Risk of lung cancer associated with residential radon exposure in south-west England: a case-control study.

Studies of underground miners occupationally exposed to radon have consistently demonstrated an increased risk of lung cancer in both smokers and non-smokers. Radon exposure also occurs elsewhere, especially in houses, and estimates based on the findings for miners suggest that residential radon is responsible for about one in 20 lung cancers in the UK, most being caused in combination with smoking. These calculations depend, however, on several assumptions and more direct evidence on the magnitude of the risk is needed. To obtain such evidence, a case-control study was carried out in south-west England in which 982 subjects with lung cancer and 3185 control subjects were interviewed. In addition, radon concentrations were measured at the addresses at which subjects had lived during the 30-year period ending 5 years before the interview. Lung cancer risk was examined in relation to residential radon concentration after taking into account the length of time that subjects had lived at each address and adjusting for age, sex, smoking status, county of residence and social class. The relative risk of lung cancer increased by 0.08 (95% CI -0.03, 0.20) per 100 Bq m(-3) increase in the observed time-weighted residential radon concentration. When the analysis was restricted to the 484 subjects with lung cancer and the 1637 control subjects with radon measurements available for the entire 30-year period of interest, the corresponding increase was somewhat higher at 0.14 per 100 Bq m(-3) (95% CI 0.01, 0.29), although the difference between this group and the remaining subjects was not statistically significant. When the analysis was repeated taking into account uncertainties in the assessment of radon exposure, the estimated increases in relative risk per 100 Bq m(-3) were larger, at 0.12 (95% CI -0.05, 0.33) when all subjects were included and 0.24 (95% CI -0.01, 0.56) when limited to subjects with radon measurements available for all 30 years. These results are consistent with those from studies of residential radon carried out in other countries in which data on individual subjects have been collected. The combined evidence suggests that the risk of lung cancer associated with residential radon exposure is about the size that has been postulated on the basis of the studies of miners exposed to radon.

Residental radon and lung cancer 395 have little or no effect on the risk of the disease (Tomasek et al. 1994: Lubin et al. 1995a).

Study subjects
At each of the five centres in Devon and Comwall where investigation and treatment of lung cancer is carried out, all subjects aged less than 75 years who were referred with a suspected diagnosis of lung cancer during a 4-year period were identified each week by local research assistants. The centres and periods involved were: Plymouth July 1988-June 1992. Barnstaple May 1989-April 1993. Truro May 1989-April 1993. Torquay June 1989-May 1993and Exeter July 1989-June 1993. Subjects were eligible for the study if they were current residents of the counties of Devon or Cornwall and had lived in either county for at least 20 years during the 30-year period ending 5 years previously. Only subjects who were ethnically white were included in the study as very few residents of Devon or Comwall are from other ethnic groups. making the identification of control subjects of similar age. sex and ethnic group extremely difficult. In all. 2959 subjects with suspected lung cancer were identified (Table 1). Of these. 1175 (39.7%) did not satisfy the residence requirements. A total of 1412 (47.7%) did satisfy the requirements and were interviewed by a local research assistant using a structured questionnaire. The remaining subjects were not interviewed for a variety of reasons: the responsible medical staff withheld permission in 260 cases (8.8%). usually because the subject was very ill; the research assistants thought a further nine subjects (0.3%) were too ill to question: 31 (1.0%) died before they could be questioned; 68 (2.3%) did not wish to participate: and four (0.1%) were non-white and therefore ineligible.
For each subject with suspected lung cancer who was interviewed, a control was sought from hospital patients of the same sex, bom within 5 years of the case. who satisfied the study residence requirements. and whose current hospital admission was for a disease not known to be strongly associated with smoking. Patients whose current hospital admission was for a disease closely associated with smoking (see Table 2 for list of diseases) were excluded so that smokers would not be over-represented in the hospital control group compared with the population from which they were drawn. As referral pattems differed between patients with suspected lung cancer and other diseases, patients at each centre were also matched on two or three broad residential areas, based on county districts appropriate for the relevant centre: namely (1) Plymouth vs elsewhere.
(3) West Cornwall (Kerrier and Penwith) vs mid-Cornwall (Carrick) vs elsewhere, (4) Torbay vs elsewhere and (5) Exeter vs elsewhere. To select hospital controls. each research assistant had a list of hospital wards. Each week. as the starting point, one ward was selected randomly, with probability proportional to the number of beds. Patients in that ward were then considered systematically. to see if any fulfilled the matching criteria, and then patients on the next ward in the list were considered. and so on. A total of 2401 subjects were approached as hospital controls, of whom 1418 (59.1%) were interviewed: 881 (36.7%) did not satisfy the residence requirements: permission to interview was withheld by the medical staff for 65 (2.7%); 35 (1.5%) did not wish to participate: and two (0.1%) were non-white and therefore ineligible. Some time after the interview, the hospital case notes of the subjects with suspected lung cancer were reviewed (see Table 2).
For 982 of the 1412 subjects. the final diagnosis was primary cancer of the trachea. bronchus or lung [International Classification of Diseases. 9th revision. code 162 (World Health Organization. 1975). but excluding carcinoids]. The original pathological slides for these patients were reviewed blind by one of us who had histopathological training (PS) and coded according to the International Classification of Diseases for Oncology (ICD-0) (World Health Organization. 1976). The ICD-O codes were then aggregated into groups. Confirmation of the diagnosis was available by histology for 696 (70.9%) of the subjects whose final diagnosis was lung cancer. and by cytology for a further 140 (14.3%): no microscopic evidence was available for 146 (14.9%). In this last group. the clinical outcomes and the proportion who were life-long non-smokers provide evidence that the majority had been correctly identified as having lung cancer. By the end of the investigation. 73% of the subjects without microscopic confirmation had died and 91% of these patients were certified as having died of lung cancer against 76% and 97%. respectively (indirectly standardized for age). of those with microscopic confirmation. The proportions of life-long non-smokers in the two groups were 0.0% of those without and 0.5% of those with microscopic confirmation in men and 8.3% and 7.1%. respectively, in women.
