Patients 18 years of age or older diagnosed with a single primary cancer at the following sites were included: anus, anal canal and anorectal, appendix, rectosigmoid junction, rectum, sigmoid colon, descending colon, splenic flexure, hepatic flexure, transverse colon, ascending colon, cecum, esophagus, stomach, pancreas, kidney and renal pelvis, larynx, liver, lung, main bronchus, melanoma of the skin, brain, thyroid, trachea, and urinary bladder.

Patients were excluded according to the following criteria: no information on the reason for the absence of surgery; diagnosis only at autopsy or death certificate; follow-up time <1 day; stage 0 or occult cancer; lack of information on the exact site of cancer; and absence of sociodemographic information on race, median household income, type of housing region in the rural-urban continuum.

The rationale for the selection of patients for each analysis is depicted in Supplementary Figure 1. The syntax used in the SEERStat software for patient selection is found in the Supplementary material.

The cancer staging systems used were the sixth and seventh editions of the AJCC for patients diagnosed between 2004–2009 and 2010–2015, respectively. The race identification used was restricted to Black, White and Other (American Indian/AK Native, Asian/Pacific Islander). The median household income adjusted for accumulated inflation for the year 2019 was used.

The staging, regardless of the edition of the AJCC system, has been reclassified as I, II, III, and IV. Unassigned or missing data regarding stage and histological grade were classified as NA (Not applied/not assigned); this category was not excluded from survival analyses or propensity score matching.

For survival, histologies (ICD-O-3) were grouped into an up to seven-category variable according to their relative frequency, in a decrease manner, for each site. An eighth variable was created grouping all remaining histologies. The codes of the first two categories are described in Supplementary Table 1.

For propensity score matching (PSM), histologies (ICD-O-3) were grouped into a three-category variable according to their relative frequency: most common, less common and rare. The codes of the first two categories are described in Supplementary Table 2, and the third category is composed of other histologies not described.

Statistical analyzes of normality (Kolmogorov-Smirnov), descriptive, crosstab (Pearson's χ² and Cramer V-test), multinomial logistic regression, survival and PSM were performed on IBM SPSS v25.0. Kaplan-Meier plots and survival tables were performed on JAMOVI 2.2.5. The graphs representing the hazard ratios with a confidence interval of 95% were built in MS Excel. When appropriate, the significance level (α) was 0.05.

The Kaplan-Meier (KM) estimator was used to analyze the proportionality of risks based on curves and the Log-Rank test, as well as median survival, percentages, and censoring distributions. Only variables with proportionality of risks were included in Cox regression analysis. The proportionality of risks for continuous variables was evaluated by the correlation between the partial residues generated by the univariable Cox regression with the observation time of the analyzed outcome; the time dependence was assumed in the presence of correlation between these variables, visually analyzed by the scatter plot. In the case of the year variable, there was a correlation; its discrete nature impedes covariation by time (time-dependent Cox regression), and its categorized form was used instead of the quantitative form. Time dependence was also observed as a function of age, opting for a multicategorical variable instead of the continuous one.

Cox regression models were elaborated containing all the variables obtained (sex, year of diagnosis, age-14 categories -, race, median household income, rural-urban continuum, the reason for no cancer-direct surgery, histological grade, staging, cancer site, histological type-−8-category variable), except for the subgroup division variable, when applicable. Two-way and three-way interactions were included in the models, if the number of instances (number of patients) allowed for the complexity of the model (number of degrees of freedom). Two-way interactions between cancer site and histology and cancer site and age were entered into analyzes of the entire cohort and by sex, respectively. Three-way interactions between sex, age, and sites, and between sex, age, and histology were performed in full-cohort and site-wide analyses, respectively. In site and cancer analyses, it was possible to include the three-way interaction just before pairing. In analyses by year of diagnosis, it was possible to only the two-way interaction between site and histology; the other interactions were included only in the analysis of periods (2004–2009 and 2010–2015). In cases where it was not possible to include the three-way interaction, it was replaced by a respective two-way interaction that excludes age.

Significance level adjustment by the Bonferroni method was applied only to the variables of interest after the independence analysis in order interpret the prognostic value of these factors. Adjustment was performed by dividing the significance value by the number of possible pairs between variables included in each model. Unadjusted p-values were presented throughout the manuscript, except on express occasions.

