Document mpREwog1orG4oLK654yNDE1Eg

.7ournal of Epidemiology and Community Health 1994;48:369-376 369 I Leukaemia clusters in childhood: geographical analysis in Britain E G Knox M ~ uCottage, Great Comberton, Wm o1r0ce3sDteUrs,hUirKe. EGb o x Commpondence to: Professor E G fiOx, Mill Cottage, Comberton, Worcestershire -10 3DU, UK. -Abstract Study objective To validate previously demonstrated spatial clustering of childhood leukaemias by showing relative proximities of selected map features to cluster locations, compared with control locations. If clusters are real, then they are likely to be close to a determining -hazard. Design Cluster postcode loci and partially matched control postcodes were compared in terms of distances to railways, main roads, churches, surface water, woodland areas, and railside industrial installations. Further supporting comparisons between non-clustered cases and random postcode controls with those map features representable as sin- -gle grid points were made. Setdng England, Wales, and Scotland -1966-83. Subjects Grid referenced registrations of 9406 childhood leukaemias and nonHodgkin's lymphomas, including 264 pairs (or more) separated by <150 myand grid references of random postcodes in -equal numbers. Main results The 264 clusters showed relative proximities (or the inverse) to several map features, of which the most powerful was an association with railways. The non-railway associations seemed to be statistically indirect. Some railside industrial installations,identified from a railway atlas, also showed relative proximities to leukaemia clusters, as well as to non-clustered cases, but did not "explain" the railway effect. These installations, with seemingly independent geographical associations, included oil refineries, petrochemical plants, oil storage and oil distribution depots, power -stations, and steelworks. Conclusions The previously shown childhood leukaemia clusters are confirmed to be non-random through their systematic associations with certain map features when compared with the control locations. The common patterns of close associationof clustered and non-clustered cases imply a common aetiological com- -ponent arising from a common environ- mental hazard namely the use of fossil fuels, especially petroleum. (3Epidemiol Community Health 1994;48:369-376) Many clusters of childhood leukaemia have been reported, including both simple geographical concentrations and groups confined by joint temporal and geographical limits. They have been extensively discussed.12Some of the former were related to point-source hazards such as nuclear installations; while others were unrelated to any previously declared geographical feature.45Some reports of space-time clusters were simple anecdotes, and although formal verification techniques were sometimes used: clustering was confirmed in some sets of data and not in others."l0 Most studies suffered from an arbitrary choice of the bounds within which the clusters were defined; and without a prior scale specific hypothesis it was difficult to separate the defining criteria from those used for statistical testing. The question therefore remained whether clusters of either kind were real, or whether the "positive" reports represented selective publication of extreme examples from a random overall set. Recent analyses of the full set of leukaemia registrations in England, Wales, and Scotland over a period of 18 years, however,' have again shown joint spatial-temporal clusters of regist- rations spanning approximately 0-5km and 60 days,"12bounds similar to those identified in other studies. These findings were then used to declare a scale specific prior hypothesis suitable for testing the presence of pure geographical clustering. This was based upon the argument that if space-time clusters and longer term geographical concentrations coexisted within the same set of data, then they could probably be regarded as separate expressions of a common determining mechanism. The approach was successful in that it showed a significant excess of geographically defined pairs and triplets within this same MITOW range." These findings were based upon map references attached by the Post Office to the postcodes of the registration addresses, and stored in the National Childhood Cancer Register. Only the map references were released for this research; not the postcodes themselves. Unfortunately, geographic precision was limited by the frequent allocation of several postcodes to one map reference. Although it was possible to identify the several postcodes attached to a given map reference, and their several address ranges, it was often not possible to identify the registration postcode itself, to check its precise location, or to calculate more exactly the distances between pairs of registration addresses. Important questions therefore remained unans- wered. Might the "clusters" represent a variation of postcode sizes only, or of child-resident addresses within certain postcode groups - for example, constellationsof tower blocks of flats. -1 I 370 Knox Clusters defined on so small a scale were too close for comfort to the resolution limits of the map coordinates themselves. Although the National Childhood Cancer Register was set up to enable research of this kind - and indeed it has no other purpose - the Department of Health (the arbiter in these matters) most unfortunately refused to sanction any research use of the recorded postcodes unless specific permission was first obtained from the ethical committees in every one of some 200 NHS health districts. This requirement was considered impractical. It was unlikely that the total response needed to justify a general release