Comparative Effectiveness of Oral Fluoropyrimidines Versus FOLFOX as Adjuvant Therapy for Stage III Colon Cancer: A Retrospective Cohort Study Using Overlap‐Weighted Restricted Mean Survival Time Analysis

ABSTRACT

Purpose

Oxaliplatin‐based chemotherapy is the standard adjuvant treatment for stage III colon cancer, but oral fluoropyrimidines remain widely used for patients unsuitable for combination therapy. Real‐world data comparing these strategies are limited.

Methods

We conducted a retrospective cohort study of patients with stage III colon cancer who received adjuvant chemotherapy at Kaohsiung Veterans General Hospital (2017–2021). Using a 12‐month landmark design, we compared disease‐free survival (DFS) between patients receiving FOLFOX (IV‐Only) versus oral fluoropyrimidines (PO‐Only). The primary analysis used overlap‐weighted restricted mean survival time (RMST) differences to quantify absolute treatment effects. Cox proportional hazards regression provided complementary hazard ratio estimates.

Results

Of 157 patients included, 58 (36.9%) received FOLFOX and 99 (63.1%) received oral fluoropyrimidines. During median follow‐up of 36 months, 22 DFS events occurred. The 3‐year DFS was 93.1% versus 76.5% for IV‐Only and PO‐Only groups, respectively. The primary analysis showed that patients receiving FOLFOX had 4.0 additional months of disease‐free time over 48 months (RMST difference: 4.0 months; 95% CI, 0.6–7.9; p = 0.014) and 2.4 months over 36 months (RMST difference: 2.4 months; 95% CI, 0.2–4.7; p = 0.024). The hazard ratio from overlap‐weighted Cox regression with covariate adjustment was 0.37 (95% CI 0.09–1.45; p = 0.154), directionally consistent with the RMST findings but not reaching statistical significance. Hazard ratio estimates were directionally consistent across sensitivity analyses but varied in statistical significance, likely reflecting limited power: unadjusted (HR 0.23; p = 0.019), stabilized weighting with covariates (HR 0.19; p = 0.046), and multivariable adjustment (HR 0.44; p = 0.249).

Conclusions

In this real‐world cohort, oral fluoropyrimidines were associated with 4.0 fewer months of disease‐free time compared with FOLFOX at 48 months among patients for whom both treatment options were appropriate. As a single‐center study with limited events, these findings should be confirmed in larger multicenter cohorts. The combination of RMST with overlap weighting provides an interpretable framework for comparative effectiveness questions where treatment selection is strong.

Key Points

• Real‐world data comparing oral fluoropyrimidines versus FOLFOX as adjuvant therapy for stage III colon cancer among patients suitable for either regimen are limited.

• Overlap weighting targets patients in clinical equipoise, providing treatment effect estimates most relevant to real‐world decision‐making.

• Using a 12‐month landmark design, oral fluoropyrimidines were associated with 4.0 fewer months of restricted mean disease‐free time compared with FOLFOX at 48 months.

• The overlap‐weighted hazard ratio favored FOLFOX (HR 0.37; 95% CI 0.09–1.45) but did not reach statistical significance; effect direction was consistent across IPTW and multivariable Cox sensitivity analyses.

• Restricted mean survival time offers clinically interpretable absolute effect measures in studies with limited events.

Plain Language Summary

After surgery for stage III colon cancer, patients typically receive chemotherapy to reduce the risk of cancer returning. Two common options are FOLFOX (an intravenous combination of oxaliplatin and fluorouracil) and oral fluoropyrimidines (tablets taken by mouth). FOLFOX is generally more effective but causes more side effects, including nerve damage. In everyday practice, doctors often choose oral therapy for older patients or those with other health conditions, while reserving FOLFOX for younger, fitter patients. However, many patients fall somewhere in between, where either option could be reasonable. We studied 157 patients with stage III colon cancer who remained recurrence‐free at 12 months after surgery to compare outcomes between these two treatments, focusing specifically on patients who could have received either regimen. Using statistical methods to account for differences between patient groups, we found that patients who received oral fluoropyrimidines had, on average, 4 fewer months of disease‐free time over 4 years compared with those who received FOLFOX. The hazard ratio pointed in the same direction but was not statistically significant, likely due to the small number of recurrence events. Our results suggest that when both treatments are viable options, choosing oral therapy over FOLFOX may come at a measurable cost in disease‐free time, which may help inform shared decision‐making between patients and their doctors.

