Stage-dependent prognostic impact of age in colorectal cancer: A population-based SEER analysis Authors: Burak Pacaci, Ibrahim Vedat Bayoglu Published: 2026 ABSTRACT AbstractThe prognostic association between age and colorectal cancer (CRC) survival may vary by stage. We analyzed adults with microscopically confirmed CRC diagnosed during 2004–2022 in the Surveillance, Epidemiology, and End Results database after excluding appendiceal primaries, death certificate-only/autopsy-only cases, and non-microscopically confirmed cases. Multivariable Cox models assessed age continuously, and age-by-stage interaction models compared patients aged 18–49 versus ≥ 50 years among those with known stage. The primary survival cohort included 454,216 patients; multivariable Cox models were fitted in 452,739 patients. Increasing age was associated with worse cancer-specific survival (CSS; HR 1.0279/year, 95% CI 1.0275–1.0284) and overall survival (OS; HR 1.0443/year, 95% CI 1.0440–1.0447). Age-by-stage interaction was significant for both endpoints. For CSS, the association of age ≥ 50 years was strongest in localized disease (HR 2.17, 95% CI 2.04–2.31) and attenuated in regional (HR 1.58, 95% CI 1.54–1.63) and distant disease (HR 1.41, 95% CI 1.38–1.44), with a similar pattern for OS. Older age was associated with worse CRC survival, but the magnitude of this association was stage-dependent.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57998-x.Keywords: Colorectal cancer, SEER, Cancer-specific survivalSubject terms: Cancer, Gastroenterology, Oncology, Risk factors Introduction Colorectal cancer (CRC) remains a major global health burden and is among the most commonly diagnosed cancers and leading causes of cancer death worldwide1. Tumor stage at diagnosis is the dominant determinant of prognosis; however, patient-level factors, including age, also influence outcomes and treatment delivery2. Age has traditionally been regarded as an adverse prognostic factor in CRC. Older adults often have a higher comorbidity burden and reduced physiologic reserve, which may limit tolerance of surgery and systemic therapy. In parallel, early-onset colorectal cancer (EOCRC; diagnosis before age 50 years) has become an increasing public health concern, with rising incidence reported across multiple regions3,4. EOCRC may also differ in clinical presentation and tumor biology compared with later-onset disease5,6. These epidemiologic shifts have contributed to recommendations to begin average-risk CRC screening at age 45 years in the United States7,8. However, whether age itself independently affects CRC prognosis remains debated. For EOCRC, prior registry studies and meta-analyses suggest that stage-adjusted cancer-specific survival is often similar to later-onset disease, although results vary across datasets and endpoints9–11. Many prior analyses have treated age as a single, uniform prognostic factor and have not formally tested whether its effect differs by stage9–11. If the association between age and survival varies from localized to metastatic disease, this has implications for counseling, risk modeling, and study design. To address these gaps, we conducted a large population-based analysis using the Surveillance, Epidemiology, and End Results (SEER) program. We aimed to (1) evaluate the association between age and cancer-specific survival (CSS) and overall survival (OS), (2) formally test whether the association between age and survival differs by stage, and (3) evaluate robustness through sensitivity analyses, including alternative handling of unknown cause of death (COD), adenocarcinoma-only, treatment-adjusted, three-category age, and colon-only analyses. Methods Data source and study population This retrospective population-based cohort study was conducted using data from the SEER program of the National Cancer Institute. We used Incidence - SEER Research Data, 17 Registries, Nov 2024 Sub (2000–2022), which covers approximately 26.5% of the U.S. population based on the 2020 census12. Data were extracted using SEER*Stat software (seer.cancer.gov/seerstat). Although this SEER Research Data product spans 2000–2022, we restricted the analytic cohort to diagnoses from 2004 onward to ensure consistent staging using the Combined Summary Stage (2004+) variable with expanded regional codes. This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cohort studies (Supplementary Table S10). Patients diagnosed with primary CRC between 2004 