Ascertainment of incident cancer cases
All participants were followed by electronic record linkage to routinely collected NHS data on cancer registrations, deaths and emigrations. This enabled virtually complete follow-up (only 1.4% of the cohort have been lost to follow-up) [
39]. Cancer diagnoses were coded to the 10
th revision of the World Health Organisation’s International Classification of Diseases. We examined the 21 most common cancers: oral cavity and pharynx (C00-C14), larynx (C32), stomach (C16), colorectum (C18-C20), liver (C22), pancreas (C25), lung (C34), malignant melanoma (C43), breast (C50), cervix (C53), endometrium (C54), ovary (C56), renal cell carcinoma (C64), bladder (C67), brain (C71), thyroid (C73), non-Hodgkin’s lymphoma (C82-C85), multiple myeloma (C90), and leukaemia (C91-C93, C95) and cancer of the oesophagus (C15), which was investigated separately by subtype, given the known differences in aetiology. Classification of oesophageal cancers by subtype was based on the following ICD-O morphology codes: oesophageal squamous cell carcinoma (OSCC): 8050–8086, and oesophageal adenocarcinoma (OAC): 8140–8145, 8190–8231, 8260–8263, 8310, 8315, 8480–8490, 8570–8575.
Statistical analyses
The baseline for analyses of alcohol intake in relation to cancer risk was the recruitment questionnaire in median year 1998. Of the 1,364,268 women eligible for the analysis, 40,640 were excluded because they had a cancer (other than non-melanoma skin cancer, C44) registered before recruitment, 10,400 were excluded because they had missing information on alcohol consumption and, given the close correlation between alcohol consumption and smoking, a further 74,597 were excluded because their smoking status at baseline was not known. A further 438,056 women were excluded because they reported drinking no alcohol or less than 1 drink per week. This exclusion was done to account for the fact that most of the women who reported drinking less than one drink per week were ex-drinkers, [
40] who may have stopped drinking due to poor health and could have different characteristics from drinkers, which might be difficult to measure but could be associated with the development of cancers [
41]. Given that we were interested in the potential interaction of alcohol with MHT use, the analysis sample was further restricted to postmenopausal women (defined as those who reported that they had experienced natural menopause or had undergone bilateral oophorectomy). Women who were premenopausal, perimenopausal, or of unknown menopausal status at recruitment were assumed to be postmenopausal after they reached the age of 55 years and were included in follow-up from age 55 years, because 96% of women in this cohort with a known age at natural menopause were postmenopausal by that age. Therefore only 5454 women (who had not previously reported natural menopause or bilateral oophorectomy) were excluded either because they were diagnosed with cancer prior to 55 years old or because they exited the cohort before they reached 55 years old.
After exclusions, 795,121 women remained, for whom person-years were calculated from the date that the recruitment questionnaire was completed to the earliest of the following: first registration with cancer, death, emigration, or end of follow-up on 31st December 2018. For analyses relating to cervical and endometrial cancers (n = 600,036), women were only included in the analysis if they had not reported a hysterectomy prior to recruitment. For analyses relating to ovarian cancer (n = 699,491), women were only included if they had not reported a bilateral oophorectomy prior to recruitment. For analyses relating to breast cancer (n = 756,462), a small proportion of women (4.9%) were excluded if they returned their recruitment questionnaire by post after their breast cancer screening appointment to ensure their answers were not influenced by any information gained at their screening appointment.
Women were grouped into four categories according to the total number of alcoholic drinks (i.e. 1 glass of wine, half a pint of beer/lager/cider, or 1 measure of spirits) reported at baseline: 1–2, 3–6, 7–14, and ≥ 15 drinks per week. Women who reported drinking 1–2 drinks per week were taken as the reference category. Cox proportional hazard models were used, with time since recruitment as the underlying time variable, to estimate the hazard ratios (henceforth referred to as relative risks [RRs]) and their 95% confidence intervals (CIs) for the various cancer sites. Analyses were stratified by year of birth and year of recruitment. Analyses were adjusted for the following variables recorded at recruitment: five regions of residence (London and Southeast, Southwest, Midlands, Northern England, Scotland), deprivation quintile (according to the Townsend index which is a score incorporating census area data for employment, car ownership, home ownership and household overcrowding [
42]), educational qualifications (tertiary, secondary, technical, none), smoking (never, past, current < 5, 5–9, 10–14, 15–19, 20–24, ≥ 25 cigarettes per day), oral contraceptive use (never, ever), menopausal hormone therapy (MHT) use (never, past, current), BMI (< 20.0, 20.0–22.4, 22.5–24.9, 25–27.4, 27.5–29.9, 30.0–32.4, 32.5–34.9, ≥ 35 kg/m
2), and strenuous physical activity (defined as exercise that is enough to cause sweating or a fast heart beat; never/rarely/ < once, once, 2–7 times per week). In order to include the same women in all analyses, a separate category was created for the small number of women with missing data for each adjustment variable (< 5% for each variable). For analyses relating to breast, ovarian, cervical and endometrial cancers, additional confounders were also included: age at menarche (< 12, 12–13, ≥ 14 years-old) and a combined variable of age at first birth (< 25, ≥ 25 years) and parity (nulliparous, 1-2, ≥ 3); and in breast cancer analyses, adjustment was also made for family history of breast cancer (yes, no). Potential violations of the proportional hazards assumption were assessed by examination of findings separately by 5 year follow-up periods. When comparing more than one group, group-specific 95% confidence intervals were calculated to allow direct comparison between any two groups [
43].
The association of alcohol intake and cancer incidence was summarized in the form of a log-linear trend in risk per increase of one drink per day, and, for the purposes of this analysis, we considered a trend with p < 0.05 to be statistically significant. To correct for measurement error, including potential changes in intake over time, the test for trend across categories of intake was based on the mean alcohol intake (drinks/week) reported by 412,971 women at resurvey in median year 2006 (IQR 2006–2006) in each baseline alcohol intake category. We also conducted a sensitivity analysis excluding the first 5 years of follow-up, to assess whether any association was affected by reverse causality bias, whereby early symptoms of cancer might cause changes in alcohol intake.
The associations between alcohol and cancer incidence were further investigated in subgroups of women according to smoking status (never, past, current), BMI (< 25, 25–29, ≥ 30 kg/m2) and use of MHT (never, past, current) at baseline. Interactions between alcohol and the other factors were tested using a likelihood ratio test, and we adjusted for false discovery rate (FDR) using the Benjamini–Hochberg method, and an FDR threshold of 0.05 was set for significance. In analyses of interactions, all upper aerodigestive cancers (defined as cancers of the oral cavity, pharynx and larynx and OSCC) were combined and their interaction with smoking status at four levels (never, past, current < 10 cigarettes/day, current ≥ 10 cigarettes/day) was also calculated and presented graphically. For hormone-related cancers (breast, cervical, endometrial and ovarian), further analyses restricted the alcohol-BMI interaction to never users of MHT and restricted the alcohol-MHT interaction to women of normal BMI (20–24.9 kg/m2).
In order to assess the likely impact of reducing alcohol intake among women who reported drinking at least 1 drink per week, we estimated the proportion of aerodigestive, breast, colorectal, and pancreatic cancers, attributable to drinking more than 1–2 drinks per week, based on the distribution of women, and estimated relative risks by alcohol (and in the case of aerodigestive cancers by alcohol and smoking).
Stata version 17 (StataCorp, College Station, TX, USA) was used for all analyses. All statistical tests were two-sided.