Analysis
Standard methods for analysis of case–control studies were used. The mean delay from diagnosis to interview was subtracted from the date of interview to calculate a reference date for duration of exposure for each control. Tertiles of cannabis use were determined by the marginal distribution ofuse for all subjects to reduce the chance of zero cell counts if just the control group was used. Relative risks (RRs) were estimated by calculating odds ratios by logistic regression and adjusted for confounding variables. Adjustment for age, joint- yrs of cannabis smoking and pack-yrs of cigarette smoking was made by including them as continuous variables in the regression models. The effects of categories of pack-yrs of cigarette smoking (quintiles of smoking for all subjects interviewed) and joint-yrs of cannabis smoking (tertiles of use for all subjects interviewed) were also assessed.
The RRs were also calculated based on cannabis use up to 5 yrs prior to diagnosis or a reference date in the controls, as exposure after that time was unlikely to have caused the malignancy. This 5-yr period was based on natural growth rates for lung cancer, in which the mean time from malignant
change to diagnosis is at least 8 yrs for nonsmall cell lung cancers, and ,3 yrs for small cell lung cancers [25, 26]. The age at which cannabis smoking started was categorised, and the RR associated with starting at ,16 yrs of age compared with o21 yrs was estimated. Various logistic regression models were fitted with potential confounders as continuous and categorical variables, and the estimates of RR and the confidence intervals (CIs) were not appreciably different from
the results presented. Differences in Akaike’s Information Criteria (AIC) were used to assess the linearity of the dose–response relationship of the risk of lung cancer after fitting parameters as continuous and categorical variables.