1. Introduction

Adjuvant chemotherapy is a cornerstone of treatment for patients with resected stage III colon cancer, with well‐established benefits in reducing recurrence and improving overall survival. The landmark MOSAIC trial demonstrated that adding oxaliplatin to fluorouracil and leucovorin (FOLFOX) significantly improved disease‐free survival (DFS) and overall survival compared with fluorouracil and leucovorin (FL) alone, with a 10‐year overall survival of 67.1% versus 59.0% in patients with stage III disease (hazard ratio [HR] 0.80, p = 0.016) [1, 2]. Similar benefits were confirmed in the NSABP C‐07 trial and the NO16968 (XELOXA) study [3, 4]. Based on these findings, oxaliplatin‐based regimens have become the standard of care for fit patients with stage III colon cancer.

However, not all patients are suitable candidates for oxaliplatin‐based chemotherapy. Concerns regarding cumulative neurotoxicity, which affects up to 92% of patients receiving oxaliplatin‐based regimens, with grade 3 sensory neuropathy reported in 12%–18% of cases [1, 3, 5], advanced age, comorbidities, and patient preference have led to the continued use of single‐agent fluoropyrimidines in clinical practice [5]. Fluoropyrimidines are antimetabolite agents that inhibit thymidylate synthase and serve as the backbone of adjuvant chemotherapy for colon cancer. Oral formulations, including capecitabine and tegafur‐uracil (UFT), offer the convenience of home‐based administration. The X‐ACT trial established oral capecitabine as a non‐inferior alternative to intravenous bolus fluorouracil/leucovorin (Mayo Clinic regimen), with comparable DFS (HR 0.88; 95% confidence interval [CI] 0.77–1.01) and a more favorable toxicity profile [6, 7].

Similarly, oral UFT plus leucovorin demonstrated non‐inferior DFS compared with intravenous fluorouracil in a Japanese phase III trial of stage III colorectal cancer [8]. However, a pooled analysis of individual patient data from four randomized trials demonstrated that the addition of oxaliplatin to fluoropyrimidines significantly improved both DFS and OS in stage III colon cancer, confirming the superiority of oxaliplatin‐based combinations over fluoropyrimidine monotherapy [9].

In real‐world clinical practice, the choice between oxaliplatin‐based intravenous regimens and oral fluoropyrimidine monotherapy is often influenced by a complex interplay of patient factors, including age, Eastern Cooperative Oncology Group (ECOG) performance status (a scale from 0 [fully active] to 4 [completely disabled] assessing daily functioning), comorbidities, expected tolerance, and patient preference [10]. However, data directly comparing these two treatment strategies in routine clinical settings are limited. While randomized controlled trials provide robust efficacy data under controlled conditions, they may not fully reflect the heterogeneous patient populations and treatment patterns encountered in everyday practice [11].

Therefore, we conducted this retrospective cohort study to compare DFS outcomes between patients receiving intravenous oxaliplatin‐based chemotherapy (FOLFOX) and those receiving oral fluoropyrimidine monotherapy as adjuvant treatment for stage III colon cancer in a real‐world setting. We employed a landmark analysis approach and propensity score overlap weighting with covariate adjustment to minimize immortal time bias and confounding by indication, respectively.