would be obtained; and specific permission relating to a subset of individual children could not be sought because it was not possible to identify them or their physicians to the committees. If the matter was to be carried further, and the spatial clustering validated, it would be necessary to devise some means of enhancing the positional and relational specificity of the map references within the censored data set. This paper shows how such a technique was devised, and it records the results of its subsequent application. Methods The registration materials have been described in detail elsewhere.' They consist of a computer file of all registrations of leukaemias and reticulo-endothelialturnours in children aged &14 years in England, Wales, and Scotland in the 18 year period 1966-83. Individual records included the date of registration, age at registration in months, the sex of the child, the ICD code of the tumour, a map reference derived from the postcode of the address at registration, and the census enumeration district code. Certain materials, including exact date of birth, the postcode itself, the address, and identifying data, were not released. This national file provided the data on which the earlier investigations, and the demonstrations of space-time and geographical clustering, had been based. The present investigation uses the same set of data and it seeks to confirm the reality of the clusters by demonstrating proximities to putative hazards associated with particular cartographic features. It achieves its enhanced specificity through postulating that if the previously demonstrated short range geographical concentrations are genuine, then close pairs and triplets of cases are also likely to be close to a determining hazard; and that these hazards might be recognised, or at least indicated, through inspecting maps or the sites themselves. The present report is based upon map inspections. This method overcomes the problem that many children may have contacted hazards some distance from the given address in the period before registration. While many day to day movements will be within a few cluster diameters of the home address, others must involve more distant travelling associated with school attendance, shopping, holidays, and house moves. This will necessarily dilute any true address-proximity effects. However, by concentrating on clusters, with the inference of a nearby hazard common to the pair or triplet of cases, the likelihood of such detection is greatly amplified, and the specificity of any findings much enhanced. The method selects only the most informativeevents and locations. The lack of positional accuracy resulting from the cen- sorship rules is thus compensated for by a much improved specificity, within this subset, of any detected proximities to apparent external hazards. Firstly, the file was searched for all pairs of registration coordinates separated by less than 0.15km. This includes all those separated by < =0.1 km in both easting and northing, and the de facto separation limit of such pairs is 0.142 km. This is close to the resolution limits of the postcode coordinates, which are recorded to an accuracy of 0.1 km over the greater part of the country. Accepted pairs were those in which both members suffered from acute lymphatic or acute unspecified leukaemia, or non-Hodgkin's lymphoma: and the more northerly registration was accepted as the position of the cluster. Some of the pairs were components of larger sets and these groups were condensed to a single coordinate. Experiments with pairs within limits greater than 0.15 km (for example, 0-3km) resulted in numbers of larger connected groups of limited positional precision, and for this reason the narrower limit was preferred. One concordant twin pair was excluded, as was an artefactual cluster resulting from the assembly of children of armed forces personnel and their families in a favoured treatment centre. This left a working list of 264 locations, each accommodating two or more cases separated by nominal distances of 0.142 km or less. Control postcodes were chosen as those filed alternately 10000 before and after the cluster postcode (the first one to carry the cluster map reference), in a serially numbered and easting- sorted file containing 1.28 million residential codes. This was designed to achieve a degree of social and epidemiologicalpair matching, while leaving sufficient distance for them to differ regarding the presence of short range hazards. The clusters and the control postcodes were plotted on Ordnance Survey (OS) maps (1 :50000 Landranger Series). They were widely scattered throughout the country. For each, distances were carefully measured to the nearest: (a) railway in current use, (b) railway in current or earlier use, (c) surface water, including stream, river, canal, pond, estuary, or coast, (d) wooded area, (e) "A" class road, and (0 church. All of them occur within a few cluster diameters of a large proportion of the general population, unlike sparse features such as nuclear power stations. Rail and road proximities were sought because of pollution associations, and water and woodland proximities because of possible infective (for example, arthropodborne) associations. Except for churches, which are "points", all these map features have extended forms. Churches were included as a standardising variable, as described later. Measurements were recorded to a precision of 0.01 km, although the dependable accuracy is Leukaemia clusters in childhood 37 1 probably no better than 0.02 km, on a map of this scale; and with the postcode release problems, it is probably no better than 0.10 km on the ground. In the event, the first