Of the remaining subjects originally suspected to have lung cancer. the final diagnosis was a smoking related disease (see Table 2) in 113. and these were excluded from the study. while the other 317 were transferred to the hospital control group. The hospital case notes of subjects selected as hospital controls were also reviewed. For 36 patients. the final diagnosis was a smoking related disease and they were excluded from the study. The final diagnoses of the remaining 1382 patients and the 317 transferred from the suspected lung cancer group are listed in broad categories in Table 2.
In addition to the hospital controls. a further population-based group of controls was selected, frequency-matched to the subjects with suspected lung cancer by age. sex and county of residence. In Comwall, these controls were randomly selected from the lists of the Family Health Services Authority (FHSA) (forrnerly Family Practitioner Committee). and permission for interview was sought from each patient's general practitioner before contact was made. Population controls in Devon were initially selected in the same way but. during the course of the study, permission to use FHSA lists was withdrawn, and the remaining controls were randomly selected using electoral rolls. A total of 5223 individuals were selected as population controls. including 2444 from FHSA lists and 2779 from electoral rolls (Table 1). Of these. 1486 (28.5%) were interviewed: 1119 (21.4%) did not satisfy the residence requirements; 304 (5.8%) did not wish to take part: 160 (3.1%) aUsually because subject or a family member was very ill. "As judged by klcal research assistants. cNon-white ethnic group. "Moved outside study area (50). subsequently found to be outside age range (31). already in study (10), non-white ethnic group (2). eAge/sex group already full (897), outside age range (622). already in study (9). tuberculosis, road traffic accidents or bums attributed to acohol consumption of subject, and cancers of lip, mouth, pharynx, larynx, oesophagus, pancreas, kidney, bladder, cervix and unknown primary site. These subjects were exciuded from the study. "These subjects were transferred to the hospita control group.
interest. pattems in the two groups were very similar for each sex (Tables 3 and 4). The two control groups were therefore combined for examination of radon-related risk. The final number of subjects included in the analysis was 4167. comprising 982 subjects with lung cancer and 3185 controls.
Information on residential radon concentrations For all subjects who were interviewed. full residential histonres covering the previous 35 years were obtained. For each dwelling at which the subject had been a resident for more than a year. information was collected on the precise address. the period it was occupied by the subject and the following housing characteristics.
which were noted by Gunby et al (1993) as having the greatest bearing on residential radon levels in the UK: tWpe of building.
floor levels of living area and bedroom. and presence of doubleglazing in living area and bedroom. Attempts were made to measure the radon concentration in every address in Devon or Cornwall at which the subjects had lived during the 30-year period of interest. Two detectors were installed for a period of 6 months.  aThose who had never smoked as much as one cigarette per day for as long as a year or smoked cigars or a pipe regularly for as long as a year. and who had smoked less in total. than 500 cigarettes. 100 cigars or 20 oz of pipe tobacco. nEx-smokers are those who had stopped smoking at the onset of their illness (lung cancers and hospital controls) or on the date of their interview (population controls). :Current pipe or cigar smokers who did not smoke cigarettes and occasional smokers. i.e. those who were not lifelong non-smokers but had never smoked as much as one cigarette per day or cigars/pipe for as long as a year. one in the living area and one in a bedroom: for the studv subjects' current address. this w-as the subject's ow-n bedroom. w hile for past addresses it w as a bedroom that w as currently in use. For subjects who had lixed in their current home for more than 5 s-ears. radon detectors were provided by research assistants. who v isited the home to check that the detectors had been correctly placed and subsequentlx retrieved the detectors. Current residents of prex ious homes of study subjects in Dex-on and Cornw all w ere contacted by the National Radiolooical Protection Board (NRPB) by post and invited to participate in the study. Radon detectors w-ere sent bv post to those w ho agreed. w ith detailed instructions on installation.
Non-responders w ere sent a second and. if necessarx. a third letter.
The postal approach w-as successful for approximatelI 50%7e of the current residents of previous homes. When it w-as not successful. personal visits were made by research assistants. The small passixve radon detectors were manufactured by the NRPB. The production and processinc methods conformed to the criteria of a formal validation scheme (Cliff et al. 1991). the accuracy of the measurements w as tested ex erv 6 months. and stringent quality assurance procedures w ere applied ( Hardcastle et al. 1996). Each detector consists of a small chamber contairnng a sensitixve plastic material. Radon diffuses into the chamber and decays through its chain of decay products. Some of the alpha particles emitted damage the sensitive plastic element. and this damage is revealed later by etching the plastic in a solution of sodium hy-droxide. The etched tracks are counted with an automatic image analyser. and their number is proportional to the exposure of the detector to radon. The detectors remained in place for 6 months before return to the NRPB for analy sis. Precautions wxere taken to prexent the detectors recording appreciable exposure to radon in the period before and after monitoring in the target address. Before despatch from the NRPB. the detectors were stored in nitrogen. w-hich prox-ides a low-radon environment. Detectors wxere transported betxween the NRPB and Devon and Cornw-all by post. A typical transit time wxould be 3 days. and a large proportion of this time wxould be spent essentially in outdoor air. Outdoor radon lexels in the UK are low (lWrixon et al. 1988. Appendix J). and the small percentage of time spent indoors. in places such as sorting offices. is unlikelv to have made a material contribution to the oxverall radon exposure recorded by the detector.
The research assistants wxere instructed to store detectors for a maximum of 6 x eeks before placement in homes and to keep them in a lox-radon enxironment. such as a well-xventilated upstairs room or a Xehicle. A control detector wxas supplied w ith each batch of detectors. and it remained in the local storage place for the total period during w hich the detectors in that batch w ere out in the field. The results from these control detectors provided reassurance that   aHousebots, caravans, etc. Radon concentratin assumed equal to outdoor level of 4 Bq m-3. 'Subject occupied several dwellings for short periods over a total penod of more than a year. Each such period is counted as one address' in the table.  (Table 5). and measurements were obtained in 9448 (72.5%) of these addresses.
