >1000 persons/km2.

72

To investigate clustering further, up to 25 small areas with the greatest evidence of clustering were selected from data for 14 countries. Identification of the cluster areas used a newly developed algorithm (Wray et al., manuscript submitted). These areas were compared with matched control areas using additional data collected for cluster and control areas by observers blind to area status. The additional data related, firstly, to demographic factors previously considered by Kinlen [12] and others (e.g. [14]) and, secondly, to opportunity for environmental exposure to known leukaemogens – including proximity to nuclear facilities, chemical and petrochemical plants, and fertilizer use. Conditional logistic regression was applied to analyse comparisons of cluster and control areas (Alexander et al., manuscript submitted).

Results

The total data set revealed statistically significant evidence of spatial clustering, but this was of small magnitude (Table 1). Clustering did not appear to be focussed in specific subgroups determined by age-at-diagnosis or cell type (data not shown). Stratification by population density did not change the results substantially. Examination of heterogeneity (clustering) within individual strata determined by population density showed that it was concentrated in the stratum 150–500 persons/km2 and, particularly, 250–500 persons/km2 (Table 2). There is no evidence of heterogeneity within any of the densely populated strata (≥500 persons/km2, 500–750 persons/km2, 750–1000 persons/km2, ≥1000 persons/km2); however, the incidence Table 1 Generalised clustering of childhood leukaemia (CL) Extra-Poisson variation (%) P (90% CI)

n cases

Adjusted for

1.65 (0.22–3.08) 1.50 (0.10–3.01)

13351 12728

country country, population density a

0.03 0.04

a

Table 3 Incidence of CL by population density strata Density stratum

Incidence ratioa and 95% CI

0–50 50–100 100–150 150–250 250–500 500–750 750–1000

1.00 1.05 1.07 1.06 1.06 1.16 1.03

(0.91–1.10) (0.97–1.14) (0.99–1.16) (0.99–1.15) (0.99–1.13) (1.06–1.26) (0.94–1.13)

a

Incidences in this stratum compared to reference strata (>persons/ km2) after adjusting for country. P-value for heterogeneity = 0.06, and for quadratic trend = 0.02 Table 4 Cluster and control areas status by levels of population density in 1975 and 1989

1975 density 1989 density

OR/single level a (95% confidence interval)

P value

0.47 (0.23–0.96) 2.50 (1.22–5.13)

0.03 0.007

a

Multivariate analysis for trends across 4 levels of population density (1 = 0–150, 2 = 150–500, 3 = 500–750, 4 = 750 and higher persons/km2); each trend adjusted for the other (see Alexander et al., manuscript submitted)

rates were significantly elevated in the moderately densely populated strata (Table 3) and there is evidence of a quadratic effect with incidence peaking at this point. When selected cluster and control areas were compared, cluster areas showed evidence of lower population density before the study period together with higher population density at the end of the period (Table 4). Similar associations were found for more subjective indicators of population mixing in relatively isolated communities (data not shown). Although more cluster than control areas were close to nuclear facilities, the association does not approach conventional levels of statistical significance (OR = 4.0, 95% CI = 0.45–35.7). No other ORs for proximity to possible environmental sources of known leukaemogens were significantly or substantially elevated [18].

Population density was not available for small areas in Slovakia and Slovenia

Table 2 Generalised clustering of childhood leukaemia within strata determined by population density Population density (persons/km2)

Extra-Poisson variation (%)

Pa

500+ 150–500 0–500 150–250 250–500 500–750 700 and above

0.69 3.94 0.36 0.89 7.31 0.03 –0.54

NS 0.01 NS NS <0.001 NS NS

a Monte-Carlo P-values >0.1 are indicated as NS (not statistically significant) For details, see [17], (Alexander et al., manuscript submitted)