analyses then demanded secondary searches for proximities to other map features. They were definable (iike churches) as cartographic points; only the map positions were recorded, and distances calculated rather than measured. Details are given later. The results in turn prompted an examination of distancesbetween these same geographical points and the full set of all registered cases, clustered or not, together with a new set of control locations: 9406 of them, excluding Orkney and Shetland. The new control locations were chosen from the full postcode file through a simple random number process, without matching. Results CLUSTER-CONTROL RELATIONSHIPS The mean eastings for the 264 clusters and the 264 cluster-controlswere 427.65E and 42769E, respectively. Mean northings were 304.11N and 334.72N. The difference asymmetry for eastings and northings arises from the prior east-ordering of the file. For the full set of all registrations, mean eastings were 429.46E (cases) and 424.42E (controls); and mean northings were 320.40N and 316.83N.There were no geographical biases here sufficient to perturb any of the subsequent cluster-control or casecontrol comparisons. The cluster-control pair matching succeeded to the extent that the mean interpair distance was 124.51km; while for the cases and their serially corresponding but unmatched random controls, the mean inter-"pair" distance was 239.19 km. CLUSTER-CONTROL COMPARISONS OF MEASURED PROXIMITIES The results of these comparisons are shown in table 1. The most striking finding is a clustercontrol asymmetry for mean distance (km) to the nearest railway; a 42% decrease for railways in current use and a 43% decrease if disused and dismantled railways are included. Many of the disused lines would have been in use during the period of the leukaemia registrations. The difference between the group means, and the deviation of the mean interpair distance from zero, are both highly significant. The churches and "A" class roads show less striking cluster-control distance contrasts although both are in the same direction as the rail difference, and the church difference is statistically significant. The surface water and woodland correlations, significant for the former, are in the reverse direction. It was recognised in planning the study that relative proximity to such as a railway might be determined by the characteristics of adjacent neighbourhoods rather than the nominated feature itself. Church distances are measured because they serve as an indicator (in cities) of high population density, older housing, and less affluent life styles: the kind of areas crossed by railways. It was known from another investigation that neural tube defects, with known social correlations, exhibited significantproximities to ch~rches.'I~t was for their controlling function in this respect that church distances were measured. There was indeed a significant correlation between distance to the nearest church and distance to the nearest railway in the combined set of cluster and control data, directly confirmingthe presence of such an effect. When church distance was used to standardise rail distance, the latter being reduced to an excess/ deficit above/below the value expected from the rail/church regression function, the rail association remained significant. The reverse standardisation diminished the church correlation to non-significant levels; the 2 statistic for rail standardised distance asymmetry was 2.3. The inverse correlationsof surface water and woodlands probably reflect incompatibilities with dense industrialisation, both in real terms and through the difficulties experienced by cartographers, with competition for map space in complex urban areas. The possibility of a postcode artefact must again be considered. As noted already, postcodes in areas with dense transport systemsmay perhaps cover larger numbers of addresses; or postcodes of a given size may accommodate relatively more children. Linear obstructions to postmen's beats (such as railways) must tend to separate individual postcodes rather than traverse them, perhaps creating an ordered "packing" of map references in their immediate vicinities. However, church distance standardisation should have eliminated much of the urban high density effect, and linear packing should affect controls as well as clusters. The absence of a strong association with "A" class Table 1 Cluster-control comparisons: measured distances to map features Nearest map feature Mean distances Active rail Any rail Church Water Km from cluster Km from control Cluster-conuol/control Pairwise t ratio Group I ratio 1.274 2.198 -0.42 3.961 3.261 0.931 1.620 -0.43 4.178 3.450 0.562 0.743 -0.24 2,960 2,143 0,695 +00..52476 3.295 2.101 Wooded 1,342 1.193 +0.12 1.%O 0.940 "A" road 0.614 0.798 -0.23 1.871 1.583 roads, which also traverse population dense areas and tend to "organise" the disposition of postcode map references, again supports the intrinsic specificity of the rail connection. As a further check, mean rail distance comparisons were recalculated separately in areas of very high or lesser population densities and rail network densities - the "very high" comprised map sheets covering London, the West Midlands, Liverpool, Manchester, and Glasgow. The rail distances were greater and the clustercontrol differences more extreme in the less dense areas, with mean cluster and control rail distances (current and disused lines) of 1.16 and 2.28 km (t=2.989). The mean distances in the higher density areas were 0.73 and 0.85km respectively, and were not significantly different. These findings do not suggest a simple artefact, but rather that hazard conditions in extremely dense conurbations