An additional 89 were houseboats. caravans. etc.. where the radon concentration would be low and %vas assumed to be equal to the outdoor level for the UK. estimated by Wrixon et al (1988. Appendix J) to be 4 Bq m-. A further 20 addresses corresponded to periods at sea. where the radon concentration is very low (INSCEAR. 1982) and was assumed to be zero. When attention was restricted to the 10-year period ending 5 years before interview.
British Joumal of Cancer (1998) 78 (3), 394-408 Table 7 Distribution of time-weighted average residential radon concentrations experienced by study subjects during the 30year penod ending 5 years before interview. based on measured values and estimates for which no measurement could be obtained. The method of estimation used was that for analyses based on observed values (see section on Information on residenbal radon concentrabons). there w as a total of 5885 addresses. and measurements w-ere obtained for 5303 (90.1 ';%). The proportions of addresses at %vhich radon measurements were obtained were xersimilar for subjects with lung cancer and for controls (see Table 5 . For 2121 subjects (5 1%7 ). measurements w-ere obtained for all the relexvant addresses.

Number of
In order to take account of the possibility that radon remedial measures had been taken at some addresses. current residents of studs subjects' past addresses w ere asked w hether any such measures had been taken. and the NRPB's database of approxi-matelI 100 000 radon measurements in Devon and Cornwall (NRPB. 1996) was searched for evidence of previous measurements at both current and past addresses of studv subjects. For nine addresses. a previous measurement w as found that was aboxve the 200 Bq m-action level (NRPB. 1990) and that was more than tw-ice the more recent measurement. For these nine. it w as assumed that remedial measures had been implemented 3 months after the earlier measurement and that the earlier value applied up until then.
For each address at which the radon had been measured. the average annual radon concentration w-as estimated assuming that 45% of indoor time wxas spent in the liming area and 55% in the bedroom (Wrixon et al. 1988. Appendix M). and using the seasonal correction factors derived bx Pinel et al (1995). Each subject's time-weighted average indoor radon concentration was calculated dunng the 30-year penrod of interest. wxhen the %veights were equal to the number of years spent at each address. The w eighted ax erage was based on measured values. when these w ere axailable. and on estimated X alues for addresses wxhen no measurement could be made. For addresses at which no measurement could be made. estimates of the ax eraae annual radon concentration were obtained bv different methods for addresses in Devon and Cornwall and for addresses elsewxhere. Different estimates were also used according to whether analy ses were based on obserxved radon concentrations ignoring the uncertainties in their assessment. or whether these uncertainties were taken into account.
For analy-ses based on obserxed radon xalues. i.e. ignoring uncertainties in the assessment of radon concentrations. estimation for addresses in Dexon or Cornmall was carried out as follows. First. the approximately 100 000 radon measurements in Devon and Cornm-all in the NRPB database w ere examined. and the geometric mean radon concentration for addresses in each 5-km grid square was obtained. The areas of Dexon and Cornwall corresponding to the grid squares were then classified into six geog,raphical groups according to xxhether the mean x-as <20. 20-32. 33-54. 55-89. 4.5-4.9 or > 5.0 log0 (Bq m-)]. resulting in geographical groups in which radon concentrations were likelv to be similar. All the addresses relating to studv subjects for xxhom no measurement could be made wxere then classified into the same six geographical groups according to the grid square in wxhich they wxere situated. Measurements obtained specifically for the study tended to be lowxer than those in the NRPB database. In consequence. the missingy values in each group were estimated by considering only measurements made specifically for the study. Estimates x ere based on measurements for control subjects only. as recommended by Weinberg et al (1996). and because the control subjects were a close approximation to the population from which the studv subjects were drawn. The estimates were calculated as the arithmetic mean in each geographical group (Weinberg et al. 1996). This method of estimation was chosen after evaluating the performance of sexeral different methods in predicting, the measurements that had been made for the study. Use of a largyer number of geographical groups or adjustment for housing charactenrstics by fitting regression models did not improve prediction performance appreciably. For study subjects who occupied several addresses in Devon or Cornxxall for short periods covering a total period of more than a year. the radon concentration for this period wxas estimated by the arithmetic mean of all measurements made throuahout Dex on and Cornwall specifically for control subjects. Values for UK addresses not in Dex-on or Cornwall were estimated from the NRPB national database. xhich includes oxer 150 000 measurements in areas other than Dexon and Cornxxall. W'hen the full postcode of the address was axvailable. these wxere based on the arithmetic mean of the nearest 20 results. and for other addresses a county or other appropriate mean wxas used. For addresses not in the UK. the world axerace of 40 Bq m-xxas used (UNSCEAR. 1993 The methods used for estimating missing radon concentrations in the anaix ses that took uncertainties into account are described in the Appendix.

