Due to the availability of data on medication since July 2005, we identified all patients with a newly diagnosed cancer (N = 403,322) between October 2005 and December 2014 from the Swedish Cancer Register, which includes almost complete information on all cancers diagnosed in Sweden since 1958 onward . All patients were cross-linked to the Swedish Causes of Death and Migration Registers using the personal identification numbers assigned uniquely to all residents in Sweden . We excluded 155 patients who died and 14,724 patients who emigrated before cancer diagnosis, leaving 388,443 patients to be followed from the date of cancer diagnosis, until emigration (Migration Register), death (Causes of Death Register), or December 31, 2014, whichever occurred first. We used 7th Swedish revision of the International Classifications of Diseases (ICD) codes to classify different cancer types (Supplementary Materials, Table S1).
Ascertainment of exposures
The Swedish Prescribed Drug Register contains information on prescribed and dispensed medications from all Swedish pharmacies since July 2005 . All pharmacies, retailers and wholesalers across the country are obligated to report the sales on monthly basis with overall very good data quality . Unused drugs are advised to be returned to the pharmacies for incineration. The proportion of all returned drugs was around 2.3–4.6% of the dispensed volume . Medications in this register are classified according to the Anatomical Therapeutic Chemical (ATC) system . The register includes information on medicine types, prescription and dispensing dates, quantity, defined daily dose and prescription text . Low-dose aspirin and most NSAIDs cannot be purchased over-the-counter without a prescription in Sweden . Patients who had medications dispensed with ATC codes B01AC06, B01AC30 and B01AC56 were considered as medicated with lose-dose aspirin (limited to daily dose of 75 or 160 mg). We focused on low-dose aspirin that tends to be used in long term in the present study, because of its potential effect of reducing stress-related outcomes [14, 21]. Patients who had medications dispensed with ATC code M01A were considered as medicated with non-aspirin NSAIDs.
We identified all low-dose aspirin and non-aspirin NSAIDs dispensed from 3 months before cancer diagnosis until the end of follow-up, because prescription drugs are dispensed for up to a three-month supply in Sweden. Multiple prescriptions at the same dispense date for the same medicine (2% for aspirin, 1.6% for non-aspirin NSAIDs) were summed up and unused medicines returned to the pharmacies were extracted from the amount of the previous dispense (0.3% for aspirin, 0.3% for other NSAIDs). Because of the time-varying nature of medication use, we constructed on- and off-medication periods for each patient after cancer diagnosis through information on dispense date and dosage according to the prescription text. The on-medication period was defined as the interval from the most recent dispense date of a specific NSAID, until the last day when the dispensed drug was estimated to be consumed. Time periods outside the on-medication periods among patients that had ever used NSAIDs, and among patients that did not use any NSAID during follow-up, were defined as off-medication periods (Supplementary Materials Fig. S1).
Because the defined daily dose does not necessarily correspond to the recommended or prescribed daily dose, we estimated days on medication as the division of the total amount of dispensed drug by the prescribed daily dosage per medicated period for each patient. The information on prescribed daily dosage for each on-medication period was extracted from the prescription text, or from the defined daily dose when the prescription text was not available (25% for aspirin, 4% for non-aspirin NSAIDs).
Ascertainment of completed suicide and death due to accident
The Swedish Causes of Death Register collects nationwide information from 1961 onward, including dates and the underlying and contributing causes of death , with high accuracy . We used the 10th Swedish revision of ICD (International Classification of Diseases) codes X60-X84, V01-X59, and Y85-Y86 to ascertain deaths from suicide and accident (Supplementary Materials Table S2). Death due to accident was further classified as deaths due to transport accident, fall, accidental threat to breathing, unspecified fracture, or others.
Patients’ use of drug as well as their general health status may confound the association between aspirin and non-aspirin NSAIDs and risk of unnatural deaths from suicide and accident. To control for these confounding factors, we included socioeconomic status (education, occupation, cohabitation status), mental health status (history of psychiatric disorder), and general health status (Chronic Disease Score) as covariates. The Longitudinal Integration Database for Health Insurance and Labour Market Studies (LISA) was established by Statistics Sweden and collects data from labor market, and educational and social sectors annually for individuals over 16 years of age . Information on the highest educational level, occupation, and cohabitation status at the time of diagnosis was retrieved from LISA for all cancer patients. Chronic Disease Score is a measure of comorbidity based on the aggregated number of prescribed medications. The score is a summary of weights from each comorbidity category represented by medication classes [25, 26]. As a proxy of general health status, we calculated a Chronic Disease Score, based on the prescribed medications before cancer diagnosis, for each patient , after excluding anxiolytics, antidepressant, antipsychotic, and anti-inflammatory medications because of their close relationship with the exposure or other covariables. We defined history of psychiatric disorders as having any inpatient or outpatient hospital visit for psychiatric disorders from 1987 onward using the ICD-9 codes 290–319 and ICD-10 codes F10-F99. Because of the close link between psychiatric disorders and suicide and death due to accident [27, 28], we updated this variable per on- or off-medication period for each patient.