and 2022 were initially identified using International Classification of Diseases for Oncology, Third Edition (ICD-O-3) site codes corresponding to colon (C18.0–C18.9), rectosigmoid junction (C19.9), and rectum (C20.9). Cases were restricted to those with malignant behavior, age 18 years or older, and a single primary cancer (SEER sequence number: one primary only). Appendiceal primaries (C18.1) were excluded because appendiceal cancers are biologically and clinically distinct from conventional colorectal cancers. We also excluded death certificate-only/autopsy-only cases using the SEER Type of Reporting Source variable and excluded non-microscopically confirmed cases using the Diagnostic Confirmation variable. Primary site was grouped as right colon (C18.0, C18.2–C18.4), left colon (C18.5–C18.7), rectosigmoid junction (C19.9), rectum (C20.9), or other colorectal sites (C18.8–C18.9). Tumor grade was harmonized across diagnosis years and categorized as I, II, III/IV, or Unknown. Patients were categorized into two predefined age groups at diagnosis: 18–49 years and ≥ 50 years. This cutoff was selected because it corresponds to the conventional EOCRC threshold and aligns with screening-related epidemiologic literature. Stage was categorized according to SEER Summary Stage as Localized, Regional, Distant, or Unknown. Unknown stage was retained as a separate category in primary multivariable models, whereas interaction and stage-stratified analyses were restricted to known stage (Localized/Regional/Distant). Race/ethnicity categories were defined according to SEER classifications. Primary site surgery was defined from the RX Summ–Surg Prim Site (1998+) variable as codes 10–80 or 90 (yes) versus 00 (no). Records coded as unknown surgery (99) or not applicable (98) were excluded from multivariable models because these values reflect indeterminate or inapplicable surgical status rather than a distinct clinical category. Conversely, “Unknown” levels for other covariates, such as grade or stage, were retained as explicit categories to minimize selection bias. Chemotherapy and radiation recodes were extracted for descriptive and exploratory sensitivity analyses but were not included in the primary models because SEER treatment variables include “No/Unknown” categories and lack details on treatment intent, regimen, sequencing, dose intensity, and completion. Year of diagnosis was categorized into three eras (2004–2009, 2010–2015, and 2016–2022) for descriptive purposes and modeled as a continuous covariate per year since 2004 in multivariable Cox models. SEER does not provide detailed information on comorbidity, frailty, performance status, geriatric vulnerability, gastrointestinal comorbidity, family history, genetic susceptibility, molecular features, recurrence, treatment intensity, or treatment completion. These unavailable variables were therefore not included in the models and were considered when interpreting residual confounding. Because SEER data are de-identified and publicly available, institutional review board approval was not required. Outcomes The primary outcome was CSS, defined as time from diagnosis to death attributable to colorectal cancer using the SEER cause-specific death classification13,14. Deaths from other causes were censored at the time of death. In the primary CSS analysis, deaths with missing/unknown COD were treated as censored at death; a sensitivity analysis excluded these records. The secondary outcome was OS, defined as time from diagnosis to death from any cause. Survival time was measured in months. Records with missing or unknown survival months were excluded. Statistical analysis Baseline characteristics were summarized overall and by age group. Age was reported as median with interquartile range and compared using the Mann–Whitney U test. Categorical variables were summarized as counts and percentages and compared using Pearson chi-square tests. Standardized mean differences were used to quantify overall imbalance between age groups; for categorical variables with more than two levels, the maximum level-specific standardized mean difference was reported. Kaplan–Meier survival curves were generated to estimate stage-specific CSS and OS stratified by age group (18–49 vs. ≥ 50 years). Five-year CSS and OS estimates with 95% confidence intervals were also calculated by stage and age group. Multivariable Cox proportional hazards regression models were constructed to evaluate associations