This nationwide and population-based cohort study included all patients who underwent gastrectomy for gastric adenocarcinoma in Sweden from 2006 to 2015 and survived for at least 5 years. These participants were considered cured of the cancer. The follow-up started 5 years after surgery and ended at death or the end of the study period (December 31, 2020). Data sources were medical records and nationwide registers. The study was approved by the Ethical Review Board in Stockholm, Sweden (2017/141–31/2), and registered at clinicaltrials.gov (identification code NCT05540119).

The source cohort is the Swedish Gastric Cancer Surgery Study (SWEGASS), which includes at least 98% of all patients who underwent gastrectomy for gastric adenocarcinoma between January 1, 2006 and December 31, 2015 in Sweden. A detailed description of SWEGASS has been published previously [6]. In brief, patients with gastric adenocarcinoma who underwent gastrectomy were identified using well-validated nationwide Swedish registers for cancer (Swedish Cancer Registry [7]) and inpatient and outpatient care (Swedish Patient Register [8]). The final cohort was selected after a review of surgical charts, pathology reports, discharge summaries, and reports from multidisciplinary meetings. Patients in SWEGASS who survived for at least 5 years after gastrectomy were eligible for the present study. We excluded patients older than 78 years at the date of gastrectomy because it was very rare that these patients received neoadjuvant chemotherapy.

The study exposure was neoadjuvant chemotherapy. Patients were considered exposed regardless of whether they discontinued the treatment or reduced the dose due to side effects, corresponding to the intention-to-treat strategy. Patients who did not receive neoadjuvant chemotherapy were categorised as unexposed (reference group). Data on neoadjuvant chemotherapy were collected from the medical records.

The study outcome was death due to any cause, i.e. all-cause mortality. Mortality data came from the Swedish Cause of Death Register, which has 100% completeness for the date of death and > 96% completeness for causes of death, including deaths of Swedish citizens who die abroad [9].

Eight variables were considered potential confounders, which are defined and categorised as follows: age (at the start of follow-up, continuous variable), sex (male or female), comorbidity (0, I, or ≥ II, according to the most well-validated version of the Charlson comorbidity index scoring system [10, 11], not counting the gastric adenocarcinoma), education (≤ 9, 10–12, or ≥ 13 years of formal education), calendar year period (at the start of follow-up, 2011–2015 or 2016–2020), in-hospital postoperative complications (0, I–II, or III–IV, according to the Clavien–Dindo classification of surgical complications [12]), tumour sub-location (non-cardia or cardia) and splenectomy (during the gastrectomy, no or yes). Data sources for these variables were medical records, the Longitudinal Integration Database for Health Insurance and Labour Market [13], and the Swedish Patient Register [8].

Survival curves were depicted using the Kaplan–Meier estimator. Cox proportional hazards regression was used to calculate hazard ratios (HR) with 95% confidence intervals (CI). Four models were analysed: (1) a crude model without any adjustments, (2) a full model with adjustment for all eight potential confounders presented above (‘Confounders’), (3) a model with adjustment for age and comorbidity, and 4) a model with adjustment for six confounders excluding adjustment for age and comorbidity. To evaluate whether any associations between neoadjuvant chemotherapy and long-term mortality were modified by age, sex, comorbidity, calendar year and tumour sub-location, an interaction term was included in the main multivariable model one by one where HR were derived within each stratum. These variables were also categorised as stated above (‘Confounders’), except for age, which was categorised into two groups divided by the median. The proportional hazards assumption was evaluated by log–log survival plots and by calculating the correlations between Schoenfeld residuals for a particular covariate and ranking of individual failure time. The correlations were low and the statistical test was not statistically significant at a level of 0.05 for each variable, indicating that the proportional hazards assumption was met for all variables. We noted that the neoadjuvant group tended to be right censored earlier than the non-neoadjuvant group because patients receiving neoadjuvant therapy underwent gastrectomy in more recent calendar years. Thus, we assessed if the results were robust in a sensitivity analysis where the main analysis approach was replicated in a subgroup of patients with the possibility for a complete 5 year follow-up and where the outcome was 5 year survival. In a second sensitivity analysis, we excluded participants who, according to the Swedish Patient Register and the Cause of Death Register, had tumour recurrence and died due to gastric adenocarcinoma more than 5 years after gastrectomy. Partial missing data were low, no more than 4.1% for neoadjuvant chemotherapy status or any of the eight potential confounders combined. The analyses were, therefore, managed by complete case analysis, i.e. exclusion of patients without complete data on all variables included in the analysis. The data management and statistical analyses were conducted according to a pre-defined study protocol and were conducted by the first author (WL) under the supervision of a senior biostatistician (FM). The statistical software was SAS version 9.4 (SAS Institute Inc., Cary, NC, USA).