pidemiology (INHANCE) Consortium (http://inhance.iarc.fr/
) was established in 2004, based on the collaboration of research groups leading large epidemiology studies of head and neck cancer that are on-going or have been recently completed. We pooled the data from 18 individual case-control studies (version 1.1), including 12,282 cases and 17,189 controls (3
). Compared to our previous publication (20
), the current dataset added a Rome study(5
), New York multicenter study(14
) and Boston study(16
). In this current analysis, we excluded from the analyses a French study (323 cases and 234 controls) that was restricted to regular smokers (4
) and the Sudan (106 cases and 151 controls) and India (576 cases and 582 controls) centers of the International Multicenter study(12
) because of the small number of subjects to represent these regions for estimation of population attributable risks. Additionally in India, contrary to other countries, betel quid and areca nut chewing are major contributors to attributable fractions of oral cavity cancer.
Characteristics of the individual studies are presented in table 1 in the appendix
. Most were hospital-based case-control studies and frequency matched their controls to the cases on age, sex and additional factors (study center, hospital, race/ethnicity). The Los Angeles study individually matched the control subjects to case subjects on age decade, gender and neighborhood, though in the analysis the matching was broken. Face-to-face interviews were conducted in all studies except for the Iowa study, in which subjects completed self-administered questionnaires.
Written informed consent was obtained from all study subjects and the investigations were approved by institutional review boards at each of the institutes involved. Questionnaires were collected from all the individual studies, to assess the comparability of the collected data and of the wording of interview questions among the studies. Data from individual studies were received at the International Agency for Research on Cancer (IARC) with personal identifiers removed. Each data item was checked for illogical or missing values and inconsistencies were resolved as necessary.
Cases and controls with missing data on age, sex, or race/ethnicity, and cases with missing information on the site of origin of their cancer were excluded (56 cases and 54 controls). Cases were included in this study if their tumor had been classified by the original study as an invasive tumor of oral cavity, oropharynx, hypopharynx, oral cavity or pharynx not otherwise specified (NOS), larynx, or head and neck cancer unspecified. Subjects with cancers of the major salivary glands (parotid, submandibular, or sublingual glands; ICD-O-2 codes C07-C08), or of the nasal cavity/ear/paranasal sinuses (ICD-O-2 codes C30-C31) were excluded from the analysis. The ICD coding used for the classification into subsites was specified in detail previously(20
). Thus the data for this analysis included 11,221 head and neck cancer cases and 16,168 controls from 17 studies. There were a total of 2,993 oral cavity cancer cases, 4,040 oropharyngeal and hypopharyngeal cancer cases (pharyngeal), 917 unspecified oral cavity/pharynx cases, 2,965 laryngeal cancer cases and 306 unspecified head and neck cancer cases. We focused our site-specific analyses on oral cavity, pharyngeal and laryngeal cancers. Three of the studies did not collect information on tumor histology. Of the studies that collected histology, 86.7% (8034/9265) of head and neck cancer cases were squamous cell carcinoma (SCC).
The questions about tobacco smoking and alcohol drinking in the study questionnaires were conceptually similar across studies, although the exact wording differed. The questions about tobacco and alcohol use were examined carefully for comparability before variables were created for this analysis (definitions for being a cigarette, cigar, or pipe smoker for each study are provided in the appendix
). Variables on the frequency (i.e., number of cigarettes, cigars, or pipes smoked per day), duration (in years), and pack-years (i.e., cumulative smoking) of tobacco smoking were available in all studies.
Information about snuff use and chewing habits was collected by the Puerto Rico study, the International multicenter studies, and all studies in North America. Snuff use and chewing are not common behaviors in Europe or Latin America, except in specific populations (e.g., Norway and Sweden) that were not included in the pooled dataset (definitions of ever chewing and ever use of snuff are provided in the appendix
). Frequency and duration variables for chewing and snuff use habits were pooled across relevant studies. For this study, never users of tobacco were defined as individuals who had not used cigarettes, cigars, pipes, snuff, or chewing products during their lifetimes. A combined tobacco frequency variable was created, where ever tobacco users were categorized as having used 1–20 cigarettes, 1–20 cigars, 1–20 pipes, 1–2 chewing products or 1–2 snuff units per day, or >20 cigarettes, >20 cigars, >20 pipes, >2 chewing products or >2 snuff units per day.