Information on other factors
During the interview. subjects %vere questioned about smoking habits. occupational histor-. carotene consumption'. exposure to radiotherapy and county of birth. Women who w-ere married or xx idoxxed w-ere also asked about their husband's current or last occupation. Smokinc, habits w-ere classified according to consumption at the onset of the illness that brought the subject to hospital (lung cancers and hospital controls) or current consumption (population controls). For current cigarette smokers w-ho also smoked a pipe. tobacco consumption >-as converted to cigrarette equivalents bv assuming that 1 oz of pipe tobacco per week was equix-alent w-ith reaard to the risk of lung cancer to two cigarettes per day (Doll and Peto. 1976) and >-vas added to their cigarette consumption. Few current cigarette smokers also smoked cigyars or cigarillos. and no allowxance w-as made for these. Each job held by a study subject for more than a xear w-as classified accordina to w-hether or not it >-as likely to incur a specific risk of lung cancer. Those considered to incur a possible risk wxere: work undergyround in a tin or other mine in Devon and Cornwall (Hodgson and Jones. 1990). and jobs x-ith asbestos exposure. includincu dockyard wxork (Harnes. 1968: Acheson andGardner. 1979). Subjects wxere classified into three social class grroupings: I & II. III non-manual or manual. and IV & V (Office of Population Censuses and Sun-ey s. 1980) based on their current or last job or. for married women. their husband's current or last job. A list of all foods consumed in the UK that are appreciable sources of carotene w as compiled by a nutritionist. and. during the interview. subjects w-ere questioned as to the frequency w-ith w hich each w as consumed. For each subject an estimate of daily carotene consumption was calculated using standard portion sizes obtained from survey and other data (Ministry of Agriculture. Fisheries and Food. 1993) and tables of the composition of foods. in w-hich carotene was expressed in the form of beta-carotene equivalents (Holland et al. 1991). Subjects w-ere then dix ided into quartiles accordinr to their estimated carotene consumption. For patients vA-ho reported that they had received radiotherapy. their statements w-ere assessed by a radiotherapist to determine whether or not they were likely to hax-e caused a dose to the lung of more than 1 gray. For all of these factors. the findings w ere in the direction expected. and the detailed results "-ill be reported elsew-here.
Statistical methods Initial analx ses wvere based on obserned radon concentrations. and they ignored uncertainties in the assessment of radon exposure. In these analy-ses. associations between lung cancer risk and obserx ed time-w-eiahted radon concentrations w-ere studied usincg the Stata statistical package (StataCorp. 1997). Relative risks are maximum-likelihood values based on unconditional logristic regression w ith adjustment for age (5-year interx als). sex. smoking status (sexen categories. as in Table 3). county of residence and social class (three categories . Estimates of excess relatixe risk (ERR) per 100 Bq mare based on linear logristic regressions and At the time the studv 'sas: started it wias widels believed that carotene A-as likels to be the agent primarily responsible for the proph% lactic salue of green and velloss s-egetables This no longer seems lihke to be true Our findings. how-ever. show that it did ser e as a good index ot w-hatever benefits the consumption ot green and vellow-vegetables might produce to be published use estimated radon exposure in indix idual subjects considered as a continuous xariable. Inclusion of additional terms in the regression models. such as interactions between the terms listed above. or additional factors. such as place of current residence (individual countv district. or urban/rural status of countx district). carotene consumption (four categories). w ork in a job incumnr , a potential lung cancer risk. exposure to radiotherapy or birth in Dexon or Cornwxall. altered the estimate of ERR at 100 Bq m-by 6'k at most. Analy ses based on linear. as opposed to linear logistic. models of radon risk or conditional on the adjustment variables also gaxe similar answers. Sianificance lexvels are based on the likelihood ratio test. and confidence intervals are based on standard errors. For estimates of relatixe risks w-ithin categories of radon concentration. cutpoints w-ere chosen on the basis of the distribution of time-weigahted average radon concentrations for control subjects and without prior knowledge of the relatixve risks. Heterogeneity tests for subject and tumour characteristics were based on the likelihood ratio. Analvses that took into account uncertainties in the assessment of radon exposure w ere carried out usina the method dexveloped by Reex es et al ( 1998). Further details are gixen in the Appendix.

RESULTS
The arithmetic mean of the seasonally adjusted radon lexels in the 9448 addresses at w-hich measurements w ere obtained w as 57 Bq in'. The -alues were approximately log-normally distributed and the quartiles of the distribution occurred at 15. 28 and 58 Bq m->. while the hiahest measured concentration was 3549 Bq min. The distribution of the measurements in subjects with lung, cancer and controls was xerx similar (Table 6).
When the 30-vear period of interest was considered for each subject. measurements were available for an axerage of 25.2 and 25.5 -ears for subjects w ith lung cancer and controls. respectixely. corresponding, to 84.0%7c and 85.1%',7 of the period of interest (see Table 5). After substitution of estimates for addresses at which no measurement could be obtained. the time-x eigahted residential radon concentration experienced by subjects with lung, cancer during the 30-y ear period of interest had arithmetic mean 58 Bq m-. while thexalue for controls AasXery close at 55 Bq m-( There is known to be uncertaintv in the measurement of residential radon concentrations. as illustrated by the fact that. when the radon concentration in a house is measured on tu-o separate occasions. the values obtained on the two occasions differ (Lomas and Green. 1994   Adjusted radon concenn (Bq m3) Figure 1 Relative risk of lung cancer according to resident radon concentratio ac §usted for age, sex, smokig status, county of residence and socal cl. In the top panel, relative risks and 95% confidec intervals are shown by mean time-weighted average concentaton durig the 30-year period ending 5 years before interview for iKdivals with observed values in categories <25, 25-49, 50-9, 100 -199, 200-399 and 400+ Bq rn-3, and the fitted regression kne is based on observed values for idvidual stbjects. In the bottom panel, the mean values and fitted reg on line have been acjsted for unetainties in the assessment of radon concentration. In Fe top panel the fted regression line correspondS to an ERR per 100 Bq nf-3 of 0.08, whie for the bottom panel the value is 0.1 2 . As a consequence of this uncertainty, observed radon concentrations at the upper end of the distuibuton will tend to be considerably higher than their true values, while observed radon concentrations at the lower end of the distribution will tend to be slightly lower than their tnie values. and ERRs based on observed radon concentrations will underestimate any risk ). When the methods described in the Appendix were used to adjust for uncertainties in the assessment of radon exposure. the mean values of the te time-weighted average radon concentration for individuals whose observed values lay in the categories 200-399 and 400+ Bq m-3 were estimated to be 202 and 371 Bq m3. respectively. considerably lower than their observed values of 259 and 662 Bq m-3 (see Table 8), and the estimated ERR per 100 Bq m 3after adjusting for uncertainty was 0.12 (95% CI -0.05, 0.33) (see Figure 1). This estimate is larger. and also has a wider confidence interval. than the estimate based on observed radon concentrations. i.e. without allowing for uncertainty.