Discussion

These results represent analyses of the largest data set of CL incidence data to have been assembled for the investigation of clustering. The generalised clustering is statistically significant, but the extra-Poisson component of variation is small (just 1.65% of Poisson variation). Thus, most of the variation in incidence of CL can be attributed to Poisson variability, and isolated intense clusters are unusual phenomena. The association with population density is interesting. A general elevation of incidence is seen for densely (but not the most densely) populated areas (500–750 persons/ km2), and this appears as a uniform increase in risk. The

73

somewhat less densely populated areas (250–500 persons/ km2) show a heterogenous pattern with some areas, but not all, having marked increase in risk; this heterogenous pattern appears as clustering within this population-density stratum. When clustered areas were selected in each national data set, a proportion of them will have had genuinely elevated risk, while others represent the play of chance on normal underlying risk. We cannot determine which group any individual area falls into. The selection algorithm was constructed and validated using computer simulations of random and non-random distributions (Wray et al., manuscript submitted) and appeared to perform well. In addition, statistically significant space-time overlaps of hypothesised times-at-risk of cases in the selected areas [19] suggest genuine excess risk in many of the selected areas. These spacetime interactions and the interplay between low population density at outset and increases during the study period (Table 4) are all consistent with hypotheses relating CL to epidemic patterns of common infectious agents. Neither proximity to nuclear facilities nor to other putative environmental leukaemogens provide convincing explanations for the status of the clustered areas. These data suggest that demographic patterns acting as modifiers of risk of exposure to infectious agents are determinants of clustering of CL rather than nuclear facilities or other industrial plants. The possibility of some degree of synergism between the two cannot, of course, be excluded. Acknowledgements This work was funded by the European Commission BIOMED programme as project number PL93-1785, and describes a lecture given at Hamburg. I should like to acknowledge the support provided by all collaborators listed in the Appendix. The main joint reports of EUROCLUS are in the collaborative papers cited, which contain all the results reported here.

9. Robert-Guroff M, Gallo RC (1983) Establishment of an etiologic relationship between the human T-cell leukemia/lymphoma virus (HTLV) and adult T-cell leukemia. Blood 47: 1–12 10. Wotherspoon AC, Doglioni C, Diss TC, Pan LX, Moschini A, Deboni M, Issacson PG (1993) Regression of primary low-grade B-cell gastric lymphoma of mucosa-associated lymphoid tissue type after eradication of Helicobacter pylori. Lancet 342: 75–77 11. Greaves MF (1997) Aetiology of acute leukaemia. Lancet 349: 344–349 12. Kinlen LJ (1995) Epidemiological evidence for an infective basis in childhood leukaemia. Br J Cancer 71: 1–5 13. Kinlen LJ, Dickson M, Stiller CA (1995) Childhood leukaemia and non-Hodgkin’s lymphoma near large rural construction sites, with a comparison with Sellafield nuclear site. Br Med J 310: 736–738 14. Alexander FE, Ricketts TJ, McKinney PA, Cartwright RA (1990) Community lifestyle characteristics and risk of acute lymphoblastic leukaemia in children. Lancet 336: 1457–1462 15. Potthoff RF, Whittinghill M (1966) Testing for homogeneity. I. The binomial multinomial distributions. Biometrika 53: 167– 182 16. Muirhead C, Butland BK (1996) The Potthoff-Whittinghill method. In: Alexander FE, Boyle P (eds) Statistical methods of investigating localised clustering of disease. IARC, Lyon 17. Alexander FE, Boyle P, Carli P-M, Coebergh JW, Draper GJ, Ekbom A, Levi F, McKinney PA, McWhirter W, Michaelis J, Peris-Bonet R, Petridou E, Pombe-Kirn V, Plesko I, Pukkala E, Rahu M, Storm H, Terracini B, Vatten L, Wray N, on behalf of the EUROCLUS project (1998) Spatial clustering of childhood leukaemia: summary results from the EUROCLUS project. Br J Cancer 77: 818–824 18. Alexander FE, Boyle P, Carli P-M, van der Does van den Berg A, Draper GJ, Ekbom A, Levi F, Brewster D, Michaelis J, Prastard C, Petridou E, Pukkala E, Storm H, Terracini B, Vatten L, on behalf of the EUROCLUS project (1998) EUROCLUS: clustering of childhood leukaemia in Europe (PL93-1785). Report to the EU BIOMED publication (in press) 19. Alexander FE, Boyle P, Carli P-M, Coebergh JW, Draper GJ, Ekbom A, Levi F, McKinney PA, McWhirter W, Magnani C, Michaelis J, Olsen JH, Peris-Bonet R, Petridou E, Pukkala E, Vatten L, on behalf of the EUROCLUS project (1998) Spatial and temporal patterns in childhood leukaemia: further evidence of an infectious origin. Br J Cancer 77: 812–817

