All residents in the study regions aged 30 to 59 were eligible for the study; additional eligibility criteria, such as citizenship and proficiency in the local language were imposed in some study centres (Web Annex Table 1). The choice of age-range aimed to maximise the likelihood of exposure. Mobile phone use is a relatively new phenomenon: until the late 1990’s mobile phone use was mainly restricted to people in the age range most likely to use the phones for business purposes [7].

Eligible cases were all residents of the study region diagnosed during the study period with a confirmed first primary glioma, meningioma, or acoustic neurinoma. Eight centres (Australia; Canada—Montreal, Ottawa and Vancouver; Denmark; Israel; Italy; Sweden) also included malignant parotid gland tumours. Because benign parotid gland tumours may be treated in a very large number of institutions, most centres found it logistically difficult to ensure complete ascertainment, and only Canada–Ottawa, Israel and Sweden included them. They will not be discussed in this paper. The ICD codes for the eligible diagnoses are presented in Web Annex Table 2.

All diagnoses were either histologically confirmed or based on unequivocal diagnostic imaging. In Australia and Germany, only histologically confirmed tumours were included. In Denmark cases found to have had any previous cancer (excluding non-melanocytic skin cancer) were excluded.

Each centre established procedures for the rapid ascertainment of cases from participating diagnostic and treatment units, which was particularly important for glioma patients, whose health can deteriorate quickly. Every effort was made to maintain a close relationship with the units to ensure that cases were not missed and that the required authorisations were obtained from treating physicians when necessary. Close monitoring of case ascertainment was essential and all study centres, except Finland and Japan, used one or more secondary source (including medical archives, hospital discharge and billing files, and hospital or regional cancer registries) to improve ascertainment levels. Enrolment of cases through secondary sources often implied longer delays in case ascertainment and consequently lower participation.

Controls were randomly selected from the source population. The sampling frame depended on the local situation (Table 1). The study design called for controls to be individually- or frequency-matched to cases, with the number of controls varying according to the tumour type: 1 control per case for brain tumours; 2 for acoustic neurinoma; and 3 for parotid gland tumours. In Germany, two controls were selected for each brain tumour case. Controls were matched on year of birth (within 5-year categories), sex and study region.

Controls were individually matched to cases in Canada–Ottawa, Vancouver; France, Israel, Japan, New Zealand and UK-North. In the other centres, individual matching was conducted post hoc, with cases being assigned controls chosen to have been interviewed as close as possible in time to the case, from among those who fit the matching criteria.

All cases for whom physician authorisation for contact had been obtained and all controls were initially informed about the study and asked to participate. The procedures varied between centres (Web Annex Table 3), depending on the requirements of local Ethics Review Boards. In seven centres, the cases were initially approached by the treating physician or a nurse for consent to be included in the study. In other study centres approaches included: active case ascertainment by the study staff followed by physician authorisation to contact each case directly; blanket approval to contact all eligible cases; or a mix of the two. In all centres participants provided signed informed consent.

Whenever possible, consenting subjects were interviewed face-to-face by trained interviewers using a computer-assisted personal interview (CAPI) questionnaire. Only Finland used a paper version of the questionnaire. In exceptional cases, telephone interviews were conducted with difficult-to-reach subjects. If subjects became too tired or confused to complete the interview in one session a second appointment was arranged; a partner or other family member could assist in the interview. When the study subject had died or was too ill to participate, a proxy respondent was interviewed where this was possible and permitted by ethics committees. In Australia and New Zealand an abbreviated questionnaire was used for proxy interviews. Controls who refused to participate in the study were asked, whenever possible, to complete a short non-respondent questionnaire in all centres, except in Denmark and UK-South, in order to evaluate whether they differed from participating controls. A small number of cases in some centres also completed the non-respondent questionnaire. Detailed results of analyses of the non-respondent questionnaires will be published separately.

The study questionnaire covered demographic factors, mobile phone use (detailed below), use of other wireless communication devices including cordless DECT telephones, occupational exposures to EMF and other potential confounders or risk factors for the diseases of interest (including exposure to ionising radiation, smoking and the subject’s personal and familial medical history). Specific questions on exposure to loud noise and hearing loss were asked of acoustic neurinoma cases and their controls (and of all controls in centres using frequency matching).

The CAPI questionnaire included many checks: the sequence of questions was constrained with little opportunity to skip questions and automatic range and consistency checks were incorporated. After completion of the interviews, routine checks were performed on the data from all centres both locally and centrally. Inconsistencies and ambiguities were identified and resolved wherever possible.

The study used two main approaches to characterising exposure from mobile phones. The first depended only on the history of use derived from questionnaire responses and the second attempted to evaluate the amount of RF energy absorbed in different areas of the brain.

In both approaches, exposure was calculated up to a given reference date, which was set to the date of the diagnosis of the case in each matched set. Evaluation of RF energy absorption required the localisation of the tumour, which was defined crudely in terms of the side of the head, or lobe of the brain, or more precisely, from the exact location of the tumour on the 3-dimensional grid. Exposure for each control was estimated at the location of the tumour of his/her matched case.