are sufficiently severe to mask the effects of railways; or per- haps they seldom permit a sufficient distance between railways and control locations, to show real contrasts. The subsidiary findings in table 1 can probably be regarded as secondary to a primary association between clusters and railways, and the mutual demographic/cartographic associations between railways, roads, churches, and other map features. The evident strength of the rail-cluster correlation, and the failure to dispose of it through a form of density standardisation, suggest that it is statistically real. However, it is not possible to say immediately whether this reflects a direct hazard from the railway itself or some industrial or demographic aspect of the peri-railway environment. There are many different ways in which railways could offer a hazard. They include atmospheric pollution from diesel fuel exhaust, evaporation of volatile cargo, or spillage of liquid cargo including fuel or toxic chemicals or radioactive waste; sewage and viral contamination of lines from passenger-care lavatories; use of herbicides and pesticides on rail side vegetation; rodent control poisons; viral hazards among railside wildlife ecosystems (for example, rats, shrews, mice, insects); theft of goods or trespass by children or members of their families; or capture of rodents by domestic pets, with resulting toxic or infective contamination of the home. The clusters were plotted on a national rail map that distinguished high voltage electrified rail lines (25 kV 50Hz) from lower voltage (750 V dc) "third rail" lines and non-electrified lines, and on regional maps that distinguished industrial from mixed passenger/industrial tracks. The clusters were very widely distributed involving all three main types of locomotive power and both goods-only and mixed goods-passenger lines. That is, the clusters were not specifically associated with either of the electrification formats or specifically with diesel-only tracks; although diesel locomotives can and do travel on any lines, electrified or not. Nor were they specifically associated with sewage leaking from passenger cars. The need to narrow the field further led to the next stage of the study. CLUSTER-CONTROL COMPARISON OF PROXIMITIES TO POINT HAZARDS A number of rail associated potential hazards, distinguishable from the railway itself, were identified through a railway atlas.I5 The particular value of the atlas was its recording of rail sidings and industrial branch lines, together with indicators of their uses and users. These include oil refineries, petrochemical complexes, liquid propane gas terminals and tank farms; local oil terminals and depots for transfer from rail to road tankers; cement works and distribution terminals; storage and distribution sidings for agricultural fertilisers (Kemira depots): fossil fuelled and nuclear fuelled power stations; other nuclear installations, sidings and branch lines assigned to the Ministry of Defence (MOD): road stone, quarry, and coal mine terminals; dock sidings, goods traffic marshalling freight, and container terminals; steel works and other (aluminium, tinplate etc) foundries and smelters; gas and coke works; and others. Many of these sidings are also marked on OS maps, furnishing appropriate coordinates, but are shown in formats which are difficult to detect and impossible to identify without the help of the atlas. Some locations are excluded altogether from OS maps under the terms of official falsification policies relating to (some) nuclear and defence installations. The atlas record was, however, incomplete for present purposes. Some elements recorded on OS maps, for example, several major power stations in the London area and in the West Midlands, are not recorded here, most having been closed down quite recently. Some installations, including docks and liquid propane gas terminals and both nuclear and oil fired power stations, have no rail links and are supplied entirely via sea and estuary, and are legitimately omitted from a rail atlas. Finally, there were railside installations with circular features understood on OS maps as ``tanks'' but without a siding. They are presumably oil, bitumen, other liquid, or powder rail to road transfer terminals; locomotive refuelling points; or railside factories loading directly from tankers halted on the main line or on rail spurs or loops too small to be recorded as sidings. Sewage work tanks are generally well labelled by the OS, although one or two had to be confirmed as such or separated from adjacent oil installations on larger scale maps (1 :25000 OS, or "A to Z" city street maps). Cooling towers are recognised from the size of the circles and from their dispositions. Gas holders are more difficult and not always distinguishable from large liquid fuel tanks. With the natural gas revolution, few coal-gas or coke works are now marked and noted. These various railway associations were identified through a systematic search on OS maps for all features recorded in the atlas; and a supplementary search along all rail lines (approx 3 km either side) on all map sheets with an atlas feature, a case cluster, or a clustercontrol) and a search of all major ports and harbours and navigable waters. Care was taken not to include features such as factories with tanks, identified only because Leukaemia clusrers in childhood 373 they were close to identified clusters and which were not close to railways or docks. This would have introduced bias to subsequent clustercontrol comparisons. Some locations, for example, power stations with oil tanks or marshalling yards incorporating cement or oil terminals, were recorded under more than one head. Small