For Cox regression analysis of sites with reduced sample number, it was used the bootstrapping technique with Bias Correction accelerated method to corroborate the findings; 1,000 resamples were performed. The interpretation of bootstrapping results was based on the analysis of confidence intervals, p-value, and bias for robust conclusions. Bootstrap p-values were reported in these cases only for multivariable models.

The directionality and location of the associations were evaluated by the analysis of standardized adjusted residuals. A Bonferroni correction was applied to adjust the level of significance to the amount of analysis by the interaction between rows and columns to obtain the significant associations. The interpretation of the effect size of the associations was adjusted by the degrees of freedom.

For logistic regression with the stage as a dependent, the parallel lines test was initially performed to assess the assumption of an ordinal regression. Due to the rejection of the null hypothesis of the parallel lines test (p < 0.05), a multinomial regression was performed. The proportionality/linearity with logits was analyzed by comparing the differences in the coefficients (βs) of the categories of the categorical variables and by the Box-Tidwell transformation for continuous variables. Continuous variables that break the assumption of significance in the insertion of the transformed variable [ln (variable)^*variable] were categorized. Categories (levels) with overlapping confidence intervals or that break proportionality were collapsed with similar ones.

For PSM, the SPSS tool was used. Briefly, a 1:1 pairing without substitution was performed to match men and women (indicator group) according to the variables age, year of diagnosis (as categorical), race, median household income, rural-urban continuum classification, stage, histological type, cancer site, reason no cancer-directed surgery, and the AJCC edition used for staging. Additionally, the age variable of 14 categories and three-level histology variable were added to improve the matching. A tolerance of 0.000001 was established for matching based on the score generated by the matching logistic regression model. For Cox regression analysis after pairing, two models were developed: one covariate sex with the other variables, as abovementioned, and the other covariate sex with the score generated by the logistic regression of pairing.

A total of 1,274,118 patients were selected, with a median age of 65 years (18–85+) and a median survival of 26 months (0–191). Of the entire cohort, 831,871 died, with 672,329 (80.8%) deaths attributed to cancer. Patients' characteristics are described separately by sexes in Table 1, which shows a considerable difference in the number of deaths and median survival between men and women.

Analyzes were performed to observe patterns and trends that may explain differences in survival between sexes, before performing survival analysis. Table 1 describes different frequencies of age groups, cancer sites, staging groups, and surgery decisions by sex. The significance of the observed difference was tested by contingency tables (Pearson's χ² test) associated with the study of effect size (Cramer's V-test).

There was a significant association between the sex and age clusters (Pearson's χ²:23,147.50; Cramer's V₍₁₎:0.135; p < 0.0005), cancer site (Pearson's χ²:76,809.54; Cramer's V₍₁₎:0.246; p < 0.0005), and with the recommendation and performance of surgery according to sex (Pearson's χ²:4,149.46; Cramer's V₍₁₎:0.057; p < 0.0005). There were also significant associations between sex and histology per site, and a positive association of both sexes with some histology per cancer site. Moreover, associations between age clusters and cancer sites were observed per sex. The Supplementary results describes the details and association of age cluster and sex by cancer site.

Regarding stage, although the female sex was positively associated with stage I cancer (std. res.: 68.7) and the male sex mainly with stage IV cancer (std. res.: 45.0), this association was weak (Pearson's χ²:5,128.74, p < 0.0005) with a negligible effect size (Cramer's V₍₁₎:0.063, p < 0.0005).

However, other associations are not directly perceived with sex, which could be potential confounding effects. While no expressive numeric differences were observed between males (Pearson's χ²:403,888.37; Cramer's V₍₄₎:0.383; p < 0.0005) and females (Pearson's χ²:385,634.62; Cramer's V₍₄₎:0.406; p < 0.0005), there was a very strong association between site and stage (Pearson's χ²:788,214.79, p < 0.0005) with a large effect size (Cramer's V₍₄₎:0.393, p < 0.0005) (Supplementary Table 3), which prevailed even when excluding brain cancer in the analysis (Pearson's χ²:439,064.97; Cramer's V₍₃₎:0.298; p < 0.0005).