with CSS and OS, adjusting for age, sex, race/ethnicity, primary site group, stage, grade, surgery, and diagnosis year. Age was modeled continuously in the primary multivariable Cox models. Diagnosis year was modeled continuously as years since 2004. Multicollinearity was assessed using variance inflation factors, and no significant collinearity was detected. The proportional hazards assumption was evaluated using Schoenfeld residual tests and graphical diagnostics (Supplementary Figure S1). Given the very large sample size, formal proportional hazards tests were interpreted in conjunction with visual assessment. Because stage and surgery demonstrated time-varying effects, piecewise Cox models were performed as sensitivity analyses by splitting follow-up at 12 and 60 months (0–12, 12–60, and > 60 months) and estimating interval-specific effects with robust standard errors clustered by individual. For interval splitting only, records with 0-month survival were assigned 0.001 months for computational purposes. Age-by-stage interaction models were fitted for CSS and OS to formally test effect modification by stage. These models included age group, stage, and age-by-stage interaction terms and were adjusted for sex, race/ethnicity, primary site group, grade, surgery, and diagnosis year. Interaction analyses were restricted to patients with known stage. Stage-specific age effects were derived from the interaction models as linear combinations of the main age effect and interaction terms. Overall interaction was evaluated using likelihood ratio tests comparing models with and without interaction terms. Sensitivity analyses included alternative handling of missing/unknown COD for CSS, adenocarcinoma-only analysis, exploratory treatment-adjusted models including SEER chemotherapy and radiation recodes, three-category age analysis (18–49, 50–64, and ≥ 65 years), and colon-only analysis excluding rectosigmoid and rectal primaries. The adenocarcinoma-only cohort was defined using ICD-O-3 histology codes 8140–8147, 8210–8211, 8220–8221, 8260–8263, 8480–8481, and 8490. All analyses were performed using R software (R Foundation for Statistical Computing, Vienna, Austria). Two-sided p values < 0.05 were considered statistically significant. Results Patient characteristics Of 488,919 initially eligible adults with colorectal cancer, 15,330 appendiceal primaries were excluded. We further excluded 6,192 death certificate-only/autopsy-only cases and 13,181 non-microscopically confirmed cases, yielding a primary survival cohort of 454,216 patients (Fig. 1). No additional records were excluded for missing or unknown survival months after these exclusions. For the primary multivariable Cox models, 1,477 records with unknown or not applicable primary-site surgery codes were excluded, resulting in a multivariable analytic cohort of 452,739 patients. Interaction analyses were further restricted to patients with known stage, resulting in a known-stage interaction cohort of 434,251 patients. Study flow diagram. Selection of the colorectal cancer cohort from the SEER database, detailing exclusions and the creation of the primary survival cohort, multivariable analytic cohort, and known-stage interaction cohort. Median follow-up was 101 months using the reverse Kaplan–Meier method. Baseline and treatment characteristics stratified by age group are presented in Table 1. Overall, 61,274 patients (13.5%) were aged 18–49 years and 392,942 patients (86.5%) were aged ≥ 50 years. Compared with patients aged ≥ 50 years, those aged 18–49 years more frequently had rectal primaries, distant-stage disease, chemotherapy recorded as “Yes,” and radiation recorded as “Yes.” Patients aged ≥ 50 years more frequently had right-sided colon primaries and localized-stage disease. Table 1Baseline and treatment characteristics of the study population stratified by age group.CharacteristicLevelOverall18–49≥ 50P valueSMDN454,21661,274392,942Age at diagnosis, yearsMedian (IQR)65 (55–76)44 (39–47)68 (59–77)< 0.001SexFemale217,198 (47.8%)28,267 (46.1%)188,931 (48.1%)< 0.0010.039Male237,018 (52.2%)33,007 (53.9%)204,011 (51.9%)Race/ethnicityNH White290,184 (63.9%)32,877 (53.7%)257,307 (65.5%)< 0.0010.243NH Black52,766 (11.6%)8,410 (13.7%)44,356 (11.3%)Hispanic (All Races)60,452 (13.3%)12,146 (19.8%)48,306 (12.3%)NH Asian or Pacific Islander43,218 (9.5%)6,504 (10.6%)36,714 (9.3%)NH American Indian/Alaska Native3,732 (0.8%)672 (1.1%)3,060 (0.8%)NH Unknown Race3,864 (0.9%)665 (1.1%)3,199 (0.8%)Primary site groupRight colon168,219 (37.0%)13,862 (22.6%)154,357 (39.3%)< 0.0010.366Left colon121,158 (26.7%)18,817 (30.7%)102,341 (26.0%)Rectosigmoid35,454 (7.8%)5,921 (9.7%)29,533 (7.5%)Rectum113,897 (25.1%)20,956 (34.2%)92,941 (23.7%)Other15,488 (3.4%)1,718 (2.8%)13,770 (3.5%)Summary stageLocalized166,666 (36.7%)18,608 (30.4%)148,058 (37.7%)< 0.0010.155Regional165,270 (36.4%)23,928 (39.1%)141,342 (36.0%)Distant102,982 (22.7%)16,774 (27.4%)86,208 (21.9%)Unknown19,298 (4.2%)1,964 (3.2%)17,334 (4.4%)GradeI44,104 (9.7%)6,096 (9.9%)38,008 (9.7%)< 0.0010.027II267,699 (58.9%)35,418 (57.8%)232,281 (59.1%)III/IV70,670 (15.6%)9,557 (15.6%)61,113 (15.6%)Unknown71,743 (15.8%)10,203 (16.7%)61,540 (15.7%)Primary site surgeryNo86,435 (19.0%)10,959 (17.9%)75,476 (19.2%)< 0.0010.035Yes366,304 (80.6%)50,147 (81.8%)316,157 (80.5%)Unknown/not applicable1,477 (0.3%)168 (0.3%)1,309 (0.3%)ChemotherapyNo/Unknown268,560 (59.1%)23,721 (38.7%)244,839 (62.3%)< 0.0010.486Yes185,656 (40.9%)37,553 (61.3%)148,103 (37.7%)RadiationNo/Unknown387,166 (85.2%)47,542 (77.6%)339,624 (86.4%)< 0.0010.232Yes67,050 (14.8%)13,732 (22.4%)53,318 (13.6%)Year of diagnosis2004–2009139,311 (30.7%)17,220 (28.1%)122,091 (31.1%)< 0.0010.0712010–2015139,451 (30.7%)18,554 (30.3%)120,897 (30.8%)2016–2022175,454 (38.6%)25,500 (41.6%)149,954 (38.2%)Values are presented as n (%) unless otherwise indicated. Age is presented as median (IQR) and compared using the Mann–Whitney U test; categorical variables were compared using Pearson chi-square tests. Standardized mean differences (SMDs) are shown for overall group imbalance; for categorical variables with more than two levels, the maximum level-specific SMD is reported. Chemotherapy and radiation are SEER treatment recodes and should be interpreted descriptively because “No/Unknown” categories combine absence of treatment with incomplete treatment ascertainment. Age SMD is not shown because age groups are defined by age.NH, non-Hispanic; SEER, Surveillance, Epidemiology, and End Results; SMD, standardized mean difference. Baseline and treatment characteristics of the study population stratified by age group. Values are presented as n (%) unless otherwise indicated. Age is presented as median (IQR) and compared using the Mann–Whitney U test; categorical variables were compared using Pearson chi-square tests. Standardized mean differences (SMDs) are shown for overall group imbalance; for categorical variables with more than two levels, the maximum level-specific SMD is reported. Chemotherapy and radiation are SEER treatment recodes and should be interpreted descriptively because “No/Unknown” categories combine absence of treatment with incomplete treatment ascertainment. Age SMD is not shown because age groups are defined by age. NH, non-Hispanic; SEER, Surveillance, Epidemiology, and End Results; SMD, standardized mean difference. CSS and OS analyses In multivariable Cox models adjusted for age, sex, race/ethnicity, primary site group, stage, grade, surgery, and diagnosis year, increasing age was associated with worse CSS (HR 1.0279 per year, 95% CI 1.0275–1.0284; p < 0.001) and OS (HR 1.0443 per year, 95% CI 1.0440–1.0447; p < 0.001) (Table 2). Stage was the strongest prognostic factor in both models. Compared with localized disease, distant-stage disease was associated with markedly worse CSS (HR 13.04, 95% CI 12.81–13.28; p < 0.001) and OS (HR 5.86, 95% CI 5.79–5.93; p < 0.001). Primary site surgery was associated with lower hazards of both CRC-specific death and all-cause death, although this estimate should be interpreted as an adjusted association rather than a causal treatment effect. Table 2Multivariable Cox Regression Models for CSS and OS.VariableCSS HR (95% CI)CSS P valueOS HR (95% CI)OS P valueAge (per 1-year)1.0279 (1.0275–1.0284)< 0.0011.0443 (1.0440–1.0447)< 0.001Male (vs. Female)1.08 (1.07–1.09)< 0.0011.16 (1.16–1.17)< 0.001Race/ethnicity: Hispanic (vs. NH White)1.01 (0.99–1.02)0.4490.99 (0.98-1.00)0.125Race/ethnicity: NH AI/AN (vs. NH White)1.19 (1.13–1.26)< 0.0011.17 (1.12–1.23)< 0.001Race/ethnicity: NH Asian/PI (vs. NH White)0.91 (0.89–0.92)< 0.0010.84 (0.83–0.86)< 0.001Race/ethnicity: NH Black (vs. NH White)1.22 (1.20–1.24)< 0.0011.21 (1.20–1.23)< 0.001Race/ethnicity: NH Unknown (vs. NH White)0.20 (0.17–0.23)< 0.0010.24 (0.22–0.27)< 0.001Site: Left colon (vs. Right colon)0.92 (0.91–0.94)< 0.0010.93 (0.92–0.94)< 0.001Site: Rectosigmoid (vs. Right colon)0.93 (0.92–0.95)< 