Studies of patterns of lung cancer in underground miners exposed to high levels of radon have suggested that exposure during the previous 5-15 years carries a greater risk per unit exposure than that received in the more distant past (see. for example, Tomasek et al, 1994). When the present analysis was repeated considering radon concentrations during the 30-year period ending 5 years before the interview, but weighting the exposure received during periods 5-14, 15-24 and 25-34 years previously in proportions 1.0:0.75:0.50, as suggested by recent analyses of data from the studies of miners , the ERR based on the observed radon concentrations was 0.07 (95% CI -0.03. 0.19). while the corresponding estimate after adjusting for uncertainties in the assessment of radon exposure was 0.11 (95% CI -0.06, 0.31: see Table 8). Both these estimates are very similar to the values obtained when all time periods were weighted equally.
When the analysis was limited to the 2121 subjects (49% of those with lung cancer and 51% of the controls) for whom measurements were available for all 30 years, the ERR per 100 Bq mI3 based on observed radon concentations was 0.14 (95% CI 0.01. 0.29), somewhat greater than the value obtaied when all subjects were included in the analysis, although the difference between the two groups was Brish Joumal of Cancer (1998)    After adjusting for uncertainties. the estimated ERR per 100 Bq mincreased further to 0.24 (95%7 CI -0.01. 0.56). As for the analvsis when all subjects x-ere included, additional peniod-weighting to give greater weight to exposure in the more recent past changed the estimate of risk bv 'ers fittle (Table 9). To see whether there "as an' e'idence that the effect of radon differed according to the characteristics of the tumour. the ERRs per 100 Bq m-based on observed radon concentrations were estimated separately by histological type and by site of tumour (Table 10).
Although there was some variation in the ERRs. with higher values occurng for small-cell tumours than for other histological types. and for tumours bevond the main bronchi. including the periphery of the lung. rather than in the main bronchi. the observed Xvariations were not greater than gould be expected bv chance (histological type x'4=4.9. P= 0.29: site of tumour`= 2.2. P= 0.13). When tumours for which no microscopic evidence w as available were excluded, the ERR per 100 Bq mfor the remaining tumours was 0.09 (95%c CI -0.02. 0.21). very similar to the value of 0.08 (95%7c CI -0.03. 0.20) obtained when thev were included. The youngest subject with lung cancer was aged 30 years while 25 subjects (two with lung cancer and 23 controls) were aged 74 years when selected but were aged 75 years (24 subjects) or 76 years (one control) at interview. The ERR for current cigarette smokers is adjusted for amount smoked in categories <15.
15-24 and 25+ cigarettes per day. The ERR for ex-smokers is adjusted for time since quitting in categories <10 years and 1 0+ years. Other smokers are current pipe or cigar smokers who do not smoke cigarettes, and occasional smokers. Years working outdoors are full-time equivalent years in the 30-year peiod ending 5 years before interview. Years of part-time work are counted pro-rata. Symbols and other details are as in Table 10.
A similar analxsis was camred out to see w-hether there w-as anv ex idence that the effect of radon differed accordinc to any knou n characteristics of the subject (Table lI). Out of the four characteristics considered (sex. age. smoking status and years spent working outdoors). there vvas ev idence of heterogeneity only for sex (Xl = 3.7. P=O.05). with women having a lower ERR per 100 Bq m-than men. Among the remaining, categories. ERRs were hiahest for subjects aged less than 55 years. for ex-smokers and for those uwho had worked out of doors for more than 20 years. but there wxas no ex idence of heterogeneity for any of these characteristics (see Table 11).

DISCUSSION
This report presents the results of a large. population-based study specifically designed to examine the relationship between residential radon concentration and luncg cancer risk. The study wvas carried out in the part of the UK where the highest concentrations of residential radon occur and. in order to identifv a group of people likely to have been exposed to high average concentrations during the previous 35 vears. was restricted to long-term residents of the area To ensure that the subjects vvith lung cancer included in the study represented as closely as possible those occurrinc in the study population. only incident cases were included and. to minimize anx biases in the information on smoking and factors other than residential radon that determine lung cancer risk. all study subjects w ere personally interviewxed bv trained research assistants using standard questionnaires.
The exposure of interest is the residential radon concentration experienced by the study subjects in the past. This cannot be assessed directly. as it is possible only to measure current concentrations in both current and previous residences. There is. however. evidence from a studv of temporal variations in residential radon concentrations that. in the high radon areas of the UK. levels has-e not increased appreciably in general. at least durincg the decade before this study (Lomas and Green. 1994). In addition. efforts A-ere made in the present study to identify any dwellings occupied by study subjects where radon remedial measures A-ere likely to have been taken and to estimate the radon concentrations appropriately.
Radon concentrations found in the present study A-ere loser than those found in the NRPBWs large database of approximately 100 000 measurements x-ithin Devon and Cornwall. This is chiefly accounted for bv a tendency for the dwellings included in the NRPB database to be preferentially located in the highest radon areas w ithin Devon and Cornmwall. An earlier surney by the NRPB of radon concentrations in UK residential addresses selected to be representative of the w hole country from files maintained by the Post Office included 37 measurements for Devon and 16 for Comwxall and oave arithmetic means of 72 (95%c CI 19. 125) and 114 (95% CI 67. 162) for the two counties respectively. The corresponding values in the present study for addresses occupied by control subjects were 42 (95% CI 40. 44) for Devon and 108 (95% CI 100. 116) for Cornwall. based on 5706 and 1538 measurements respectively. The values observed in the present study are therefore consistent with those observed in the NRPB representative survey. Although strenuous attempts were made to measure the radon concentrations at the addresses of all study subjects during the 30year period of interest, there were inevitably gaps in the measurement histories. corresponding to 15% of the period of interest. and estimates for these addresses were therefore constructed using a validated methodology (Weinberg et al. 1996) which took into account the location of the address. When the analysis was limited to individuals for whom it was possible to obtain radon measurements for the entire 30-year period of interest. the estimated excess relative risks were larger than for the entire group. This may be a chance finding or it may be a reflection of the fact that more information is available regarding the exposure histories in this subgroup.