tank installations on rail lines not covered by the terms of the above search will have been missed; nor was it possible to identify gas and coke works with sufficient consistency to warrant their inclusion. Coal mines, quarries, and stone terminals seemed unlikely candidates both on prior and on intuitive/visual grounds, and were not listed. Distances from these hazards to the caseclusters, to their matched control locations, to all registrations, and to their unmatched random controls, were subsequently computed (rather than measured) and attached to clustercontrol lists and case-control lists for subsequent analysis. Table 2 gives the main results for the cluster-control comparisons. The numbers of hazard points of different classes are given in the final row of this table. Very large installations such as oil refineries were sometimes represented as several points. The outstanding contrast between the clusters and their matched control locations relates to oil and petroleum depots and terminals, petrochemical factories, oil storage and unloading farms, and refineries. The cluster-control differences are highly significant on several criteria including between group t tests, paired t tests, and within pair asymmetries tested with ,f. The other putative hazards show much weaker and mainly non-significant differences, some or all of which might be explained as secondary geographical associations. The mean distances shown in table 2 were greater than those shown in table 1, and there was little overlap between the distributions of the distances with which these two tables are concerned. Although the search for rail side hazards was undertaken to "explain" the rail cluster associations in indirect terms, the proximity patterns shown in tables 1 and 2 seem to be separate phenomena. There may be a common causal mechanism, related, for example, to diesel and other oil products but if this is so the two proximity distributions are separate manifestations, statistically independent of each other. INDIVIDUAL CASE-CONTROL COMPARISONS OF POINT HAZARD DISTANCES Confirmation of the findings in table 2 was sought by comparing all 9406 individual cases and their control postcodes, in similar terms. The results are given in table 3. The group means again show highly significant differences for oil installations. The earlier positive fmdings for power stations, rail yards, and steel works are confirmed, as is the absence of any effect for docks and harbours or for nuclear or MOD installations. The earlier positive finding for fertiliser depots is not confirmed and the earlier negative finding for cement works now appears as a hazard. Significance levels are generally higher than in table 2, due to the larger numbers; although the hazard intensities are reduced compared with the studies using cluster enhancement. The refinery hazard showed a rather irregular variation of relative risk (RR) with increasing distances up to about 10km: 422 cases and 360 controls (RR=1.17) within 3.0 km, 256/215 (1.19) at 3 to 5 km, 798/631 (1.26) at 5 to lOkm, and 1600/1491 (1.07) at 10 to 20km. The risk gradients of the lesser oil hazards were more tightly localised: 552/443 (1.25) within 2 km, 485/409 (1.19) at 2 to 3km, 978/874 (1.12) at 3 to 5km, and a reversed ratio thereafter. The Table 2 Cluster-control comparisons: calculated distances to point map features Nearest maP feature Mean distances 'Oil refinery Oil depot Oil either Kemira Cement Power station Foundry/ Docks/ Rail steelwks harbour yards Nuclear imraln Km from cluster Km from control Cluster-control/ control Painvise t ratio Group t ratio Asymmetry x2 45.80 49.49 -0.07 1.384 0.700 2.981 (n=51) 11.36 14.19 9.90 12.73 -0.20 2.821 1,854 7.670 (n=183) -0.22 3.159 2.102 10-281 (n=234) 54.70 64.06 -0.15 2.678 1.444 2.367 (n=16) 25.37 26.92 -0.06 0.909 0,501 0.0 15 (n=44) 23.75 29.88 -0.20 2.860 1.900 2.570 (n=67) 41.42 47.19 -0.12 1,950 1.147 2.004 (n=44) 31.42 33.79 -0.07 1.203 0.615 3.186 (n=82) 15.29 18.80 -0.19 2.518 1.612 0.380 (n=107) 57.20 61.70 -0.07 2.012 0.771 0,015 (n=19) *Oil refinery includes major storage installations. n number Asymmetry of x2 hazard points used in is based on: (cl <co - analysis. cl >co)*/(cl< co + cl > co) Where cl is distance from feature to cluster, and co is distance from feature to control, and cl> co is number of cases where cl is greater than co. Ministry of defence 62.03 63.97 -0.03 0,964 0.303 2.207 (n=25) Table 3 Individual case-control comparisons: calculated distances to point map features Mean distances Nearest map feature 'Oil refinery Oil depot Oil either Kemira Km from cluster &n from control Cluster-control/ control Group t ratio 47.94 49.78 -0.04 2.063 13.47 14436 -0.09 4.976 11.71 13.14 -0.11 5.818 'oil refinery includesmajor storage mstallations. 6059 60.81 - 0.198 Cement 26.93 30.37 -0.11 5,989 ~ Power station 26.21 28.68 -0.09 4.456 Foundry1 sreelwks 45.25 48.82 -0.07 4.032 Docks/ harbour 33.60 33.65 - 0.080 Rail yards 17.15 19.20 -0.11 5.430 Nuclear instaln 60.82 61-95 -0.02 1.115 Ministry of defence 62.27 64.07 -0.02 1.669 I -t , -i 374 Knox matched cluster-control comparisons had shown a similar range of effectiveness with 24/ 10within 3 km of an oil installationof any kind, 23/13 at 3 to 5 km,and no excess beyond 5 km. Refineries and other large installations often spanned several kilometres on the map, sometimes with several refineries in the same area. The difficulties of representing them as single, independent map points may account for the relative irregularity and increased range of their associated gradients, compared with smaller sites. The relative risks for cement works/terminals, rail yards, and steelworks showed no clear distance