0.0010.92 (0.91–0.94)< 0.001Site: Rectum (vs. Right colon)0.85 (0.83–0.86)< 0.0010.82 (0.81–0.83)< 0.001Site: Other (vs. Right colon)1.23 (1.20–1.26)< 0.0011.17 (1.15–1.20)< 0.001Stage: Regional (vs. Localized)3.02 (2.96–3.07)< 0.0011.64 (1.62–1.66)< 0.001Stage: Distant (vs. Localized)13.04 (12.81–13.28)< 0.0015.86 (5.79–5.93)< 0.001Stage: Unknown (vs. Localized)3.62 (3.52–3.73)< 0.0011.94 (1.89–1.98)< 0.001Grade: I (vs. II)0.78 (0.76–0.80)< 0.0010.86 (0.84–0.87)< 0.001Grade: III/IV (vs. II)1.54 (1.52–1.56)< 0.0011.38 (1.36–1.39)< 0.001Grade: Unknown (vs. II)1.01 (0.99–1.03)0.1920.96 (0.95–0.98)< 0.001Surgery: Yes (vs. No)0.37 (0.36–0.37)< 0.0010.39 (0.38–0.39)< 0.001Diagnosis year (per year since 2004)0.978 (0.977–0.979)< 0.0010.983 (0.982–0.984)< 0.001Models were fit in 452,739 patients (OS events = 227,925; CRC-specific deaths = 150,905).CSS, cancer-specific survival; OS, overall survival; HR, hazard ratio; CI, confidence interval; NH, non-Hispanic; AI/AN, American Indian/Alaska Native; PI, Pacific Islander. Multivariable Cox Regression Models for CSS and OS. Models were fit in 452,739 patients (OS events = 227,925; CRC-specific deaths = 150,905). CSS, cancer-specific survival; OS, overall survival; HR, hazard ratio; CI, confidence interval; NH, non-Hispanic; AI/AN, American Indian/Alaska Native; PI, Pacific Islander. Stage and surgery showed evidence of time-varying effects in proportional hazards diagnostics; therefore, piecewise Cox analyses were performed and are provided in Supplementary Tables S1–S3. Piecewise CSS estimates were materially unchanged when deaths with missing/unknown COD were excluded rather than censored (Supplementary Tables S2–S3). Stage-stratified Kaplan–Meier curves for CSS and OS according to age group are presented in Figs. 2 and 3, respectively. Stage-stratified cancer-specific survival curves according to age group. Panels A-C correspond to localized, regional, and distant disease, respectively. Risk tables show numbers at risk at 36-month intervals. Non-CRC deaths were censored for CSS. Stage-stratified overall survival curves according to age group. Panels A-C correspond to localized, regional, and distant disease, respectively. Risk tables show numbers at risk at 36-month intervals. Age-by-stage interaction Age-by-stage interaction was evaluated for both CSS and OS among patients with known stage (Supplementary Table S4). For CSS, adding age-by-stage interaction terms improved model fit (likelihood ratio test: χ²=198.41, df = 2, p < 0.001), and both interaction terms (Age ≥ 50 × Regional and Age ≥ 50 × Distant) were statistically significant (both p < 0.001). For OS, adding interaction terms also improved model fit (likelihood ratio test: χ²=2237.7, df = 2, p < 0.001), and both interaction terms were statistically significant (both p < 0.001). Derived stage-specific estimates showed that age ≥ 50 years was associated with worse outcomes, with attenuation across stages for both CSS and OS. For CSS, the association of age ≥ 50 years was strongest in localized disease (HR 2.17, 95% CI 2.04–2.31; p < 0.001), attenuated in regional disease (HR 1.58, 95% CI 1.54–1.63; p < 0.001), and was smallest in distant disease (HR 1.41, 95% CI 1.38–1.44; p < 0.001). OS estimates showed a similar pattern, decreasing from localized disease (HR 4.19, 95% CI 3.99–4.39; p < 0.001) to regional disease (HR 2.30, 95% CI 2.24–2.36; p < 0.001) and distant disease (HR 1.46, 95% CI 1.43–1.49; p < 0.001) (Table 3). Table 3Effect of Age on CSS and OS Within Each Stage.StageAge comparisonCSS HR (95% CI)P valueOS HR (95% CI)P valueLocalized≥ 50 vs. 18–492.17 (2.04–2.31)< 0.0014.19 (3.99–4.39)< 0.001Regional≥ 50 vs. 18–491.58 (1.54–1.63)< 0.0012.30 (2.24–2.36)< 0.001Distant≥ 50 vs. 18–491.41 (1.38–1.44)< 0.0011.46 (1.43–1.49)< 0.001Interaction models were restricted to patients with known stage (n = 434,251; OS events = 216,893; CRC-specific deaths = 143,423). Stage-specific age effects were derived from fully adjusted age-by-stage interaction models. Models were adjusted for sex, race/ethnicity, primary site group, grade, surgery, and diagnosis year.CSS, cancer-specific survival; OS, overall survival; HR, hazard ratio; CI, confidence interval. Effect of Age on CSS and OS Within Each Stage. Interaction models were restricted to patients with known stage (n = 434,251; OS events = 216,893; CRC-specific deaths = 143,423). Stage-specific age effects were derived from fully adjusted age-by-stage interaction