Radon concentrations that have been estimated rather than measured are inevitably subject to uncertainty. Measured radon concentrations are. however, also subject to uncertainty in the sense that. when a dwelling is measured twice. values that differ appreciably will usually arise, even when high-quality long-term measurements are carried out and appropriate seasonal corrections applied. A study carried out by the same laboratory as that in the present study. and using similar techniques. indicated a coefficient of variation for repeated measurements in the same house of around 50% (Lomas and Green. 1994). Unless taken into account. this measurement variability will distort the results, in that the highest observed radon concentrations will tend to be overestimates of their true values, and the lowest will tend to be underestimates. so that regression coefficients based on the observed radon concentrations will tend to underestimate the strength of any relationship between true radon concentration and risk of lung cancer, with the extent of the attenuation depending on the size of the measurement variability. Special methodology has therefore been developed that takes appropriate account of the uncertainties due to both measurement and estimation variability in the assessment of time-weighted average radon concentrations ). In the present study. the relationship between radon concentration and risk of lung cancer has been estimated twice. first in the standard way based on the observed radon concentrations and then after taking the uncertainties into account. The effect of taking account of the uncertainties was to increase both the magnitude of the estimated radon-related risk and the size of the associated confidence interval. Estimates based on observed radon concentrations are appropriate for comparison with the results of other studies of residential radon in which similar uncertainties are likely to be present but have not been taken into account; while estimates in which the uncertainties have been taken into account are more appropriate for comparison with risk estimates derived in different ways and when considering the amount of lung cancer likely to be caused by residential radon.
The risk of lung cancer is determined by other factors as well as residential radon concentration. In the present analysis. logistic regression has been used to adjust for the effects of these factors, the most important of which is smoking status. In order to be sure that no appreciable residual confounding with smoking status remains. seven categories of smoking status have been used in the adjustment. with life-long non-smokers and ex-smokers of durations <10 and 10+ years in separate categories. and three separate categories for current smokers of cigarettes (see Table 3). Previous studies have demonstrated that very little residual confounding remains after this degree of stratification for cigarette consumption (Breslow and Day, 1980). Errors in the assessment of smoking status are also likely to be present. It would be possible from the theoretical point of view to take them into account. but there are few data available with which to quantify such errors. In any case. as there is little confounding between radon and smoking status in the present study, adjustment for errors in the assessment of smoking status would have little effect on the estimated risk from radon. At the present time. nine case-control studies of indoor radon and lung cancer have been carried out that have each included at least 200 subjects with lung cancer and measured at least one residence for most subjects. These studies have been carried out in Canada. China. Finland. Sweden. the USA and westem Genrany (Blot et al. 1990;Schoenberg et al. 1990: Pershagen et al. 1992Alavanja et al. 1994: Letoumeau et al. 1994: Pershagen et al. 1994: Auvinen et al. 1996Ruosteenoja et al, 1996: Wichmann et al, 1997. For eight of these studies, the published relative risks after adjusting for confounding variables have been combined using weighted linear regression to give an estimated excess relative risk of 0.09 (95% CI 0.0. 0.2) per 100 Bq m-based on observed radon concentrations (Lubin and Boice. 1997). while for the study in western Germany (Wichmann et al. 1997) the estimated excess relative risk per 100 Bq min radon-prone areas based on observed radon concentrations is 0.13 (95% CI -0.12. 0.46). The totality of the evidence from other studies of residential radon and lung cancer therefore suggests an excess relative risk of around 0.1 per 100 Bq m-3. based on observed radon concentrations. Thus. the estimated excess relative risk based on observed radon concentrations in the present study of 0.08 per 100 Bq m-3 (95% CI -0.03. 0.20) is in close accordance with the findings from other studies. Although the 95% confidence interval for the excess relative risk in the present study just includes zero. the combined evidence suggests that a zero effect would be an inappropriate interpretation of the study results.
The impact of measurement variability on the excess relative risk has been assessed for only one of the nine previous studies (Lagarde et al. 1997). For that study it was also concluded that a coefficient of variation for repeated measurements in the same house was of the order of 50%. and that the excess relative risk of 0.10 per 100 Bq m-3 based on the observed concentrations should be corrected to about 0.15-0.20 per l00Bq m-3 when measurement variability was taken into account. This conclusion is very similar to that of the present study. in which accounting for uncertainties increased the estimated relative risk per 100 Bq m-3 from 0.08 to 0.12 (95% CI -0.05. 0.33).
The results of the ten studies of the effects of residential radon that are based on individual data together provide strong empirical evidence that the results of ecological regressions, whereby lung cancer rates in geographical areas are related to area-specific average residential radon level and in which a significant negative relationship between residential radon and lung cancer has often been observed. are highly misleading (see for example Piantadosi et al. 1988;Stidley and Samet. 1993: Cohen. 1995: Lubin. 1998).