The main analyses will be based on conditional logistic regression for matched sets. This simplifies the assignment of the reference date, laterality and “tumour location” for the controls, which are important when analysing the effects of an exposure that is very localised. In addition, for an exposure that increased rapidly during the course of the study, and considering that subjects’ recall of their past exposures may be influenced by their current and recent use patterns, the matching ensures that cases and controls have been interviewed relatively closely in time.

The proportion of low to high-grade glioma cases ascertained was quite consistent across most centres where this could be determined: 66% high-grade and 28% low-grade, with 6% unknown overall. Overall 1% of the meningiomas were malignant and 5% of unknown behaviour. This was consistent across all study centres (not shown).

Participation was calculated as the proportion of all eligible ascertained cases that were interviewed. The denominator includes cases whose physician denied authorisation to contact them: 5% of glioma, 2% of meningioma and acoustic neurinoma and 9% of malignant parotid gland tumour cases. This proportion was relatively small for most centres and in eight centres there were none at all.

There was little difference between centres in participation of glioma cases according to grade of tumour: 67% overall for cases with high-grade tumours and 71% for cases with low-grade tumours.

Overall participation amongst controls was 53% (Table 4) but showed large variation across centres, ranging from 35% to 74%. Eight of the study centres achieved control participation of 60% or higher. The major reasons for non-participation were refusal (64% of non-participants) and inability to contact (27%).

The overall proportion of proxy respondents for glioma cases was 13%. This varied considerably across centres, from 2% to over 40% (Table 6). A small proportion of interviews was conducted with the study subject accompanied by another person. The proportion of proxy interviews was less than 2% for meningioma and even less for acoustic neurinoma and parotid gland cases. As would be expected, there were virtually no proxy respondents amongst the controls except in New Zealand where proxy interviews were conducted for the controls matched to cases who could not be interviewed themselves.

Overall, 60% of the face-to-face interviews with glioma cases were conducted at home, 33% in hospital, and 7% elsewhere (Web Annex Table 4). This varied greatly: in some centres, nearly all interviews were conducted at home; in others, nearly all were in hospital. The distribution of interview location was similar for meningioma and acoustic neurinoma cases (not shown). In contrast, 7% of control interviews took place in hospital (mainly in Finland and, to a lesser extent in Norway and Sweden, where study subjects were invited to treatment institutions for interview), 70% in the subject’s home and 22% elsewhere (Web Annex Table 4).

After an interview had been completed the interviewer recorded his or her impression of the reliability of information on a 5-point scale, overall and for each specific section. The percentage of subjects judged by the interviewer to be unresponsive or uncooperative overall was very low for both cases and controls (1.8 and 1.2% respectively), ranging by centre from 0 to 5.6% among glioma cases and up to 4.7% among controls. The percentages of cases and controls who were mobile phone users and were judged by the interviewer to have had little or no difficulty in remembering past phone use were 80, 86, 91 and 94%, respectively, among glioma, meningioma, acoustic neurinoma and parotid gland tumour cases, and 91% among controls (not shown).

Interviews were conducted by 230 different interviewers, the number ranging from 2 in Canada–Montreal to 39 in Denmark. About 35% of the interviewers conducted less than 20 interviews and 25% of the interviewers conducted fewer than 10, mostly with cases. While 84% of the subjects were interviewed by interviewers who had a balanced workload between cases and controls, the workload in three centres was particularly unbalanced (Web Annex Table 5).

The proportion of subjects in the lowest educational level was somewhat higher for cases than controls in several centres (Web Annex Table 6). There were little differences in marital status between cases and controls for all types of tumour. Women were less likely than men to be married (not shown).

The INTERPHONE study is no exception to the apparently inexorable decline in participation rates amongst controls selected from the general population for epidemiological studies [48]. The source population is younger than in many other cancer studies and at an age when response rates tend to be lower. The youngest men proved particularly difficult to recruit. Another factor influencing the participation of controls is the difficulty of finding a sampling frame with sufficiently accurate, up-to-date and complete information, which resulted in large numbers of subjects who could not be traced or could not be contacted using the methods authorised by ethics committees.

The possibility that participation among controls might be selective with respect to phone use was of concern, given the low participation rate (53%). Mobile phone users could be over-represented among non-participating subjects as they may be more difficult to trace (fewer with listed telephone numbers for land-lines) or too busy to participate; this could lead to overestimation of the true OR. Alternatively, we have some evidence from the non-respondent questionnaires—which were completed by 57% of controls who refused to participate and may not be representative of all non-participants in the study—that non mobile phone users may be more likely to refuse to participate, perhaps in the mistaken belief that non-users are of no interest to the study [49]. Such a bias could artificially increase the proportion of users among interviewed controls and reduce the likelihood of finding an effect should it exist. As ordained by local ethics committees, the presentation of the study differed somewhat by centre. We estimate that 41% of all controls were recruited in centres that used an approach in letters and information material that explicitly indicated that the primary objective concerned mobile phones, 46% were recruited in centres that mentioned mobile phones, without highlighting them, and 13% in the three centres that made no explicit mention of mobile phones. Thus there is a potential for differential participation between users and non-users and between users by level of use. The impact of a possible selection bias with respect to controls has been evaluated in a simulation study and shown to be potentially important [50]. Thus, it will have to be taken into account in interpreting the INTERPHONE results. In particular, analyses will be conducted by level of participation and by mode of presentation of the study.

