Experimental design and ethical considerations

This retrospective population-based cohort study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. ^{Reference von Elm, Altman, Egger, Pocock, Gøtzsche and Vandenbroucke12} The authors state that all procedures used in this study were in accordance with the ethical requirements of the appropriate national and institutional human experimentation committees, as well as the Helsinki Declaration of 1975, as revised in 2013. Our institutional review board approved all procedures involving human participants and patients (approval number: X-2303-819-902). The Big Data Center of the National Health Insurance Service (NHIS) (NHIS-2023-1-526) authorised access to the data, including sharing rights. As the data were anonymised and the study was retrospective, it was not necessary to obtain informed consent from patients.

NHIS database

This study used data sourced from the NHIS, a public insurance programme in South Korea. The NHIS database, as required by law, includes detailed records of all disease diagnoses and associated prescription information for treatments and medications. To be eligible for government-funded health insurance programmes, individuals must enrol in the NHIS, and all diagnoses are reported and categorised on the basis of the ICD-10. Individuals residing in South Korea for more than 6 months must complete their registration with the NHIS. Furthermore, the NHIS database offers precise information about socioeconomic status and date of death. ^{Reference Lee, Lee, Park, Shin and Kim13}

Study population

Data were initially obtained from the NHIS concerning adult patients (≥18 years old) diagnosed with psychiatric disorders and receiving treatment during the period from 1 January to 31 December 2017. The ICD-10 codes for the psychiatric disorders examined in this study are listed in Supplementary Table 1 available online at https://doi.org/10.1192/bjo.2025.783. Upon obtaining data indicating that the population exceeded 5 million, we requested a stratified 10% sample categorised by age and gender. In the present study, adults identified as having psychiatric disorders in 2017 were assigned to the psychiatric disorder group. Following identification of all individuals eligible for inclusion in this group in South Korea in 2017, we requested data for individuals who had never been diagnosed with psychiatric disorders, to form our non-psychiatric-disorder group. This was accomplished using a 1:1 stratified random sampling technique, considering both age and gender. Owing to tracking of cancer diagnoses from 2018 to 2022, individuals who died in 2017 were excluded from the study. Furthermore, individuals with a cancer diagnosis before 2018 were excluded, as the focus of the study was cancer cases diagnosed from 2018 onward.

Experimental endpoints

The primary study outcome was new cancer diagnoses, assessed over a 5-year period from 1 January 2018 to 31 December 2022. Cancer diagnoses were made using ICD-10 codes (C00–C96). Patients with cancer in South Korea are registered in the NHIS database, which includes the assignment of a specific code known as the C code (i.e. a cancer diagnosis), entitling such individuals to unique insurance benefits. The following cancers were uniquely categorised: gastric cancer (C16), oesophageal cancer (C15), colorectal cancers (C18–C20), gallbladder and biliary tract cancers (C23–C24), head and neck cancers (C00–C14), brain cancer (C71), liver cancer (C22), pancreatic cancer (C25), lung cancer (C34), bone and articular cartilage cancers (C40–C41), breast cancer (C50), female genital organ cancers (C51–C58), male genital organ cancers (C60–C63), urinary tract cancers (C64–C68), thyroid cancer (C73), and lymphoma and leukaemia (C81–C96).

Collected covariates

The primary demographic variables of the patients were age and gender. Socioeconomic status was determined on the basis of household income, residence and employment status (including self-employment). Residences were classified as urban (Seoul and other large cities) or rural (all other regions). Household income was divided into five groups, one of which comprised individuals engaged in a medical aid programme, whereas the remaining groups followed a four-quartile distribution. Individuals unable to afford insurance are classified by the government as eligible for medical assistance.

To evaluate the occurrence of comorbidities, the Charlson Comorbidity Index was constructed using the most frequently occuring ICD-10 codes from the NHIS database (Supplementary Table 2). In South Korea, people must declare any disability in the NHIS database in order to be eligible for social welfare benefits. Thus, information about underlying disabilities was also gathered. In South Korea, specialists in relevant professions conduct disability evaluations. A qualified physician assesses the impact of a condition on everyday life, and it is classified as mild to moderate or severe. Supplementary Table 3 provides the full classification of disabilities.

Statistical methodology

Continuous variables were assessed through calculation of means and standard deviations. Categorical variables are expressed in terms of percentages and absolute values. To reduce the variation in variables between the psychiatric disorder and non-psychiatric-disorder groups, we performed propensity score matching (a method to decrease bias in observational research ^{Reference Austin14} ) via the nearest neighbour method, with a calliper width of 0.25 and a 1:1 ratio without replacement. In the propensity score modelling, age, gender, residence, household income level, underlying disability, Charlson Comorbidity Index and underlying diseases were included. Following propensity score matching, the absolute standardised mean difference (ASD) was used to determine the balance between the two groups. An ASD of less than 0.1 was considered to indicate proper propensity score matching.

Cox regression analysis was performed on the propensity-score-matched cohort to compare the frequency of newly diagnosed cancers between the psychiatric disorder and non-psychiatric-disorder groups. An event was defined as a cancer diagnosis made on or after 1 January 2018. Time was defined as the period between 1 January 2018 and the diagnosis. In addition, a time-lag analysis was performed in the propensity-score-matched cohort to determine whether psychiatric disorders preceded the cancer or were diagnosed in response to early symptoms. Herein, a cancer diagnosis made on or after 1 January 2019 was considered as an event. A hazard plot for cancer diagnoses in the propensity-score-matched cohort was constructed using Kaplan–Meier estimation, and the log-rank test was used to compare the cumulative incidence of total cancer diagnoses between the two groups.

A sensitivity analysis was performed using multivariable Cox regression modelling to compare the results for the propensity-score-matched group with those for the entire cohort. The model incorporated all covariates and was adjusted for several factors. To enhance our understanding of the risks associated with a cancer diagnosis, we conducted further multivariable Cox regression analyses to compare individuals diagnosed with two, three or more psychiatric conditions with those diagnosed with only one psychiatric disorder. Additional multivariable Cox regression analyses were performed to investigate whether the likelihood of a cancer diagnosis differed among groups with specific psychiatric disorders.

Subgroup analyses were conducted on the basis of gender and age. Log–log plots were used to verify adherence to the foundational assumptions of the Cox proportional hazards models. The results are presented as hazard ratios accompanied by 95% confidence intervals. The multivariable model demonstrated no multicollinearity among the variables, as indicated by a variance inflation factor of 2. For the statistical analyses, we used R version 4.0.3 for Windows (R Foundation for Statistical Computing, Vienna, Austria; see https://cran.r-project.org/bin/windows/base/old/4.0.3/), with the significance threshold set to P < 0.05. The ‘MatchIt’ and ‘survival’ R packages were used for propensity score matching and survival analyses, respectively. The Coxph and Surv(time, event) functions in the ‘survival’ package were used for the Cox hazard modelling.