The findings from the studies of residential radon that are based on individual data are also consistent with the findings from a pooled analysis of 11 studies of underground miners occupationally exposed to radon (Lubin et al, 1995a). For miners exposed to, at most. 50 working-level months. which would result in approximately the same bronchial dose as living in a house with a radon concentration of around 400 Bq m-3 for 30 years, an excess relative risk of 0.09 per 100 Bq m-3 has been estimated based on 468 deaths (Lubin and Boice, 1997;. When miners receiving higher exposures were also included in the analysis, a somewhat lower estimate was observed (Lubin et al, 1995a), corresponding to an excess relative risk of around 0.05 per 100 Bq m-3. For miners exposed to more than about 50 workinglevel months. an inverse dose-rate effect has been observed. whereby, for a fixed total exposure, greater risks are associated with exposures occurring at a low exposure rate and spread over a long duration than for exposures occurring at a high exposure rate with short duration (Darby and Doll. 1990: Lubin et al, 1995b. The inverse dose-rate is likely to occur for exposure levels at which lung epithelial cells are likely to be traversed by more than one alpha particle. Multiple alpha particle traversals are likely to occur in heavily exposed miners, but are rare within the range of radon concentrations usually experienced residentially (National Research Council, 1998). The risks of residential radon exposure are therefore unlikely to be affected by the inverse dose rate effect.
Analyses of mortality paterns in underground miners receiving substantially higher cumulative exposures than would normally occur residentially have demonstrated a tendency for exposures received in the previous 5-15 years to carry a greater risk than exposures received in the more distant past Moreover, such analyses have also shown that the relative risk associated with a given level of exposure tends to be higher in younger subjects and among nonsmokers compared with smokers (Roscoe et al, 1989;Tomasek et al. 1994;. In addition, there is considerable evidence that relative risks for small-cell cancers are higher than for other histological types of cancer (National Research Council, 1998). In the present study, there was no evidence to suggest that a higher risk of lung cancer is associated with exposure received in the more recent past (Table 8). However. there is little power of discrniination in a study such as this, in which a large proportion of subjects with the highest observed radon concentrations during the 30-year period of interest had lived at the same address for most of the period. Tests for heterogeneity between tumour and subject characteristics suggested a difference in risk only between men and women (Table 11). This was unexpected. and the result may be due to random variation: the chance of finding one out of six independent heterogeneity tests to be significant at a nominal level of 5% when in fact no heterogeneity is present is approximately one in four. Conversely, some of the variation observed between the other subgroups may represent real differences that are not statistically significant because of the limited power of the study. For example, a higher excess relative risk was seen for small-cell tumours than for other types of lung cancer and a higher relative risk was seen in those aged under 55 years than in older subjects. Both of these results would be predicted from the studies of miners receiving much higher exposures. In addition, the higher risks associated with tumours outside the main bronchus may be a result of radon progeny being more liable to be deposited peripherally than in the main airways; and the tendency for the excess relative risk to increase with increasing number of years spent working outdoors may be a reflection of the fact that the time-weighted average radon concentration has been more accurately estimated for dtese individuals.
Nothing of value can, however, be learnt from the interaction with smoking as the number of lung cancers in life-long non-smokers (26) was very small.
Although this study was large in size, with nearly 1000 cases of lung cancer and over 3000 controls, and was carried out in the area of the UK where the highest residential radon concentrations are found, as well as having a highly effective measurements programme covering on average 85% of the 30-year period of interest, it has only limited power to assess the risk associated with residential radon. Plans are in hand for formal pooled analyses of the data from both European and North American studies of lung cancer and residential radon. When these analyses are complete, a more precise estimate of the lung cancer risk should be available. together with clearer evidence on any variation in risk with subject and tumour characteristics.

CONCLUSION
In the present study, the estimated excess relative risk associated with a 100 Bq m-3 increase in residential radon concentration is 0.08 (95% confidence interval -0.03. 0.20) when uncertainties in the assessment of radon exposure are ignored and is 0.12 (95% confidence interval -0.05, 0.33) when these uncertainties are taken into account. Although the confidence intervals for these estimates just include zero, the estimates are similar in magnitude to those derived from other studies of residential radon in which data have been collected on individual subjects, and also from studies of underground miners occupationally exposed at low concentrations. The combined evidence therefore suggests that a zero effect would not be an appropriate interpretation of the study's results and that there is a risk of lung cancer associated with residential radon exposure of about the size that has been postulated on the basis of studies of miners occupationally exposed to radon.
Residential radon and lung cancer 407 where Y is the binary response variable. A(x) = e-/( I + e) is the logistic function. at and , are the intercept and slope. respectively.
of the relationship between the logarithm of disease odds and the true residential radon concentration, the index j runs over the addresses belonging to a particular subject. w; is the weight given to each address and usually represents the proportion of the 30year period lived at each address (1sK = 1). theare dummy vanables representing the different levels of the covariates (age. sex. smoking status. etc). 11k are their associated regression coefficients. and k = 0.588 is a multiplicative constant that arises when approximating the logistic by the probit function. X91, ... X9J, are the surrogate (i.e. observed) values of residential radon for the subject. and these and the remaining quantities in equation (1) differ according to whether or not a measurement is available for a particular address and are explained in the following two sections.
Addresses for which a measurement was available When the measured radon concentration at the jth address of a subject was available. X9j, was set equal to it in equation (1). The remaining quantities in equation (1) involve the measurement error variance and the mean and variance of the distributions from which the log radon measurements. i.e. the logX9j,. are drawn. It is the relationship between these parameters that determines the extent to which the measurement errors affect the estimated relationship. If the log radon measurements are drawn from a distribution with mean j and variance aJ2. and the variance of the logs of repeat measurements at the same address is a2. then from Reeves et al (1998).YL9J = (J-am2)/at. aU9j = am2 , , 9 j , and v*j is given by the expression V= {exp(q)}'I-, exp(a2 ,/2) As discussed in Reeves et al ( 1998). when covariates z are included in the regression the mean j and variance at' should be those of the conditional distribution of log radon measurements given the values of the covariates for the individual in question. In practice. there was appreciable correlation between only one of the covariates (county of current residence) and the log radon measurements. and so it is enough to estimate j and a1t' separately for the two values of this particular covariate. Thus, for an individual currently living in Devon. j and (at2 were taken as the mean and variance of all the log radon measurements (in whichever county they were measured) for all individuals also currently living in Devon, and similarly for Cornwall. In the present analysis. a12 and t took values 0.82 and 3.24. respectively. for subjects living in Devon and 1.10 and 4.08. respectively, for subjects living in Comwall. while a 2 was estimated extemally from a study in which repeat radon measurements had been made (Lomas and Green, 1994). It was found not to differ significantly between dwellings that had the same occupier for both measurements and dwellings with a different occupier. and took value 0.23. This indicates a coefficient of variation on the original scale of 51%.