Despite the differences described above, it is possible to observe a consistent increase in cancer diagnosis in earlier stages (Figure 1A). Additionally, an increase in the performance of surgeries, with a reduction of no recommendations, refusals, and lack of surgery performance, was observed over time (Figure 1B). There is also a reduction in the prevalence of respiratory, brain, thyroid, and stomach cancers, with a compensatory increase in melanomas of the skin, liver, and urinary system (Figures 1C, D). There is no apparent difference in trends by sex.

Multinomial logistic regression models were used to confirm the impact of sex and year on staging and surgery performance.

Using the “surgery performed” classification as a reference and adjusting for age, year of diagnosis, race, household income, rural-urban continuum, site of cancer, stage, and histological grade, men are more likely to die before surgery recommendation [OR: 1.342 (1.232–1.462), p < 0.0005], being contraindicated for surgery [OR: 1.188 (1.159–1.218), p < 0.0005], and being recommended but not performing surgery [OR: 1.291 (1.252–1.332), p < 0.0005], but are less likely to refuse surgery [OR: 0.949 (0.917–0.982), p = 0.003], compared to females.

For analysis of stage predictive classification, only patients with stage were included (n = 1,134,221). Using the “stage I” classification as a reference and adjusting for age, year of diagnosis, race, household income, rural-urban continuum, cancer site, and histological grade, men are more likely to be classified as “stage II” [OR: 1.173 (1.158–1.189), p < 0.0005], III [OR: 1.226 (1.211–1.240), p < 0.0005], and IV [OR: 1.283 (1.269–1.298), p < 0.0005], compared to females.

The association of cancer site with stage, in addition to the association of sex with age and cancer site, the excess of stage IV in men and stages trends over time can generate highly biased results in survival analyses. For computational limitations, the PSM was performed separately by tumor sites and then merged after matching. Due to the small number of patients with tracheal cancer, there was no matching. Cancer-specific survival was used instead of overall survival to suppress the longer life expectancy bias of women compared to men (9).

After matching, there was no further association between sex and site (Pearson's χ²: 0; p = 1.000; Cramer's V₍₁₎: 0.000), but other associations, although negligible, still exist (Supplementary results). Importantly, the difference in survival by site of the cancer remained practically the same after pairing as before pairing (Supplementary Table 4 and Supplementary Figure 2).

By the Kaplan-Meier and Log-Rank comparison (χ²: 8,335.54; p < 0.0005) it was possible to observe significantly higher survival of female patients in the entire cohort (Figure 2A) and in almost all cancer sites (Supplementary Table 5). Female patients had a median survival of 74 months, compared to 30 months (29.67–30.33) of the male counterpart, which had an excess of 7.5% of deaths in relation to women (female: 48.7%; male: 56.2%). This difference remained after adjustment by other variables in multivariable Cox regression [adjusted HR: 0.876 (0.866–0.886), p < 0.0005] (Figure 2B). The 1-, 3-, 5-year, and longer survival tables, with the number at risk, are described in Supplementary Table 6.

After matching by propensity score (n = 441,916), this difference in median survival of 44 months was reduced to only 6 months [female: 29 months (28.38–29.62); male: 23 months (22.60–3.40)], although this difference was still significant (Log-Rank χ²: 431.13; p < 0.0005) (Figure 2C). The excess death in men compared to women dropped to 2.3% (female: 56.1%; male: 58.4%). Again, this difference remained after adjustment by other variables in multivariable Cox regression [adjusted HR: 0.888 (0.864–0.912), p < 0.0005] (Figure 2C). The 1-, 3-, 5-year and longer survival tables, with the number at risk, are described in Supplementary Table 7.

The site was analyzed separately because the association of sex with survival may vary depending on the type of cancer. The effect of sex in multivariable analyzes before and after matching lost its significance in the p-value adjustment for some situations (Figure 3A; Supplementary Tables 8, 9).

A consistent increase in cancer-specific survival was observed over time for both sexes (Figures 3B, C; Supplementary Tables 10, 11), with no difference in the trend by sex. Despite the absence of a difference in the association between sex and survival in some years, in general, a positive association between women and survival was observed, including analyzes by periods (2004–2009 and 2010–2015) (Figure 3D; Supplementary Tables 12, 13), which could imply some bias due to the change in the staging system. KM and adjusted Cox regression curves (Supplementary Figures 3, 4), as well as 1-, 3-, 5-year and longer survival tables, with number at risk, are described in Supplementary Tables 14, 15.