In the alcohol section of the study questionnaires, subjects were asked if they were alcohol drinkers (definitions by study in appendix
); for those who responded that they were, subsequent questions were asked about the frequency of drinking, the duration of drinking, and the different types of alcoholic beverages consumed (i.e., beer, wine, hard liquors, and/or aperitif). Details on the pooling of frequency and duration variables on alcohol are provided in the appendix
The interactions between tobacco and alcohol on the risk of head and neck cancer were assessed by estimating odds ratios (ORs) and 95% confidence intervals (95%CIs) using unconditional logistic regression models for each study. To assess interactions on the multiplicative scale, we estimated odds ratios for joint effects (OR11
= OR for ever tobacco/ever alcohol use, OR01
=OR for never tobacco/ever alcohol use, OR10
= OR for ever tobacco/never alcohol use). The multiplicative interaction parameters & 95%CIs [ψ=OR11
)] were also estimated by including variables for ever alcohol use, ever tobacco use and a product term (equivalent to the multiplicative interaction parameter) of those two variables in the logistic regression model. ψ>1 is suggestive of a joint effect that is greater than expected under the multiplicative model. When a joint effect greater than multiplicative was not observed, interactions on the additive scale were assessed with relative excess risk due to interaction (RERI), attributable proportion (AP, proportion of disease among those with both exposures that is attributable to their interaction) and synergy index (SI)(21
). We estimated 95% confidence intervals for each of these measures. The null values of RERI and AP are 0, while the null value for SI is 1.
The logistic regression models included age (<40 years, 40–44 years, 45–49 years, 50–54 years, 55–59 years, 60–64 years, 65–69 years, 70–74 years, or ≥75 years), sex, education level (no formal education, less than junior high school, some high school, high school graduate, vocational/some college, or college graduate/postgraduate), race/ethnicity (non-Hispanic white, Black, Hispanic/Latino, Asian/Pacific Islander, other, Latin American), and study center to adjust for potential confounders. We tested for heterogeneity among the study ORs by conducting a likelihood ratio test comparing a model including the product terms between each study (other than the reference study) with the variable of interest and a model without the product terms (degrees of freedom = number of studies −1), for the risk of head and neck cancer combined and for the risk of each of these head and neck cancer subsites. Heterogeneity was detected consistently; therefore, to calculate the summary estimates of association, the study-specific estimates were included in a two-stage random effects logistic regression model with between-study variability and the common odds ratio being estimated using maximum likelihood estimation.
Information on ethnicity was not collected in the Central Europe and Latin America studies. In the Central Europe study, all subjects were classified as non-Hispanic White, since the large majority of these populations are expected to be White. In the Latin American study, we categorized subjects as Latin American. For the Latin America study only, study center was used as a proxy variable for race/ethnicity in all logistic regression models because each center had an expected predominant ethnic group distribution.
For subjects with missing data on education level (655 cases and 544 controls), we applied multiple imputation with the PROC MI procedure in SAS statistical software 9.1. We assumed that the education data were missing at random (MAR) with respect to unmeasured covariates; whether or not education level was missing did not depend on any other unobserved or missing values (22
). We used a logistic regression model (23
) to predict education level for each of the geographic regions separately (North America, Europe, Latin America) using age, sex, race/ethnicity, study, and case/control status as the covariates. The logistic regression results to assess summary estimates for cigarettes and alcohol drinking for five imputations were combined by using the PROC MIANALYZE procedure in SAS statistical software.
Stratified analyses were conducted by cancer site (oral, pharynx, larynx), sex, age (<45 years, 45–60 years, ≥60 years), education (<high school, ≥high school), geographic region (Europe, North America, Latin America), type of controls (hospital-based, population-based), study size (<500 cases, ≥500 cases), BMI 2 to 5 years before diagnosis (<18.5 kg/m2, 18.5 –<25 kg/m2, 25 – <30 kg/m2, ≥30 kg/m2) and BMI at age 20 or 30 and after restriction to squamous cell carcinoma cases.
The population attributable risks (PAR) were estimated based on the formula AF = p(ec) × (OR−1)/OR, where p(ec)is the proportion exposed among the cases (24
). Odds ratios adjusted for potential confounding factors were used in these equations. The confidence intervals for the AFs were calculated from the lower and upper limit ORs. The PAR for tobacco and/or alcohol exposures (PARtotal
) were estimated with the equation below where a00
=never tobacco/never alcohol users, a01
= never tobacco/ever alcohol users, a10
= ever tobacco/never alcohol users, and a11
=ever tobacco/ever alcohol users, m = total number of cases.
ORs from multivariate analysis were used. The PARs for the tobacco and alcohol were estimated as (31