models. Models were adjusted for sex, race/ethnicity, primary site group, grade, surgery, and diagnosis year. CSS, cancer-specific survival; OS, overall survival; HR, hazard ratio; CI, confidence interval. Additional sensitivity analyses PH diagnostics and time-split model results are provided in Supplementary Figure S1 and Supplementary Tables S1–S3. A forest plot summarizing multivariable associations with CSS and OS is provided in Supplementary Figure S2. In an exploratory treatment-adjusted sensitivity model additionally including SEER chemotherapy and radiation recodes, the association between increasing age and worse outcomes persisted, although the magnitude was modestly attenuated (CSS HR 1.0240 per year, 95% CI 1.0236–1.0245; OS HR 1.0403 per year, 95% CI 1.0400–1.0407; both p < 0.001) (Supplementary Table S5). Because SEER treatment variables are limited and include “No/Unknown” categories, these estimates were interpreted as sensitivity analyses rather than causal treatment-effect estimates. To address whether the primary age grouping obscured heterogeneity within later-onset CRC, we performed an additional sensitivity analysis using three age categories (18–49, 50–64, and ≥ 65 years). The age-by-stage interaction remained statistically significant for both CSS (LRT χ²=1031.0, df = 4, p < 0.001) and OS (LRT χ²=7121.8, df = 4, p < 0.001). Within each stage, both later-onset age strata had worse outcomes than the 18–49 group, with the largest relative hazards observed among patients aged ≥ 65 years and in localized disease (Supplementary Table S6). Given the distinct treatment pathways and recurrence patterns of rectal cancer, we also performed a colon-only sensitivity analysis excluding rectosigmoid and rectal primaries. In this cohort (n = 304,092; OS events = 161,611; CRC-specific deaths = 103,683), the age-by-stage interaction remained significant for both CSS (LRT χ²=182.95, df = 2, p < 0.001) and OS (LRT χ²=1678.0, df = 2, p < 0.001). Stage-specific estimates again showed attenuation of the age association from localized to distant disease, supporting the robustness of the primary findings (Supplementary Table S7). In an adenocarcinoma-only sensitivity analysis, the stage-dependent attenuation pattern was again observed. The age-by-stage interaction remained significant for both CSS (LRT χ²=124.15, df = 2, p < 0.001) and OS (LRT χ²=1950.8, df = 2, p < 0.001), with the strongest age association in localized disease and smaller relative hazards in distant-stage disease (Supplementary Table S8). Consistent with the Cox models, five-year survival estimates were lower in the ≥ 50 group across all stages. The absolute difference was particularly evident in localized and regional disease, whereas both age groups had poor absolute survival in distant-stage disease (Supplementary Table S9). Discussion In this population-based SEER analysis, older age was associated with worse CSS and OS after adjustment for available clinicopathologic factors, but the magnitude of this association varied substantially by stage. The age-by-stage interaction was statistically significant for both endpoints. In derived stage-specific estimates, the association of age ≥ 50 years with worse CSS was strongest in localized disease and attenuated in regional and distant disease; a similar pattern was observed for OS. These findings indicate that age should not be interpreted as a uniform prognostic increment across all stages of CRC, but rather as an effect modifier whose apparent impact depends on the stage context. These findings are clinically relevant in the setting of rising EOCRC. EOCRC has been described as a distinct clinical and epidemiologic entity, with potential differences in presentation and tumor biology3–6. These trends have contributed to U.S. guideline changes recommending average-risk CRC screening beginning at age 45 years7,8. We retained the 18–49 versus ≥ 50 grouping because it corresponds to the conventional EOCRC threshold and remains clinically intuitive for screening-related and epidemiologic interpretation. However, because this dichotomy may obscure heterogeneity within later-onset CRC, we also performed a sensitivity analysis using three age groups (18–49, 50–64, and ≥ 65 years), and the stage-dependent pattern persisted. Taken together, these results caution against assuming that younger age at diagnosis itself implies worse stage-matched prognosis. The persistent association between older age and