Study population

Details of the study design and population have been described previously.30^–32 In brief, subjects were recruited and enrolled in a hospital-based case-control study of brain tumours between 1994 and 1998 in Boston, Massachusetts, USA (Brigham and Women's Hospital), Pittsburgh, Pennsylvania, USA (Western Pennsylvania Hospital), and Phoenix, Arizona, USA (St. Joseph's Hospital and Medical Center). All three hospitals were regional referral centres for brain tumours.

Eligible cases included patients diagnosed with a first primary glioma or other neuroepitheliomatous neoplasm (codes 9380-9473 and 9490-9506 of the International Classification of Diseases for Oncology, 2nd edition (ICD-O-2)), meningioma (ICD-O-2 codes 9530-9538) or acoustic neuroma (ICD-O-2 code 9560) within the 8 weeks during or preceding hospitalisation. For this analysis, only glioma (n = 489) and meningioma (n = 197) cases were included. Participation rates were 92% and 94% for glioma and meningioma cases, respectively. All participating case diagnoses were confirmed by microscopy. The majority of cases (80%) were enrolled and interviewed within 3 weeks after diagnosis.

Controls for this study (n = 799) were patients admitted to the same hospitals for non-malignant conditions, including injuries (25%), circulatory system disorders (22%), musculoskeletal disorders (22%), digestive disorders (12%) and other conditions (19%) and were frequency-matched to cases (1 : 1) by sex, age at interview (18–29, 30–39, 40–49, 50–59, 60–69, 70–79, 80–99 years), race/ethnicity (non-Hispanic white, Hispanic white, African-American, other), hospital (Boston, Pittsburgh, Phoenix) and proximity to the hospital (0–4, 5–14, 15–29, 30–49, 50+ miles). The participation rate among controls was 86%.

This study was reviewed and approved by the institutional review boards of each participating institution, and written informed consent was obtained from all study participants or their proxies.

Exposure assessment

Patients were interviewed in person by a trained research nurse who administered a computer-based questionnaire to collect information on lifetime occupational histories as well as other potential risk factors for brain tumours, including demographic variables (education, household income, marital status, religion) and history of radiotherapy to the head. For patients who were too ill to respond, functionally impaired or had died, interviews were conducted with a proxy, usually the spouse. Proxy interviews were conducted for 16% (n = 78) of glioma cases, 8% (n = 15) of meningioma cases and 3% (n = 23) of controls.

Information on occupational histories was collected for all jobs held for at least 6 months since the age of 16 years and included the name of the employer, dates of employment, job title, amount of time worked (full vs part time), type of business or service, tasks conducted, and materials and equipment used. For jobs with expected potential for exposure to specific chemicals, additional occupational information was collected through industry- and job-specific interview modules of questions.33 Sixty-four modules were developed to elicit detailed information on workplace determinants of exposure to a variety of agents, including chlorinated solvents. The information collected in the modules included the average frequency of various solvent-related tasks (converted to times per week), the average length of time it took to perform given solvent-related tasks (converted to hours per instance), sensory descriptions, dermal exposure, work practices, engineering controls and personal protective equipment use.33 Electronic versions of the occupational history data were sent to an expert industrial hygienist (PAS) within 2 weeks of the interview and reviewed for completeness and consistency. Up to 10 supplemental questions were developed to resolve missing or ambiguous information, and brief telephone re-interviews were conducted within 6 weeks of the initial interview. Thirty-six subjects did not complete the occupational histories. Of these, 29 (including nine glioma cases, five meningioma cases and 15 controls) never worked outside of the home and were considered unexposed. The remaining seven patients (five glioma cases and two controls) were excluded from our analyses.

Based on the occupational histories and the industrial hygiene literature, the industrial hygienist assessed levels of exposure to six chlorinated solvents distinguished by the numbers of carbon atoms (1 or 2) and covalently bound chlorine atoms (2, 3 or 4): dichloromethane (DCM) (synonymous with methylene chloride), trichloromethane (chloroform), tetrachloromethane (carbon tetrachloride), 1,1,1-TCA, TCE and tetrachloroethylene (perchloroethylene (PERC)). We did not include singly chlorinated methane or ethane, as these agents have different properties and, thus, different uses, namely, as refrigerants and gasoline additives rather than as industrial cleansers.

The exposure assessment for the six solvents was conducted after an extensive review of the occupational health literature. Approximately 350 papers, reports and documents on these solvents were reviewed. Values representing more than 5000 measurement results were abstracted and compiled,34 ,35 and determinants of exposure (mechanism of release (eg, evaporation), process condition (ie, open or closed), temperature, usage rate (quantity used), type of ventilation, location (indoors or out of doors), work in a confined space and proximity to the source) associated with the measurements were identified. If the exposure determinants listed above were not identified, rules were developed for assigning the determinant values by decade. For example, if a study measured exposure to a chemist but did not indicate the determinants of exposure such as the type of location where the subject worked, the measurements were assigned the ‘indoor’ location. Confidence ratings for the determinants values were assigned on a scale of 1 (low) to 4 (high).

