The Danish Cancer Registry [17, 18] has recorded incident cases of cancer on a nationwide basis since 1943 and has been shown to have an accurate and almost complete ascertainment of cancer cases [17, 19]. Cancer diagnoses in this register were recorded according to the International Classification of Diseases (ICD) for Oncology from 1977 to 2003 (ICD-O-1–3), and ICD-10 since 2004.
The Danish National Patient Register contains data on all non-psychiatric hospitalizations in Denmark since 1978 and out-patient visits since 1995. Discharge/contact diagnoses have been coded according to ICD-8 from 1978 to 1993 and ICD-10 since 1994 . Virtually all medical care in Denmark is furnished by the national health authorities, whereby this data resource allows true population-based studies, covering all inhabitants of Denmark.
The Danish National Prescription Registry  contains data on all prescription drugs redeemed by Danish citizens since 1995. Prescription data includes the type of drug, date of dispensing and quantity. The dosing information and the indication for prescribing are not available. Drugs are categorized according to the Anatomic Therapeutic Chemical index, a hierarchical classification system developed by the WHO for the purposes of drug use statistics , and the quantity dispensed for each prescription is expressed by the defined daily dose (DDD) measure, also developed by the WHO .
The Danish Civil Registration System  contains data on vital status (date of death) and migration to and from Denmark, which allowed us to extract controls and to keep track of all subjects.
All data sources were linked by use of the personal identification number, a unique identifier assigned to all Danish residents since 1968 that encodes gender and date of birth . All linkages were performed within Statistics Denmark, a governmental institution that collects and maintains electronic records for a broad spectrum of statistical and scientific purposes [21, 24, 25].
Our study base initially consisted of all Danish residents alive on 1 January 2002. However, we applied a new user design by excluding all persons from our study base who redeemed a prescription for any anxiolytic, hypnotic or sedative (ATC-codes, N05B and N05C) during the first 2 running years of the prescription database, i.e. 1995 and 1996 [16, 21]. The rationale behind this approach was based on the assumption that users of BZRD drugs within these 2 years were likely to have used the drugs prior to 1995 as well. Inclusion of individuals with a long term exposure history prior to 1995 would potentially cause substantial misclassification in dose/intensity/duration sub-analyses. Since the possibility of a carcinogenic effect of small amounts of BZRD were raised recently , it was considered vital to minimize misclassification of small volume exposure.
For a subject to be eligible for sampling at a given date (index date) as either case or control, we required that the subject had lived in Denmark continuously from 1995 to the index date. Furthermore, the subject was required to be without any history of cancer (except non-melanoma skin cancer) prior to the index date.
BZRD were defined as any drug within the ATC groups N05BA or N05CD (benzodiazepine derivates), or N05CF (benzodiazepine related drugs).
Cases and controls were considered long term users of BZRD if they had redeemed a cumulative amount of BZRD equal to or greater than 500 DDD within a period of 5 to 1 year prior to the index date. The 1 year latency period was introduced since the use of certain drugs is known to increase within the last year prior to a cancer diagnosis, most likely because of early symptoms related to a yet undetected cancer . This would inflate the drug prevalence among cases and thereby introduce reverse causation bias , i.e. an artificial association between the use of BZRD and cancer. Another reason for requiring this latency was that we considered it very unlikely that recent exposure within the last year would contribute to the cancer risk.
One problem with studies on the use of BZRD is that treatment can be either chronic or episodic. We therefore performed exploratory analyses to define the duration that should be assigned to each prescription. An analysis of waiting time distributions  revealed that prescriptions for BZRD that were more than 15 weeks apart were unlikely to pertain to the same treatment episode. We thus assigned each prescription an exposure period of 15 weeks, i.e. 105 days. If the next prescription for a BZRD occurred within this exposure period, we assumed that the treatment episode had continued. If it occurred later, we assumed that treatment had been paused. No adjustment was made for overlap between prescriptions. Similarly, the exposure period assigned to single prescriptions or the last prescription in a treatment episode was 105 days.
The analysis was performed as a conventional matched case–control study based on a new user design. Odds ratios (ORs) for cancer associated with use of BZRD were calculated using conditional logistic regression adjusting for potential confounders. In all analyses, use of BZRD within a given time window was compared with no use, i.e. in our main analysis we compared use of more than 500 DDD of BZRD within a period of 5 to 1 year prior to the index date with no use (0 DDD) within the same period.