gradations at short ranges. The variations were irregular and suggested secondary associationsrather than diffusing toxic hazards. Fossil fuelled power stations also showed an irregularity, but one which could be real. There was a modest ratio of 158/144 (1.10) cases/ controls within 2 km, a sharp jump to 232/155 (1.50)at 2 to 3 km, then a reversion to 540/516 (1.05)at 3 to 5 km. These results might feasibly reflect a pattern of significant fallout from tall chimneys. These relationships were explored further through identifying those specific sites whose relative risks, within successive radii, exceeded arbitrary "toxicityyycriteria. The most "toxic" sources are listed in table 4, and the criteria are defined in a footnote. This confirms the high risks associated with severe industrial pollution, notably on Merseyside, in north east London, the steel manufacturing area between Sheffield and Rotherham, and parts of the West Midlands; and it c o n k s the predominance of the oil and petrochemical hazards. It confirms that fertilizer depots, MOD depots, and nuclear installations can be disregarded from this point of view. It shows that the dock and railyard associations are probably indirect: also that the three "toxic" cement works and terminals were close to other more likely sources. An additional examination of the Grimethorpe-Bolsover Coalite plant in Derbyshire, much maligned in the media as a supposed source of dioxin contamination of surrounding farmland, showed no evidence of a leukaemia generating effect. In contrast with the indirect associations, the high risk power stations were often geographically independent of oil installations, and this apparent hazard cannot be dismissed so readily. The high case-control ratios were generally within 5km, where intrinsic effects are less likely to be confounded with those of other sources. The search showed two important heterogeneities. Firstly, certain steelworks showed powerful and probably independent effects while others, notably in Glasgow, the Midlands, and South Wales, failed to do so. Secondly, despite the predominant effects of refinery and petrochemical installations, a group of large refineries seemed to be relatively innocent. This might partly be due to their isolation from large populations, with small numbers of cases and controls, but a grouped assessment of several such refineries(shown in table 4) showed only a short range effect involving relatively few extra cases. Both heterogeneities suggest the likely irnport- ance of manufacturing variations and of special processes within these industries or in nearby ancillary plants; or perhaps different crude oil or other raw materials sources. Within so complex a set of geographical relationships it is difficult to identify causeeffect possibilities among these industrial sources. Caution is clearly necessary. Only the oil and petrochemical installations, the power stations, and possibly some steelworks, show a sufficient strength, consistency, and independence in their leukaemia associations, and sufficient coherence of their risk/distance relationships, to be regarded as genuine potential hazards; and even within these groups there seem to be heterogeneities. Discussion The later steps of this investigation were not envisaged when the study began. The original postulate was that close pairs of leukaemias should themselves be close to a responsible hazard, and appropriate map measurements relating the various cartographic candidates to cluster postcodes and to control postcodes showed a powerful local association with railway lines. A weaker association with churches, and negative associations with open water and wooded areas, were probably the indirect effects of their sharing the high population densities surrounding the rail lines. The question then arose as to whether the high densities or the rail lines themselves should be treated as statistically primary or whether some unconsidered industrial association of the lines might be the true cause; or indeed, whether the findings reflected only some hazard related artefact of the postcode map reference allocation process. The remainder of the study hinged on these questions. In the event, the systematic nature of the findings largely excluded the question of an artefact. A search for rail associated industrial features was conducted using a railway atlas, and their positions subsequently tested against the clusters and control locations. In a supplementary examination, these features were tested against the full set of leukaemia registrations and a fresh set of controls. The most powerful and consistent findings related to oil and petrochemical installations and to power stations. Other findings were interpreted as secondary to these primary relationships and to the mutual geographical associations of different industrial plants. Although these searches had been designed to see whether the railway effect was secondary to those of nearby installations, the disparate scales of the respective proximity distributions showed them as separate phenomena. Neither the rail proximities nor the oil installation or power station proximities readily "explained" the other. Although independent statistically, they could each represent a single general class of environmental hazard, namely the spillage or evaporation of petroleum/tar products from diesel locomotives, rail tankers, or static storage tanks together with the partial combustion, fractionation, cracking, and other chemical pro- Leukaemia clusters in childhood Table 4 ~~ Place Locations with high casejcontrol ratios Grid rej E a t Norrh Case,control &5 km Refineries & oil-srorage/processi~: Stanlow Cheshire 343.9 Stanlow Cheshire 