From this information, over 20 task exposure matrices were developed for common tasks associated with the solvents (eg, degreasing, dry cleaning, lab work, hospital care, plastic manufacture) that assigned probability and frequency of exposure by decade (1930s–1990s). For intensity, the measurement data abstracted from the literature were modelled from their exposure determinants using maximum likelihood methods to estimate the arithmetic mean exposure level in parts per million (ppm) for each unique combination of exposure determinants and year.36

Probability, frequency and intensity of exposure to each solvent were assessed for each job according to decade after review of all the available information using the task exposure matrices but modifying them based on the subject-specific reported information. The probability that the subject was likely exposed given the available information was categorised as 0%, 1–10%, 10–49%, 50–89% or ≥90%. For all subjects with a probability >0%, a frequency was assigned as ≤1 h/week, 2–10, >10–20 or >20. Last, the industrial hygienist assigned a confidence score, which reflected the quality of the background information on which the probability and frequency estimates were made, using a 4-point scale ranging from 1 (low) to 4 (high).

To estimate intensity, each job with a probability >0% was evaluated for the same determinants of exposure following the same rules as for the measurement data. A model for each solvent was then used to assign the estimated exposure level to each of the six solvents to each job based on its assigned determinants. The intensity level was estimated on a continuous scale in parts per million.

For each of the six solvents, subjects were categorised as unexposed if the probability of exposure to a given solvent was 0% in all of their jobs. Subjects were categorised as possibly exposed if the maximum probability of exposure among all jobs held was >0% and ≤49%. Subjects were categorised as probably exposed if there was at least a 50% probability of exposure in any of the jobs held; thus, the categories to which subjects were assigned as unexposed, possibly exposed and probably exposed were mutually exclusive for a given solvent. Additional exposure metrics for probably exposed subjects included: duration of exposure (years of employment in jobs with exposure probability ≥50%), cumulative exposure (estimated as ppm-hours (intensity×frequency×years) summed across all jobs with exposure probability ≥50%), average weekly exposure (estimated as ppm-hours per week using cumulative exposure divided by years exposed) and highest exposure (estimated as ppm, the highest intensity achieved in any job) for each of the six solvents.

Possibly exposed subjects were included in initial analyses in which exposure was assessed as unexposed versus possibly exposed, but they were excluded from all subsequent analyses. This was due to the likely poor specificity of exposure classification for this group, as indicated by the unrealistically high prevalence of exposure (24%–33% possibly exposed) among controls. Given the importance of using high specificity exposure metrics when evaluating risks related to rare exposures37 we restricted the analyses for more detailed exposure metrics to unexposed and probably exposed subjects only.

Statistical analysis

Unconditional logistic regression models were used to calculate OR and 95% CI for risk of glioma or meningioma. Models were adjusted for age at diagnosis, sex, race/ethnicity, hospital site and residential zone defined by proximity to the hospital. Additional adjustment for cigarette smoking and education was made, but point estimates did not change materially, and, thus, we present results from the more parsimonious models adjusted only for the matching factors described above. Additionally, we adjusted for estimated cumulative occupational exposures to lead, magnetic fields, herbicides and insecticides, each of which has been implicated, to some degree, in previous studies of brain tumour risk.31 ,32 ,38 Separate models were fitted for each of the six solvents for each exposure metric. For ever/never analyses for particular solvents, we additionally included all other solvents in the model to account for possible confounding by other solvent exposures. We also evaluated risks related to exposure to any solvent, in which subjects were classified as exposed if they were probably exposed to any of the six solvents and unexposed if they were not exposed to any of the six solvents. To explore possible effect modification, we stratified ever/never analyses of glioma risk by gender. Too few cases of meningioma were exposed to have the power to conduct stratified analyses or analyses of more detailed exposure metrics (ie, duration of exposure, cumulative exposure, average weekly exposure and highest exposure).

For analyses of risk of glioma related to the more detailed exposure metrics, we used the median level of duration, cumulative, average weekly and highest exposure among controls as a cut-off point to classify all probably exposed subjects as low versus highly exposed for each exposure metric for a given solvent. We conducted analyses using unexposed subjects as the referent group as well as using the low exposed subjects as the referent group. The latter analyses accounted for the possibility that unexposed persons may be substantially different from exposed persons in ways that cannot be adjusted for in our analysis. Thus, comparing highly exposed subjects with those who had lower exposure levels may be a more accurate comparison for evaluating risks, assuming that a dose–response relationship exists. We also conducted analyses exploring the potential effects of a 10-year exposure latency period by adjusting the exposure metrics excluding any work performed within 10 years of diagnosis (interview).

For sensitivity analyses, we reran all the models excluding subjects categorised as probably exposed if their confidence was 1. Four subjects (two controls, one glioma case and one meningioma case) were dropped. To account for possible selection bias, we also reran our analyses systematically excluding each major subgroup of controls, as defined above. All models also were reanalysed excluding subjects for whom information was collected from a proxy respondent (n = 117). All analyses were performed using Stata V.11.0 (College Station, Texas, USA).