The following potential confounders were included in the regression model: a) use of drugs known or suspected to modify the risk of some cancers, including aspirin (ATC: B01AC06, N02BA01, N02BA51), non-aspirin NSAIDs (M01A, excluding M01AX), 5-α-reductase inhibitors (G04CB), statins (C10AA), angiotensin-II antagonists (C09C and C09D), antidepressants (N06A), antipsychotics (N05A, excluding lithium N05AN) and oral contraceptives and hormone supplements (G02BB01, G03AA07, G03AA09, G03AA10, G03AA11, G03AA12, G03AA13, G03AB, G03C, G03D, G03F, G03HB01). Exposure to a confounder drug was defined by a cumulated dose of at least 500 DDD within a period of 5 to 1 year prior to the index date; b) prior diagnoses of diseases known or suspected to modify the risk of some cancers, established more than 1 year prior to the index date, including inflammatory bowel disease (IBD) (ICD-8: 563.01, 563.19, 569.04; ICD-10: K50, K51.0-K51.3), chronic obstructive pulmonary disease (COPD) (as a crude marker of heavy smoking) [composite measure of diagnoses (ICD-8: 490.00, 491.00, 491.01, 491.03; ICD-10: J42, J43, J44) or more than 500 DDD for drugs for obstructive airway diseases (ATC-group R03)], alcohol abuse [composite measure of diagnoses (ICD-8: 291, 303, 577.10, 979, 980; ICD-10: F10, G31.2, G62.1, G72.1, I42.6, K29.2, K86.0, R78.0, T51, Z72.1) or any prescription for a drug used to treat alcoholism (N07BB01, N07BB03, N07BB04)] and diabetes [composite measure of diagnoses (ICD-8: 249.00, 249.09, 250.00, 250.09; ICD-10: E10-E14) or any prescription for an anti-diabetic (A10)]; c) Charlson Comorbidity Index (CCI) score [30, 31], in which each disease category has an associated weight based on the adjusted risk of 1 year mortality. We disregarded diagnoses included in the CCI score established less than 1 year prior to the index date. We defined the level of comorbidity as none (CCI score: 0), low (CCI score: 1) and medium/high (CCI score: ≥2).
To evaluate the potential influence from confounding by lifestyle factors, we categorized cancers as related to tobacco smoking, alcohol consumption or obesity. Cancers related to tobacco smoking were defined as cancers of the buccal cavity and pharynx, oesophagus, stomach, colorectum, liver, pancreas, nasal cavity and paranasal sinuses, larynx, lung, cervix, ovary, kidney, renal pelvis or ureter, urinary bladder or myeloid leukaemia . Cancers related to alcohol were defined as cancers of mouth, pharynx, oesophagus, colorectum, liver, larynx or breast . Cancers related to obesity were defined as cancers of the oesophagus, colorectum, pancreas, breast, endometrium or kidney .
We performed sub-group analyses, specified by (i) age and gender; (ii) various exposure measures, including cumulative use, intensity and duration of use; (iii) comorbidity; and (iv) type of BZRD. In addition, we performed some pre-planned sensitivity analyses. First, the main analysis was repeated with exposure limits of 100 or 250 DDD, instead of 500 DDD. Secondly, we performed the analyses according to separate exposure groups of either benzodiazepines or BZRD. Lastly, the 1 year latency period, i.e. the exclusion of exposure to BZRD or confounders within the last year prior to the index date, was varied from 0 to 2 years.
In the recent cohort study of hypnotic use and cancer risk or mortality , the authors excluded all individuals from the unexposed population who at any time during the complete observation period redeemed a prescription for a hypnotic drug. Thereby, cancer cases among unexposed individuals who redeemed a prescription for a hypnotic after their cancer diagnosis were excluded from all analyses. This potentially created a spurious association between hypnotic use and cancer risk . To evaluate the magnitude of this error, we performed a biased analysis, mimicking this approach by repeating our analysis, only this time excluding all unexposed cases or controls who redeemed a prescription for a BZRD at any time after their index date.
All analyses were performed using Stata Release 12.0 (StataCorp, College Station, TX, USA).