Potential confounding factors

Potential confounding factors.
We used data on smoking, alcohol consumption, education, body mass index, daily intake of fruit/vegetables, red meat, fat and dietary fibers, skin reaction to sun, hormone replacement therapy use, reproduction, and occupation collected at enrollment into the cohort to adjust for potential confounding factors for each type of cancer. Further, data was analyzed both with and without adjustment for enrollment area.
Statistical analyses.
Statistical analyses were carried out with the PHREG procedure of SAS 8.2 software (SAS Institute Inc., Cary, NC, USA). Cox proportional hazards models were used to estimate incidence rate ratios (IRRs) for cancer associated with the time-weighted average and cumulative arsenic exposure. Age was used as the time axis to ensure that the estimates were based on comparisons of individuals of the same age. The analyses were corrected for delayed entry, so that individuals were considered at risk only from the age at entry into the cohort. All analyses were stratified by sex. The age of 41 years was chosen as the starting point for calculating exposure because the oldest person in the cohort was 41 in 1970 when the residential histories began. This ensured that the exposure assessment started at the same age for all individuals. Both exposure measures—time-weighted average exposure per micrograms per liter since age 41 and total cumulated exposure per 5 mg since age 41—were included in the model as time-dependent variables.
Cohort members were censored at age of death, emigration, cancer diagnosis or if they had moved to an address with unknown arsenic concentration.
The associations between arsenic exposure measures and cancer risk were modeled as straight lines, and the analyses were adjusted for known risk factors to control for potential confounding. The linear assumption was evaluated with a linear spline (Greenland 1995), with three boundaries placed at the quartiles of the time-weighted average exposure at the time of enrollment into the cohort. These were included as covariates in the Cox model, and linearity was assessed graphically and by a numerical test. To evaluate linearity, we also tested if an additional quadratic term improved the model fit. Further, we analyzed risk in association with quartiles of cumulated exposure at the time of diagnosis for cancer cases to investigate if the risk was higher or lower in association with the highest exposures.
Analyses were repeated using time-weighted average exposure to arsenic on the basis of addresses occupied from 1970 for all cohort members regardless of age. An additional analysis included a 5-year lag between exposure and cancer diagnosis, that is, excluding the 5 years before the cancer diagnosis in the exposure assessment of cases and a similar period for the non-cancer cohort members. We also repeated analyses after exclusion of individuals who had lived at an address for which the closest water utility was used as the expected source of drinking-water.
Arsenic concentrations were generally higher for persons enrolled in the Aarhus area than those in the Copenhagen area. This might imply confounding from risk factors that were not accounted for that differed between the Aarhus and the Copenhagen areas. Therefore, we repeated analyses with adjustment for enrollment area, knowing that such adjustment limited the exposure contrasts evaluated in the analyses.
For initial analyses (without adjustment for area) providing p-values ≤ 0.2, separate estimates of the association with arsenic exposure were calculated for the Aarhus and Copenhagen enrollment areas by including an interaction term in the model. This was to test the consistency of the results. A persistent association in both areas would strengthen the confidence in the result. We tested the null hypothesis that the effect of arsenic exposure was the same in the two areas (test for interaction).
On the basis of the results of previous studies, the associations between exposure to arsenic and cancers of the lung and bladder and non-melanoma skin cancer were estimated for never, former, and current smokers separately, and we tested for different effects of arsenic between the three smoking status categories.