References

Appendix

1. Alexander FE (1993) Viruses, clusters and clustering of childhood leukaemia: a new perspective? Eur J Cancer 29: 1424–1443 2. Caldwell GG (1990) Twenty-two years of cancer cluster investigations at the centre for disease control. Am J Epidemiol 132: 543–547 3. Gardner MJ (1989) Review of reported increases of childhood cancer rates in the vicinity of nuclear installations in the UK. J R Stat Soc 152: 307–325 4. Michaelis J, Keller B, Haaf G, Kaatsch P (1992) Incidence of childhood malignancies in the vicinity of West German nuclear power plants. Cancer Causes Control 3: 255–263 5. Lagakos SW, Wessen BJ, Zelen M (1986) An analysis of contaminated well water and health effects in Woborn, Massachusetts. J Am Stat Assoc 81: 583–596 6. Mulder YM, Drijver M, Kreis IA (1994) Case-control study on the association between a cluster of childhood hematopoietic malignancies and local environmental factors in Aalsmeer, The Netherlands. J Epidemiol Commumity Health 48: 161–165 7. Draper GJ (1990) The geographical epidemiology of childhood leukaemia and non-Hodgkin lymphomas in Great Britain 1966–83. OPCS, HMSO, London 8. Temin HM (1992) Keynote address: why are there so many leukaemia viruses? Leukaemia 6: 54–55

Collaborators in the EUROCLUS Project Australia

Cancer Registry of Queensland Dr W McWhirter

Denmark

Danish Cancer Registry Dr H Storm Dr JH Olsen

England & Wales

Childhood Cancer Research Group Dr GJ Draper Dr CA Stiller

Estonia

Department of Epidemiology and Biostatistics Institute of Experimental and Clinical Medicine Professor M Rahu

Finland

Finnish Cancer Registry Dr E Pukkala Dr L Teppo

France

Registry of Haematopoietic Malignancies Professor PM Carli Dr G Couillault Dr M Maynadié

74 Germany

National Register of Childhood Malignancies Professor Dr J Michaelis Dr I Schmidtmann

Greece

Special Data Collection Dr E Petridou

Italy

European Institute of Oncology Professor P Boyle Childhood Cancer Registry of Piedmont Professor B Terracini Dr C Magnani

Netherlands

Dutch Childhood Leukaemia Study Group Dr A Van Der-Does-Van Den Berg Department of Epidemiology and Biostatistics, Erasmus University Dr JW Coebergh

Norway

Norwegian Cancer Registry Dr L Vatten

Scotland

Co-ordinating Centre Dr FE Alexander Dr N Wray Scottish Cancer Registry Dr D Brewster Dr PA McKinney

Slovakia

The National Cancer Registry of Slovakia Dr I Plesko

Slovenia

Cancer Registry of Slovenia Professor Dr V Pompe-Kirn

Spain

Childhood Tumour Registry of Valencia Dr R Peris-Bonet

Sweden

Department of Cancer Epidemiology, University of Uppsala Dr H-O Adami Dr A Ekbom Swedish Cancer Registry Dr J Bring

Switzerland

Registres Vaudois et Neuchatelois des Tumeurs Dr F Levi