Selection bias with regard to severity of illness may also arise in cases, particularly glioma cases. For example, as would be expected from the poor prognosis and strong impact on communication skills, the participation of glioma cases (65%) was less than that of patients with meningioma, acoustic neurinoma and malignant parotid gland tumours (78, 82 and 75% respectively). If RF exposure were related to the severity and prognosis of cancer, differential participation due to severe illness, early death or cognitive impairment could lead to bias. Despite considerable effort to ensure rapid ascertainment to avoid these difficulties, late ascertainment of a proportion of cases because of logistic reasons and, in some countries difficulties in complying with the requirements of ethics committees, resulted in lower participation than expected. Comparison of response by grade of tumour for gliomas, however, shows no major difference across study centres with different delays between diagnosis and interview. The possibility of severity or survival bias will nevertheless have to be considered when interpreting results.

Self-reports of historical mobile phone use may be prone to substantial error. If such errors occur randomly, they usually bias risk estimates towards the null (no effect). They also increase the uncertainty of risk estimates, making it more likely that real associations are not detected. Results of short-term validation studies with volunteers indicate that recall of phone use is subject to moderate systematic error, but substantial random error: a substantial proportion of subjects markedly over- or under-estimated their mobile phone use [42].

Cases may spend time after the diagnosis of their tumour trying to understand why they have developed this disease, which might introduce a differential bias (sometimes referred to as rumination bias) in comparison with controls in recall of the amount and side of phone use. In addition, some of the patients with glioma might have recalled their phone use less accurately because of severe illness or cognitive impairment. Information about possible differences in recall of amount of use between cases and controls was obtained from retrospective validation studies. Analyses are underway and will be taken into account in the interpretation of results. The results of these analyses will be published separately.

Analyses of the INTERPHONE data will include various approaches to examining the potential for recall bias related to mobile phone use.

Possible confounding effects of region, age and sex will be taken into account systematically by the matching of cases and controls. Indeed, the mean ages of cases and controls are very close (glioma cases and controls 47.3 years; meningioma cases and controls 49.3 years; acoustic neurinoma 47.5 and 47.7 years respectively for cases and controls; parotid gland tumour cases and controls 46.3 years).

Because SES may well be correlated with mobile phone usage and with brain cancer risk [51], our primary indicator of SES, education status, will be included as a confounder in the analyses.

A priori, we do not have strong grounds for believing that other possible causes of the tumours studied, such as family history of brain tumour, past medical radiation exposure, smoking history and occupations in jobs with potential for ionising and non-ionising radiation exposure, would be related to mobile phone use. Nonetheless, the possibility of confounding by these factors will be examined empirically and they will be included in risk models where their inclusion results in a change in the ORs for the mobile phone use variables of 10% or more [52].

The fact that controls tended to be interviewed later than cases may also be a source of bias: because of the dramatic increase of mobile phone use during the study period, subjects interviewed later are more likely to have been mobile phone users. This will be handled by the matching, by truncating the exposure history of controls at the reference date, and, where appropriate, by adjustment of analyses for dates of interview and by analyses restricted to cases and matched controls interviewed close in time.

Symptoms of the disease could affect the use of mobile phones. Thus the year before diagnosis will not be included in the exposure period, and control exposure time will be truncated accordingly. For some tumours, however, early symptoms may appear much earlier. This will need to be considered carefully for acoustic neurinoma patients, as the early symptoms comprise hearing loss, tinnitus or buzzing sounds in the ear, which could lead patients to use mobile phones less frequently or to change the side of use.

There is also concern about the possibility of bias due to differences between cases and controls and between short term and long-term users in the extent of missing information and imputations for missing values were therefore conducted. Sensitivity analyses will be conducted excluding subjects with imputations, telephone and proxy interviews, and study subjects who were judged by the interviewer to be uncooperative or to remember their phone use poorly.

Differential assignment of cases and controls to interviewers in some centres and the use of many interviewers who conducted few interviews may be another potential source of bias. Sensitivity analyses will be conducted excluding interviewers with fewer than 20 interviews and those who interviewed nearly exclusively cases or controls.

Finally, despite the fact that INTERPHONE was jointly planned and based on a common core protocol, there was some heterogeneity in the methods used. Sensitivity analyses will be conducted excluding, in turn, different study centers. Additional analyses of patterns of results across study centres are planned to evaluate, in particular, the impact of the way the study was presented (whether a study of mobile phone use or a more general study) and of the participation levels among cases and controls. These will be helpful in addressing the potential for bias that might affect the overall findings.