We first described the demographic and clinical characteristics of the cancer patients, with and without medications, including sex, age at cancer diagnosis, calendar period of diagnosis (2005–2008, 2009–2011 or 2012–2014), highest educational level (> 12 years as after secondary school, 9–12 years as secondary school, < 9 years as primary school, or missing), occupation (blue collar, white collar, farmers, self-employed, retired or unemployed, or unclassified or missing), cohabitation status (cohabitation, non-cohabitation, or missing), cancer type (prostate cancer, breast cancer, colorectal cancer, non-melanoma skin cancer, hematopoietic malignancy, lung cancer, severe cancer [esophagus, liver and pancreas] and others), cancer stage (localized limited, localized advanced, regional spread, distant metastasis, unknown, or not applicable), history of psychiatric disorders (no or yes), and Chronic Disease Score (0, 1–2, 3–5, or > 6).
To assess the impact of taking aspirin and non-aspirin NSAIDs separately, we used Cox proportional hazards regression models to estimate the hazard ratios (HRs) and 95% confidence intervals (CIs) of suicide and death due to accident after cancer diagnosis, by comparing the on-medication periods with the off-medication periods of aspirin, and of non-aspirin NSAIDs, separately. In the first model, we used time since cancer diagnosis as the underlying timescale, and additionally adjusted the estimates for age at diagnosis (continuous variable), sex, highest educational level, occupation, cohabitation status, and calendar year of diagnosis. A continuous function of time was modeled with restricted cubic splines. In a second model, we additionally adjusted for cancer type, cancer stage, and Chronic Disease score (continuous variable). In a third model, in addition to all variables adjusted for the second model, we also adjusted for history of psychiatric disorders. The three models were designed to demonstrate the roles of different covariables on the studied associations. To illustrate the temporal pattern of the association, we also estimated the hazard as a function of time using restricted cubic splines with three knots that were evenly distributed along time at risk, based on the third model.
We performed sensitivity analyses. First, to assess the validity of our definition for on- and off-medication period, we defined the first month after each on-medication period also as on-medication period rather than off-medication period. Second, to separately evaluate the association for first-time use and repeated use of the medications, for each on- and off-medication period, we defined as with previous use if there was a previous on-mediation period since July 2005, and as without previous use if there was no previous on-medication period. Third, because a proportion of cancer patients had unknown stage at diagnosis, in a sensitivity analysis, we imputed unknown cancer stage to assess the influence of such missingness on the main results. Fourth, to exclude potential influence from other NSAIDs when investigating the effect of aspirin use (and vice versa), we compared the on-medication periods of aspirin or non-aspirin NSAIDs with off-medication periods with neither aspirin nor non-aspirin NSAIDs. Finally, to assess the role of other medications with potential impact on cognitive function and psychiatric symptoms, we performed additional analysis with further adjustment for use of opioids (ATC code N02A), use of anxiolytics (ATC code N05B), or use of antidepressants (ATC code N06A).
Because an association was mainly noted between the use of low-dose aspirin and accidental death, we also performed several secondary analyses to assess the robustness of this finding. First, because patients are often asked to discontinue aspirin use during surgical treatment to avoid major bleeding , we performed additional analysis to separately assess the studied associations within first year (as a proxy for the time window of primary cancer treatment including surgery) and beyond first year after cancer diagnosis. Second, to further alleviate the concern of residual confounding, we separately compared the on-medication periods with off-medication periods of the same individuals (within-individual comparison). Third, because individuals with different characteristics have been shown to have different risk of accidental death following a cancer diagnosis [30, 31], we separately analyzed the association by sex, age, cancer type, cancer stage, history of psychiatric disorders, Chronic Disease Score, highest educational level, cohabitation status, and employment status (employed vs. retired or unemployed). Finally, in addition to any death due to accident, we separately studied the association by major causes of accidental death.
We found no major violation of the proportional hazards assumption in all analyses by plotting Schoenfeld residuals. All analyses were performed in SAS 9.4 (SAS Institute) and STATA 14.1 (StataCorp LP, College Station, USA).
The study was approved by the Regional Ethical Review Board in Stockholm, Sweden (Dnr 2015/1574–31). Individual informed consent was waived in this approval.