343.9 Stanlow Cheshire 342.5 Ellesmere Port 336.6 Birkenhead 333.4 Teesmouth 453.0 Canvey ICoryton 575.2 West Thurrock 560.4 Purfleet 555.1 Dagenham 548.6 Runcorn 350.3 Partington 373.5 Saltend/Hull 516.2 Mexborough 449.2 Barry, S Glamorgan 3140 Oldbury 399.2 Glazebrook 371.8 Rainham, Essex 551.5 Capenhurst 336.7 Doncaster 460.0 Brentwood 556.5 Edmonton 536.8 Romford 550.7 Waltham Abbey 536.7 Newton Aycliffe 428.4 Widnes 352.6 Famborough 4885 Swindon 416.6 Brornford Bridge 41 1.7 Kingsbury, Warks 422.2 Peterborough 518.5 Staines 503.3 I Brownhills 40343 Ferry Rd, Cardiff 317.4 Colnbrook 503.6 Skellow/Adwick 455.2 St Helen's 351.3 Sunderland 440.8 Sh&eld/Aldwarke 444.5 Swinton 446.3 Rowley Regis 398.1 Bescott/Walsall 400.7 Worksop 460.7 Stanford-le-hope 568.6 Helsby Grouped refineries 34-7.7 375.8 376.8 376.7 380.2 387.1 5234 182.4 178.6 178.4 182.1 380.2 392.3 421.1 399.6 168.6 288.6 392.1 182.3 374.5 406.1 191.4 193.6 187.9 199.4 523.3 385.6 154.2 187.2 290.1 296.8 298.8 172.2 304.2 174.3 175.6 409.5 394.6 557.8 394.4 398.5 286.7 296.6 179.0 181.6 37-4.6 9,l 814 1315 16!11 55/35 21/11 1116 11/12 1216 48/33 12/10 19,'lO 2019 20114 1216 45/51 1617 20jl8 19/12 24/13 20115 61/33 50133 38/15 1216 15/17 22/12 3017 70/40 8/11 26/13 24113 28/13 37/19 20/9 619 23/28 28/23 33116 27/17 33/53 49/29 1116 113 313 90166 Fertilizer depots Helsby Gloucester 347.4 384.5 376.5 218.6 414 24/12 Cement works/terminals Swindon Widnes Northfleet 415.9 351.6 562.4 185.6 385.1 174.4 34/14 21/14 30120 Power starions Drakelow, Burton Willingtn, Burton Leicester Gt Yarmourh Nechels, Birmingham West Bromwich Ocker Hill Fiddlers Ferry Ince Doncaster Mexborough Methil Tilbury Gravesend Dartford Erith Dagenham 423.3 430.6 457.9 653.0 409.9 413.5 3988 354.4 347.5 4567 448.8 338.2 566.2 563.5 555.9 549.9 546.8 319-6 328.9 302.6 305.1 289.8 290.6 294.2 386.5 376.2 403.6 399.8 700.3 175.6 174.4 1765 180.7 182.4 1816 217 53/30 1517 71/43 73/35 11/11 21/15 1016 33118 23/18 1517 13/15 20119 1414 43/21 61/28 Docks Runcorn Garston Seaforth Sunderland Tyne Dock Methil Tilbury West Thurrock 349.5 339.8 332.6 440.9 4344 3375 563.0 558.5 381.6 384.0 397.3 557.5 565.6 699.3 176.0 176.0 2OjlO 35/15 53/34 26/20 32/32 1417 17/21 19116 Railyards Radyr, Llandaf New Yard, Gloucester Cocklebury Yard Bescott Yards Washwood Heath Yard West Yard Worksop Holmes Yard Ditton Edge Hill Yard Tinsley Yard 313.8 384.7 414.2 401.4 411.0 517.9 457.2 497.0 348.4 337.2 441.5 180.0 218.4 185.5 296.0 289.6 300.7 380.6 371.2 384.7 389.8 389.6 29/17 24/11 33/15 50134 14/38 23111 1316 22114 24/12 79/41 35/18 Cawconrrol 0-10 km 56 33 59143 60135 101165 158'107 70/55 51/52 69/37 101/71 2161137 53/37 107!78 48/40 77/61 41/31 1871169 86/68 175196 52/40 45/32 68/45 2441166 141192 123187 44/28 83/70 59/51 40121 1911122 59/35 36/15 100/76 101/54 60/39 117/80 49/32 94/73 65/56 89/58 77/51 1651162 1761136 21/17 34/41 53/32 3041314 62/41 38/22 47/20 87/67 51/29 23/16 58/37 76/40 22/13 2191139 185/120 62/55 84/71 65/42 58/41 76/65 17/12 49/32 54/34 100/70 208/117 2341146 59/39 145175 155195 63/56 127184 16/13 55/35 76/47 66/39 38/22 45/23 1871147 2 121124 37/16 21/16 24119 96/67 198/104 117166 136198 1491109 1641112 214,152 2471149 94/62 130/109 1671123 260j173 4251296 1671129 253/216 66/61 1491108 65/49 4631352 2311203 3421232 141192 70158 1731125 42811389 3021190 2921226 62/49 1631132 108191 42/27 291/215 150180 39/24 220/182 2151145 70151 2151179 88/79 2381174 1501104 168/116 1691124 3441272 3411264 49/35 80159 100172 7341698 130194 57/51 52/24 18511% 133185 41/30 73/60 89/52 30122 3131236 2891207 184152 1561126 160/111 103175 1541111 31/27 116187 127186 2571163 3981274 4651340 185/132 2521162 220/141 1481100 1981143 30129 138188 188/132 85/58 57/49 51/25 3511273 3051129 38/24 63/44 24/25 2191156 2631166 174/103 375 Notes Refinery centroid Shoreline Flare stack Oil storage depot ROeilfisneearyte+rmchinemalical works Refinery E centroid Tank farm Tank farm Tank farm Chemical works Chemical works Chemical works Chemical works Chemical works Chemical works British Tar Products Factory with tanks Factory with ranks Factory with tanks Factory with tanks Factory with tanks Factory with tanks Factory with tanks Factory with tanks Oil terminal/factory Oil terminal (Hartwell) Oil terminal (Shell) Oil terminal (Esso) Terminal (Warks Oil) Oil terminal (Cory) Oil terminal (Cory) Terminal (Charrington's) Oil terminal (BP) Oil terminal (Elf) Oil tenninal (Elf) Oil terminal (pikington) Oil terminal (petroha) Oil-term/steelworks Oil-term/chemical works Tanks nr rail Tanks nr rail Tanks nr rail Tanks nr rail Tanks nr rail See footnote Kemira (nr tanks above) Krmira (nr railyard, below) Castle Cement terminal Blue Circle terminal Blue Circle Works Safhon Lane, disused Oil tanks,no rail link Disused Disused Wednesbury Nr Widnn Nr Runcom Nr oil source, above Nr chemical works, above Thames estuary Thames estuary Thames estuary names estuary Nr oil terminal, above Mersey estuary Mersey estuary Meney estuary Nr oil terminal, above Nr power SUI, above Nr power SUI, above N r tank farm, above North Cardiff Swindon Wednesbury Birmingham Peterborough Lincoln Nr Widnn Liverpool Shelfield nr steelworks 376 Knox Table 4 Locations with high uselcontrol ratios Contd. Place Grid ref East North Case/control Case/control Case/control 0-S km 0-10 km 0-15 km Notes Beighton West Yard Trafford Hainault Stratford Ilford 443.9 457.9 380.0 545.2 538.3 544.6 384.6 403.1 395.6 192.0 184.8 186.7 24/13 27/15 62/32 35/27 95/52 75/47 100/57 50140 1841129 161191 3501308 2471139 180/113 90173 3301242 3721247 6631597 4881377 Nr Sheffield steel complex Doncaster nr power station Freight liner Terminal Essex, NE London E London E London Steelworks Staveley 