For the present analysis, version 3.2 of the StoP dataset was used, which includes 34 case–control or nested case–control studies totaling 13,121 GC cases and 31,420 controls. The StoP Project was established in 2012 as an international collaborative effort with the aim to elucidate determinants of risk and outcome of GC [17]. Data harmonization of core variables, such as those related to sociodemographic and some lifestyle factors, was carried out at the Coordinating Center in Milan, Italy after research teams shared the data from their previously conducted case–control or nested case–control studies.

Fourteen studies participating in the StoP Project had available information on dietary intake of vitamin C and total energy intake and were thus included in this analysis (5362 cases, 11,497 controls [the flowchart describing the inclusion of study participants is reported in Supplementary Fig. 1]). Among them, seven were from Europe (two from Italy [18, 19], two from Spain [20, 21], and one each from Greece [22], Russia [23], and Portugal [24]), three from Asia (one each from China [25], Japan [26], and Iran [27]), and four from America (one from the USA [28] and three from Mexico [29,30,31]). Of the studies included in the analysis, eight were community-based [19, 20, 24, 25, 27,28,29,30] and six were clinic-based [18, 21,22,23, 26, 31] (Supplementary Table 1).

The participating studies were conducted in accordance with applicable laws, regulations and guidelines for the protection of human subjects. Furthermore, the StoP Project was approved by the University of Milan Review Board (reference 19/15, April 1, 2015).

In this analysis, cases were individuals with histologically confirmed GC. Information about histological type (intestinal, diffuse, other, the latter including mixed, undifferentiated, and unclassified type) and subsite (cardia, noncardia) of GC was available for ten [18,19,20,21, 23, 24, 27,28,29, 31] and eleven [18,19,20,21,22,23,24, 26,27,28,29] studies, respectively.

Within each study, participants’ intake of vitamin C (mg/day) was computed from food-frequency questionnaires (FFQs) using country-specific food composition tables. Subsequently, based on the study-specific distribution of intake among controls, dietary vitamin C was categorized into quartiles. Intake of vitamin C from supplements was excluded from the analysis since information on supplements was available only for four studies (Russia [23], China [25], Portugal [24], and Japan [26]) among those with available data on dietary vitamin C intake, with variable degree of detail on the specific type of supplement taken and with very limited data on the dose.

In each study, structured questionnaires were used to collect information on participants’ sociodemographic and lifestyle characteristics.

We estimated odds ratios (ORs) and corresponding 95% confidence intervals (CIs) for the association between dietary intake of vitamin C and GC using mixed-effects logistic models, with a random intercept for each study. Model 1 was adjusted for sex (male, female), age (≤ 40 years, 41–50 years, 51–60 years, 61–70 years,  > 70 years), socioeconomic status (low, intermediate, or high, according to study-specific definitions based on education, income, or occupation), tobacco smoking status (never, former, current), alcohol drinking status (never, ever), and total energy intake (continuous). Model 2 was adjusted for the same set of covariates as model 1 and, additionally, for body mass index (BMI: underweight [< 18.5 kg/m2], normal weight [18.5–24.9 kg/m2], overweight [25–29.9 kg/m2], obese [≥ 30 kg/m2]) and intake of fruit and vegetables (low, intermediate, high, according to study-specific tertiles). Model 3 included the same covariates as model 2 and individuals in the highest and lowest 1% of vitamin C intake were excluded from the analysis. Model 4 was the same as model 2, except for total energy intake being modelled using a second-order fractional polynomial [32]. For all models, missing values of covariates were coded as a separate category for categorical variables, or replaced with the study-specific median values among controls together with dummy variables indicating replacements for continuous variables. Linear trends were assessed in similar models with intake of vitamin C in quartiles considered as a continuous variable. Additionally, for both the point estimate of the OR and the limit of its 95% CI closer to unity, we computed the E-value, which represents the minimum strength of association that an unmeasured confounder should have with both the exposure and the outcome to fully explain away the observed association between the exposure and the outcome, conditional on measured confounders [33].

Moreover, we stratified results according to the following: subsite (cardia, noncardia), histological type (intestinal, diffuse, other), type of study (clinic-based, community-based), sex (male, female), age (≤ 60 years,  > 60 years), socioeconomic status (low, intermediate, high), BMI (underweight, normal weight, overweight, obese), smoking status (never, former, current), drinking status (never, ever), H. pylori seropositivity (negative, positive), and intake of fruit and vegetables (low, intermediate, high, according to study-specific tertiles). Data about H. pylori seropositivity were available for seven studies, and in four of them, it was assessed through enzyme-linked immunosorbent assay (ELISA) [23, 26, 29, 31], in one using multiplex serology [20], and in two with both ELISA and Western blot testing [24, 27]. Furthermore, we performed a sensitivity analysis by comparing all cases to seropositive controls, assuming that H. pylori infection be a necessary cause of GC, regardless of results of serological tests [17, 34]. Stratified analyses were adjusted for sex, age, socioeconomic status, tobacco smoking status, alcohol drinking status, BMI, intake of fruit and vegetables, and total energy intake. We tested for heterogeneity across strata using likelihood ratio tests comparing the models including the interaction terms between quartiles of vitamin C intake and the stratification variable with those excluding them.

Furthermore, we evaluated potential additive interactions between dietary vitamin C (higher vs. lower than the study-specific median intake among controls) and the following variables: age (≤ 60 years,  > 60 years), sex (female, male), socioeconomic status (high, low/intermediate), tobacco smoking status (non-current, current), alcohol drinking status (never, ever), BMI (overweight/obese, normal weight), H. pylori seropositivity (negative, positive), intake of fruit and vegetables (high, intermediate/low), and intake of salt (low, intermediate/high). To this aim, we estimated the relative excess risk due to interaction (RERI) and the attributable proportion (AP) due to interaction and computed their 95% CI using the delta method [35, 36], after recoding variables of interest to use the stratum with the lowest OR of GC when both factors were considered jointly as the reference category [37].

Additionally, we assessed the dose–response relationship between dietary intake of vitamin C and GC with similar logistic mixed-effects models. We assessed its linearity by treating dietary vitamin C as a continuous variable in the model, and nonlinearity with first- and second-order fractional polynomials. We ran two different models adjusted for the same covariates as model 1 and model 2 described above for the main analysis, and in both cases we excluded from the analysis individuals in the lowest and in the highest 1% of intake of vitamin C. The model considered as the best fitting was the one with the lowest deviance [32].

All analyses were carried out using Stata software version 17 (StataCorp LLC. College Station. TX).