We conducted this population-based case–control study in North Jutland County, with a population of about 500
000 inhabitants (approximately 9% of the Danish population) (Sorensen et al, 2008
). Use of the unique 10-digit civil registry number, assigned to all Danish residents (Frank, 2000
), allowed us to link data on cancer incidence, drug prescriptions, and hospital in- and outpatient diagnoses.
From the Danish Cancer Registry (DCR), we identified all patients registered with a diagnosis of skin cancer or NHL in North Jutland County during 1989–2003. The DCR, established in 1943, records primary cases of cancer on a nationwide basis with a high degree of accuracy and completeness (Storm et al, 1997
). Data in the DCR include cancer type, site, morphology, and date of diagnosis. Tumours are coded according to a modified version of the seventh revision of the International Classification of Diseases (ICD-7). Since 1978, tumours have also been classified according to the first version of the International Classification of Diseases for Oncology (ICD-O-1), which includes a four-digit code for tumour morphology.
We used ICD-7 codes 1910–1919 to identify all BCC and SCC cases registered in North Jutland County during 1989–2003. For MM we used ICD-7 codes 1900–1909, and for NHL we used ICD-7 codes 2000–2009 and 2020–2029. We excluded patients with cancer diagnosis or organ transplantation before the skin cancer or NHL diagnosis. For BCC we included only patients with ICD-O-1 morphology codes 80903, 80913, 80923, 80933, and 81233. For SCC we included only patients with ICD-O-1 codes 80513, 80703, 80713, 80743, 80763, 80943, and 80953. We identified 5422 BCC, 935 SCC, 983 MM, and 481 NHL cases.
The Civil Registration System contains information about the vital status, date of death, and the area of residence of all Danish residents, and is updated daily. Using the Civil Registration System, we selected four population controls for each case. Cases and controls were individually matched by age and gender on the basis of risk-set sampling (Wacholder et al, 1992
); that is, the controls were sampled among county residents who were at risk of a first skin cancer or NHL at the time the corresponding case was diagnosed (index date assigned to controls). We excluded controls with cancer diagnosis or organ transplantation before the index date. A total of 29
664 controls were identified. Using risk-set sampling, the estimated exposure odds ratio in a case–control design is an unbiased estimate of the incidence rate ratio (IRR) in a corresponding cohort study (Szklo and Nieto, 2000
Data on oral glucocorticoid prescriptions were obtained from the Prescription Database of North Jutland County, a research database based on National Health Service data (Gaist et al, 1997
; Nielsen et al, 1997
). The database collects data on all prescriptions filled by ambulatory patients and forwards data on reimbursable medicines to the local regional Health Service section on a monthly basis. This Health Service, in turn, refunds 50–75% of costs. The database was established in 1989 (with complete coverage since 1991), and includes patients' civil registry number as well as information on the type of drug prescribed (according to the Anatomical Therapeutical Chemical (ATC) classification system) (Dukes, 1993
), the date when the prescription was filled, packaging size, the number of pills in each package, and the amount of drug in each pill. Indications for drug use and dosing schedules are not included. The ATC codes for glucocorticoids used in this study are listed in Appendix 1
Individuals with certain chronic medical conditions are more likely to receive prescriptions for glucocorticoids. Some of these conditions are also associated with an elevated risk of skin cancer (Jensen et al, 2008
) and NHL (Kinlen, 1992
). To control for the potentially confounding effects of chronic medical conditions, we used the Danish National Registry of Patients to retrieve all hospital diagnoses recorded among our study population from January 1, 1977 to December 31, 2003. Diagnoses are coded according to the ICD-8 classification through 1993 and the ICD-10 thereafter. The Danish National Registry of Patients contains information on all non-psychiatric hospital admissions since 1977 (Andersen et al, 1999
) and outpatient visits since 1995. We classified the diagnoses of chronic diseases into the four categories listed in Appendix 2
. In addition, we retrieved information on prescriptions redeemed for azathioprine (ATC code: L04AX01) and methotrexate (ATC code: L01BA01), as use of these drugs has been associated with an increased risk of skin cancer (Glover et al, 1997
; Jensen et al, 1999
; Marcen et al, 2003
) and NHL (Kinlen, 1992
For each subject, we identified all prescriptions for oral glucocorticoids before the date of primary skin cancer or NHL diagnosis, or before the corresponding index date for the matched population control. We computed the total amount of drug prescribed before the index date by multiplying the package size, the number of pills in each package, and the amount of drug in each pill. When the amount of drug could not be calculated (for instance, due to missing information in the prescription database), the average amount dispensed for that particular drug was estimated. Average amounts were applied to 13% of the identified prescriptions for oral glucocorticoids.
Initially, we constructed contingency tables for cases and controls by demographic characteristics (age and sex), anatomic site of the skin tumour, prior hospitalisations for the selected comorbid conditions, and previous prescriptions for azathioprine and methotrexate. Three models were used for the various combinations of oral glucocorticoid exposure and outcome (i.e., BCC, SCC, MM, and NHL). The first model treated the exposure as a dichotomous variable (i.e., any/no prescription for glucocorticoids before the index date). The second model assumed a linear effect of the exposure (i.e., as a continuous variable based on the estimated milligrams of prescribed glucocorticoids, with ‘no prescriptions before the index date' as reference). The third model was fitted to test for a nonlinear effect by treating the exposure as a restricted cubic spline (Harre et al, 1988
). Trend tests were used to evaluate the statistical significance of the effect of increasing amounts of prescribed oral glucocorticoids on the risk of these cancers (dose–response relationship). P
-values <0.05 were considered statistically significant. We included the four chronic disease categories, listed in Appendix 2
, and prescriptions for azathioprine and methotrexate as confounding factors. These variables were included in the models as dichotomous variables, that is, yes/no disease and yes/no prescription.
As cancer diagnoses might be differentially related to glucocorticoid use due to surveillance bias (Rothman, 2002
), we also conducted analyses excluding prescriptions issued within one year of the skin cancer diagnosis or index date. These analyses might also point to possible effects on the duration of glucocorticoid use. We also conducted analyses stratified on the anatomic site of the skin cancers (i.e., head and neck and other sites), to evaluate the effects of glucocorticoids by level of sun exposure.
Non-reporting of diagnosed non-melanoma skin cancer (NMSC) to the DCR has been estimated to range from 12 to 40% (Frentz, 1996
; Jensen et al, 2007
). Therefore, we conducted a sensitivity analysis (Fox et al, 2005a
; Greenland, 2005
) to explore the magnitude of effects of the non-reporting of NMSC cases on our results. We expected the non-reporting to be differential between users and non-users of glucocorticoids; NMSC is rarely fatal, and therefore in severely diseased patients (such as users of glucocorticoids) clinicians could potentially deem these cancers trivial and thereby omit registration. We used the SAS-macro written by Fox, Lash and Greenland, which was adapted to perform conditional logistic regression (Fox et al, 2005b
This study was approved by the Danish Protection Agency (Record no. 2004-41-4693). The statistical software packages R, version 2.4.1, and SAS, version 9.1 (SAS Institute Inc., Cary, NC), were used for all statistical analyses.