Addresses for which no measurement was available For addresses for which no measurement of the radon concentration was available. X9;,in equation (1) was estimated using one of the following six methods: (1) For addresses in Devon or Cornwall for which there was sufficient information to classify the address into one of the six geographical groups described in the section Information on residential radon concentrations the radon concentration was estimated by the geometric mean of all the measurements taken specifically for control subjects in the same geographical group.
(2) For addresses in Devon and Cornwall with insufficient information to assign to a particular geographical group the radon concentration was estimated by the geometric mean of all the measurements taken for control subjects throughout Devon and Cornwall.
(3) For addresses such as houseboats or caravans for which the radon concentration could be assumed to be close to outdoor levels. it was taken to be equal to 4 Bq m-3. the typical outdoor concentration in the UK (Wrixon et al. 1988).
(4) For periods at sea, the radon concentration was assumed to be equal to zero (UNSCEAR. 1982).
(5) For addresses in the UK but not in Devon or Cornwall. the radon concentration was assumed to be equal to the estimated geomietric mean for the UK. namely 15 Bq m- (Wrixon et al. 1988. Appendix K).
(6) For addresses outside the UK the radon concentration was assumed to be equal to 30 Bq m-3. which was the best available estimate of the world geometric mean concentration (O'Riordan. 1993).
For addresses for which the radon concentration was estimated, y,,j, represents the uncertainty due to measurement error associated with the estimate. For concentrations estimated using method (1) above. from Reeves et al (1998). y, is given by the expression (Yf b_ where aY2', is the between geographical group variance of the logarithms of all the radon measurements relating to control subjects, + 'M) is the within group variance. which was found to differ between the groups and was therefore estimated separately for each group, 2m is the variance of the logarithms of repeat measurements. as for addresses for which a measurement was available, and nO is the number of measurements in the geographical group g. In fact. yLvj was found to be very close to unity for all six geographical groups. and it was therefore taken to be equal to unity for all radon concentrations estimated by method (1). For concentrations estimated by methods (2)-(6) above. y was also assumed to be unity. For addresses for which the radon concentration was estimated using method (1) above, from Reeves et al (1988). a2L; is given by ='~J a +a,' -Y'~{a', + (a2' +cs / (; s4j Y t b_ + Y s-e L94j 1( be-+(Y Ym)n Therefore. when .y2 = 1 and ne is large. (Y', =a' . For the six geographical groups. the estimated within group variances were 0.52. 0.61. 0.66, 0.92. 0.92 and 1.01. respectively. and, as a2 m was 0.23. the corresponding estimated values of a21, were 0.29. 0.38. 0.43. 0.69. 0.69 and 0.78 respectively. For addresses for which the radon concentration was estimated by method (2). the within group variance based on all the radon measurements taken for control subjects in Devon and Comwall was 0.98. leading to an estimated (a2n of 0.75.
For addresses for which the radon concentration was estimated using methods (3) and (4). a2 was assumed to be zero. For addresses with radon concentration estimated using method (5) ((Y2 + (2m) was estimated from the UK survey (Wrixon et al. 1988) and took value {log(2.17))2 = 0.60. Therefore. a2w, was estimated to be 0.37. For addresses estimated using method (6). a2,( mg was taken to be the estimate from the geographical group with geometric mean closest to the estimated world average concentration, namely 0.43.

Radon concentrations adjusted for uncertainties
The mean time-weighted average radon concentrations adjusted for uncertainties that are given in Tables 8 and 9 are average values of I w;v*XNtL'S for the subjects in question. For each subject, this quantity is the expected value of the true time-weighted average radon concentration given the observed value, conditional on current residence in either Devon or Cornwall.
Model fittng Maximum likelihood estimates for the parameters in equation (1) were derived by iterative application of the logistic regression command in the Stata statistical package (Statacorp, 1997). For the first step of the iteration, the denominator of the logistic function was assumed to be equal to unity, and in subsequent iterations the numerator was adjusted for the current value of the denominator. In practice, the denominator remained very close to unity for all the models fitted, and therefore confidence intervals for the parameters could be based on the standard errors computed by Stata.

Sensitiv analysis
Additional analyses were carried out to determine the sensitivity of the results to some of the assumptions made above. Firstly. for addresses for which no measurement was available and when Xsj, was estimated using methods (5) or (6) above, the analysis shown in the top line of Table 8 was repeated first doubling and then halving a2s.. The former increased the estimated ERR of 0.12 per 100 Bq m-3 to 0.13, while the latter did not change it. Secondly, y' was first increased and then reduced by 20% for all addresses for which there was no measurement available, regardless of the method of estimation of X 1j,. The latter increased the estimated ERR from 0.12 to 0.13 per 100 Bq m-3, while the former did not change it. Thirdly. the assumed value of a2m was first increased and then reduced by 20%. The former increased the estimated ERR from 0.12 to 0.13 per 100 Bq m-3, while the latter reduced it from 0.12 to 0.11. Fmally, the assumed value of &m was first doubled and then halved. In the former case, the estimate of 0.12 per 100 Bq m-3 was increased to 0.16, while in the latter case it was reduced to 0.10.
It was therefore concluded that the results did not depend strongly on the assumptions made in the uncertainty analysis, although, as would be expected, large increases or decreases in OY2m the variance of the logs of repeat measurements at the same address, increased or decreased the effect of accounting for uncertainties in the analysis.