442.4 375.3 16/10 55/44 122176 SE of Shdield Catcliffe 441.0 389.5 37/23 119166 1691103 Sheffield/Rotherham Tinsley Park 440.2 389.6 31/20 115/65 162199 Sheffield/Rotherham Templeborough 4413 391.4 28/19 109175 168/108 Sheffield/Rotherham Roundwood 445.0 396.2 30/22 78/50 165/119 S~ h~~ ef~~f~ iel~d,l~Rotherham Aldwarke 445.2 395.1 37j23 86/56 1691122 Sheffield/Rorherham Shildon Works 422.6 525.6 1516 27/19 48/39 Nr Bishop Auckland Wednesbury 397-8 2934 44/44 1881170 3371275 Nr Ocker Hill power station - Sites in this list are selected on basis of case/control ratios (RR)and minimum numbers of controls at successive ranges; (RR>= 1.5 ______and controls> =5 at 0-5 km). OR (. RR> = 1.4 and controls > = 25 at 0-10km)OR I (RR . > = 1.3 and controls > =75 at &1~5 h-...,). s.,. duplicates removed. The Grouped refineries comprise several which were not selected or listed on these criteria, namely Fawley, Isle of Grain, Milford Haven, Neyland, Angle, Immingham, Grangemouth, Thameshaven/Canvey (W), and Neath. cessing of fossil fuels. Space-time clustering could presumably have resulted from major discrete escapes. This, if confirmed, would reverse the earlier hypothesis'* that it might represent a response to an infection and perhaps reinfection, rather than a toxic exposure. The subject of petroleum exposure in relation to childhood leukaemia has received only sparse attention in published reports. This is surprising, since the benzene component is a well established leukaemogen. The parallel increases in petroleum and other fossil fuel usage and the incidence of acute leukaemia in both children and adults, are also suggestive. The relationships between leukaemia and toxic exposures were studied in a recent case-control study of a cluster of 14 childhood cases in a restricted area in The Netherlands. This showed excessive exposuresboth to insecticides and to petroleum products.16Other locally published Dutch studies incriminating petroleum were mentioned. There are also several studies showing that petroleum and fuel exhaust exposures are leukaemia hazards in industrial workers, and that not all of it is explained by However, no large scale systematic examinations of the question in relation to childhood leukaemia have been found. The present investigation is probably the first clear demonstration of such a relationship within a total population over an extended period, and at the same time provides a confirmation and a general explanation of the geographical clustering associated with this disease. Thanks are due to Dr G J Draper and the Childhood Cancer Research Group for providing the data on which these analyses were based. The expenses of the work were defrayed through a Leverhulme Emeritus Research Fellowship. 1 Draper G. The geographical epidemiology of childhood leukaemia and non-Hodgkiin lymphomas in Great Bntain, 196643. London: OPCS, 1991. 2 Cuzick J, Hills M. Clustering and clusters -.-S In Draper G ed. The geographical epidemiology of childhood leukaemia and m-Hodgkin lymphomas in Great Britain, 196683. London: OPCS, 1991. 3 Black D. Inuestigntion of the posrible incretued incidence of cancer in West Cumbria. London: HMSO, 1984. 4 Openshaw S, Craft A. Using geographicalanalysis machines to search for evidence of dusters and clustering in child- hood leukaemia and non-Hodgkin lymphomas in Britain. In Draper G ed. Thegeographical CpidnniorOgy of childhood leukaemia and tum-Hodgkin lymphomas in Great Britain, 1%683. London: OPCS, 1991. 5 Alexander FE. Investigations of localiscd spatial clustering, and extra-Poisson variation. In Draper G ed. The geo- graphical epidenhlogy of childhood leukaemia and rum- Hodgkin lymphomas in Great Britain, 1966-83. London: OPCS, 1991. 6 Knox EG. Epidemiology of childhood leukaemia in N o d - umberland and Durham. B r J FIN SOCMed 1964;18:17- 24. 7 Glass AG, Mantel N, G u m FW, Spears GFS. Time-space clusterinrz of childhood leukaemia in New Zealand. '7 Nut Cancer Ikt 1971;47:329-36. 8 Ederer F, Myers MH, Mantel M. Do leukaemia cases come in clusters?Biometrics 196420626-38. 9 Lock SP,Merrington M. Leukaemia in Lewisham(195763) BMJ 1967;3:759-90. 10 Glass AG, Mantel N. Lack of space-time clustering of childhood leukaemia in Los Angela County 196064. Cancer Res 1969;291995-2001. 11 Gilman EA, Knox EG. Temporo-spatial distribution of childhood leukaemia and non-Hodgkin Lymphomas in Great Britain. In Draper G ed. The geographical epidemi- ology of childhood leukaemia and mm-Hodghinlymphomas in Great Britain, 196683. London: OPCS,1991. 12 Knox EG, Gilman EA. Leukaemia clusters in Great Britain. 1. Space-timeinteractions. J Epidemol Community Health 1992;46:566-72. 13 Knox EG, Gilman EA. Leukaemia clusters in Great Britain. 2. Geographical concentrations. J Epidemiol Community Health 1992;46:57>7. 14 Knox EG, Lancashue RJ. Epidemiology of congenital mayor- matiom. London: HMSO, 1990. 15 Baker S K . Rail atlas. Great Britain and Ireland. 7th Ed. Oxford Oxford Publishing Company. 1992. 16 Muld? YM, Drijver M, Kreis IA. Case-controlstudy on the associauon between childhood haemopoietic malignancies and local, environmental factors in Aalsmeer, The Nether- lands. J Epi&l Community Health 1994:160-4. 17 Lidquist R, Nilsson B, Eklund G, Gahnon G. Acute leuk- aemia in professional drivers exposed to gasoline and diesel. Eur J Haematol 1991;47:9&103. 18 Wongsrichchanalai C, Delzell E, Cole P. Mortality from leukaemia and other diseases among workers at a petro- leum refinery.JOccup Med 1989;31:106-111. 19 Austin H, Cole P, McCraw OS. A case-control study of leuksemiaatanoilrdinery.JOccup Med 1986;28:1169-73. 20 McCraw OS, Joyner RE, Cole P. Excess leukaemia in a refinery population. J Occup Mad 1985;27:220-2.