worse outcomes likely reflects multiple overlapping mechanisms rather than a single biological effect of age. Older adults may have higher comorbidity burden, frailty, reduced physiologic reserve, and lower tolerance of curative-intent surgery or systemic therapy. They may also experience differences in treatment selection, dose intensity, perioperative risk, and completion of adjuvant treatment. Older adults are underrepresented in cancer clinical trials15, and pooled analyses suggest that the benefit and tolerability of modern adjuvant regimens may differ by age, particularly for oxaliplatin-based therapy in stage II/III colon cancer16. These factors are especially relevant in localized and regional disease, where long-term outcomes depend heavily on the ability to tolerate surgery and complete curative-intent therapy. In distant-stage disease, prognosis is dominated by metastatic burden, treatment responsiveness, and tumor biology. Therefore, the smaller relative age effect observed in distant disease should not be interpreted as evidence that age is unimportant in metastatic CRC. Rather, the attenuation may partly reflect compression of relative hazard ratios in a high-risk population with poor baseline prognosis, where the absolute risk of death is high in both age groups. This interpretation is supported by the five-year survival estimates, which showed poor absolute survival for both age groups in distant-stage disease despite persistent relative differences by age. This interpretation is consistent with contemporary metastatic CRC literature emphasizing metastatic disease burden, tumor biology, and treatment responsiveness as dominant determinants of outcome in advanced disease17. Endpoint choice is important when interpreting age gradients. As expected, the association between age and OS was stronger than that between age and CSS, reflecting the increasing contribution of non-cancer mortality with age and the role of competing risks in older patients18. For CSS, we used cause-specific Cox models based on the SEER cause-specific death classification. Fine–Gray subdistribution hazard models were not performed; therefore, CSS estimates should be interpreted as associations with the instantaneous hazard of CRC-specific death among individuals still at risk, rather than cumulative incidence estimates for prediction. This approach is aligned with our primary objective of quantifying covariate associations and testing age-by-stage effect modification. SEER cause-specific mortality also relies on death certificate–derived classification, which can be affected by misclassification and coding changes over time13,14. Nevertheless, the consistent stage-dependent pattern across CSS and OS supports the robustness of the central inference. This study has several strengths. The large population-based cohort provided substantial power to evaluate interaction effects and allowed adjustment for anatomic subsite and other available clinicopathologic variables. The cohort definition excluded appendiceal primaries, death certificate-only/autopsy-only cases, and non-microscopically confirmed cases, directly addressing key concerns regarding SEER cohort validity. The analytic strategy formally tested age-by-stage interaction rather than treating age as a uniform prognostic factor. Findings were also supported by multiple sensitivity analyses, including adenocarcinoma-only, treatment-adjusted, three-category age, colon-only, alternative COD handling, and piecewise Cox analyses. Several limitations should be acknowledged. First, SEER Summary Stage lacks the granularity of AJCC TNM and does not capture important prognostic details such as T/N substage, nodal burden, metastatic pattern, margin status, recurrence, or disease-free interval. Second, SEER does not provide detailed information on comorbidity, frailty, performance status, geriatric vulnerability, nutritional status, or gastrointestinal comorbidities such as inflammatory bowel disease, Crohn’s disease, or ulcerative colitis. These factors may differ by age and may influence both treatment selection and survival. Third, SEER lacks molecular and hereditary risk information, including mismatch repair/MSI status, RAS/BRAF status, Lynch syndrome, family history, and other genetic susceptibility factors that are particularly relevant in EOCRC. Fourth, although exploratory models including SEER chemotherapy and radiation recodes were performed, these variables are limited