Methods
The Surveillance, Epidemiology, and End Results (SEER) database12 was established by the National Cancer Institute, covering approximately 28% of the US population. The Public Use version of data collected from the SEER18 registries12 between January 1, 1973, and December 31, 2016, was used for this study. As a comparison, the mortality data of the general US population collected by the National Centre for Health Statistics spanning from 1969 to 2016 was used.
To avoid the influence of the second primary tumours on the outcome of cancer patients, we only included patients diagnosed with the first primary ovarian cancer at the beginning of this study. Patients were excluded if the diagnosis was made at autopsy or obtained solely from the death certificate, and those without definite data on age at diagnosis, income, educational level, and survival time were also excluded in this study (Supplementary Figure.1 ).
Available data about demographic characteristics from the SEER database included age at diagnosis, race, and calendar year of diagnosis. The tumour-related variables encompassed pathological type, stage, and grade. Income (median family income) and educational level (percentage of residents >25 years of age with at least a bachelor’s degree) were categorized into quartiles. Residence types (rural/urban) were obtained at the county level from the SEER program. Data on ovarian cancer treatment, follow-up time, and cause of death were also available. Ovarian cancer was categorized into five histological subgroups according to the International Classification of Disease for Oncology third revision (ICD-O-3) codes: type I epithelial, type II epithelial, germ cell, sex cord-stromal, and other1314. Surgeries for ovarian cancer were categorized into “Oophorectomy without hysterectomy”, “Oophorectomy with hysterectomy”, “Surgery, not otherwise specified (NOS)”, “None”, “Debulking”, and “Pelvic Exenteration”.
Patients with the following causes of death codes in the SEER database were considered to have committed suicide: International Classification of Diseases, Eighth Revision codes (ICD-8): 950-959; International Classification of Diseases, Ninth Revision codes (ICD-9): 950-959; and International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10) codes U03, X60-X84, and Y87.0 and recode 50220. Patients with the following causes of death codes in the SEER database were considered to have committed accidental death: ICD-8: 800-949; ICD-9: 800-949; and ICD-10: V01-X59 and Y85-Y86 and recode 50210.
The number of suicides or accidental deaths divided by person-years of survival was calculated as the mortality of suicidal or accidental death. Among subgroups of cancer patients stratified by different characteristics, SMRs were calculated as the ratios of observed to expected number of deaths5 15 6, which provided the relative risk of death from suicide and accident injury for cancer patients compared with all general US population with the same distribution of age, sex, and race. The observed values represented the number of suicides or accidental deaths in patients with ovarian cancer, and the expected values represented the number of individuals who died by suicide or accident in the general population. A five-year age range was used for standardization, and the 95% CI of SMR was determined by using the Poisson distribution approximation. Interaction tests were further carried out to investigate the potential differential effect of cancer subtypes on suicide risk during different follow-up periods. Given that other causes of death were considered as competing events, we also designed the Fine-Gray model to identify demographic and tumour-related characteristics associated with a higher risk of suicide and accidental death. Observations were censored if patients did not die from suicide or accidental death at the time of the last follow-up. The survival time recorded as 0 month in the SEER database was converted to one-half of a month based on accepted epidemiologic practices15.
All statistical tests were two-sided, and values with P<0.05 were considered statistically significant. The SEER database was accessed using SEER*Stat software 8.3.812. The calculation of SMRs, cumulative incidence curves, Gray tests, and interaction tests were conducted in R version 3.51 statistical software. These histograms summarizing the results of subgroup analyses were drawn by Microsoft Excel and GraphPad Prism 8.0.