by “No/Unknown” categories and lack information on treatment intent, regimen, dose intensity, sequencing, completion, toxicity, and recurrence-directed therapy. Therefore, treatment coefficients should not be interpreted as causal treatment-effect estimates. Similarly, surgery should be interpreted primarily as a marker of treatment selection, disease resectability, and patient fitness, and adjustment for surgery may partially account for pathways through which age influences survival. Additional limitations relate to missingness and generalizability. Grade was unknown for a substantial proportion of cases; although an explicit Unknown category was modeled, missingness may be nonrandom. Coefficients for missingness-related categories, such as unknown race/ethnicity, should therefore be interpreted cautiously because they may reflect data completeness patterns rather than biological or clinical effects. The cohort was restricted to individuals with a single primary cancer, which improves interpretability but may limit generalizability, particularly among older adults with prior malignancies. Finally, because of the very large sample size, formal statistical tests, including proportional hazards diagnostics, may detect small departures from model assumptions. We therefore emphasize effect sizes, confidence intervals, graphical diagnostics, and consistency across sensitivity analyses rather than relying solely on statistical significance. Clinically, these findings support stage-aware counseling and prognostic modeling that includes age-by-stage interaction rather than treating age as a single uniform risk increment. For patients diagnosed before age 50 years, age alone should not be used as a shorthand for biologic aggressiveness when stage is comparable. Conversely, the pronounced age gradient in localized and regional disease highlights the need for careful treatment planning, geriatric assessment, optimization of comorbidities, and supportive interventions to help older adults tolerate curative-intent therapy19. At the research level, incorporating age-by-stage interaction may improve risk adjustment in registry studies, refine trial stratification, and better identify where age-associated outcome gaps are most clinically actionable. These findings extend prior SEER-based analyses of EOCRC and age-related CRC survival differences by stage by formally testing age-by-stage interaction and deriving stage-specific age effects20,21. Conclusion Older age was associated with worse CSS and OS in CRC, but the magnitude of this association was stage-dependent. The relative age-associated survival difference was greatest in localized disease and attenuated with advancing stage. These findings support incorporating age-by-stage interaction into prognostic modeling and encourage stage-aware interpretation of age-related survival differences, while recognizing the residual confounding inherent to registry-based analyses. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions All authors contributed to the conception and design of the study. BP led the data processing and conducted the primary statistical analyses, with input from IVB. All authors contributed to the interpretation of the results and reviewed and critiqued the statistical analysis. BP drafted the manuscript. IVB reviewed, edited, and critically revised the manuscript for important intellectual content. All authors read and approved the final version and agree to be accountable for all aspects of the work. Data availability The datasets analysed during the current study are available in the SEER (Surveillance, Epidemiology, and End Results) repository, [https://seer.cancer.gov/data/access.html](https:/seer.cancer.gov/data/access.html). Code availability No custom or bespoke computational tool, algorithm, software, or pipeline was developed for this study. Statistical analyses were performed in R using standard packages. The analytic code is available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets analysed during the current study are available in the SEER (Surveillance, Epidemiology, and End Results) repository, [https://seer.cancer.gov/data/access.html](https:/seer.cancer.gov/data/access.html). No custom or bespoke computational tool, algorithm, software, or pipeline was developed for this study. Statistical analyses were performed in R using standard packages. The analytic code is available from the corresponding author upon reasonable request.