2. Methods

2.1. Study Design and Population

This retrospective cohort study was conducted at Kaohsiung Veterans General Hospital, Taiwan. We identified consecutive patients with pathologically confirmed stage III colon cancer who underwent radical surgical resection between January 2017 and December 2021. All clinical data were extracted through manual chart review of electronic medical records. Recurrence events (local recurrence and distant metastasis) were identified based on pathological confirmation, imaging findings documented by treating physicians, or multidisciplinary tumor board decisions, rather than administrative codes. Patients were included if they received adjuvant chemotherapy with either oxaliplatin‐based regimens (FOLFOX) or oral fluoropyrimidines (capecitabine or UFT). Patients who received other regimens (e.g., FOLFIRI, bevacizumab, or cetuximab) or those with incomplete medical records, defined as missing essential baseline variables required for the propensity score model (e.g., pathological staging, ECOG performance status) or outcome data (e.g., follow‐up dates, recurrence status), were excluded; such cases comprised < 5% of the screened cohort. This study was approved by the Institutional Review Board of Kaohsiung Veterans General Hospital (IRB No. 251201‐2), and the requirement for informed consent was waived due to the retrospective nature of the study. The study flow is shown in Figure 1.

FIGURE 1.

Patient selection and study design. (A) CONSORT flowchart illustrating patient selection from the initial stage III colon cancer cohort (n = 331) to the primary analysis cohort (n = 157). The primary analysis compared IV‐Only (FOLFOX) versus PO‐Only (oral fluoropyrimidine) regimens; the Switch group was reserved for secondary analysis. (B) Study design timeline showing the 12‐month landmark approach. Treatment exposure was assessed during the first 12 months post‐surgery; all survival analyses used the landmark date as time zero, with restricted mean survival time (RMST) estimated at τ = 48 months.

2.2. Landmark Analysis

To avoid immortal time bias, we employed a 12‐month landmark approach [12, 13]. The landmark time point was set at 12 months post‐surgery, corresponding to the expected completion of adjuvant chemotherapy and ensuring at least one surveillance computed tomography scan per institutional protocol. Patients who experienced a DFS event (recurrence or death) or were lost to follow‐up before the landmark date were excluded from the analysis. Survival time was measured from the landmark date onward, and treatment exposure was determined based on the treatment received within the first 12 months post‐surgery. All survival analyses, including Kaplan–Meier estimation and restricted mean survival time (RMST) calculations, used the landmark date as time zero, ensuring that the comparison was free of immortal time bias. Follow‐up continued through the date of last clinical contact or data extraction (June 2024). There was no prespecified maximum follow‐up duration.

2.3. Outcome and Exposure Definitions

The primary endpoint was landmark disease‐free survival (landmark DFS), defined as the time from the 12‐month landmark date to the first occurrence of local recurrence, distant metastasis, or death from any cause. Patients who remained alive and disease‐free were censored at the date of last follow‐up. For the primary analysis, patients were categorized into two treatment groups based on their entire adjuvant chemotherapy course: intravenous‐only (IV‐Only; patients who received FOLFOX exclusively) and oral‐only (PO‐Only; patients who received oral fluoropyrimidines exclusively). Patients who switched between regimens during treatment were excluded from the primary analysis.

2.4. Propensity Score Estimation and Weighting

Given the non‐randomized nature of this study and anticipated differences in baseline characteristics between treatment groups, we employed propensity score (PS) methods to reduce confounding. The probability of receiving IV‐Only treatment (versus PO‐Only) was estimated using logistic regression. Covariates included in the PS model were selected a priori based on clinical relevance and included: age (continuous), sex, pathological T stage (pT4 vs. others), pathological N stage (pN2 vs. N1), lymph node ratio (continuous), tumor location (right colon vs. left colon), lymphovascular invasion, perineural invasion, tumor differentiation (poor vs. well/moderate), ECOG performance status (≥ 1 vs. 0, reflecting the clinical threshold between fully active patients and those with any degree of functional limitation), and year of surgery (continuous).

We applied overlap weighting, which assigns weights equal to the probability of receiving the opposite treatment (i.e., 1 − PS for treated patients and PS for control patients) [14, 15]. This approach optimizes covariate balance by emphasizing patients in the region of clinical equipoise while maintaining the effective sample size (ESS). Covariate balance was assessed using standardized mean differences (SMD), with a threshold of |SMD| < 0.1 considered indicative of adequate balance. The ESS after weighting was calculated using Kish's approximation: ESS = (∑w)²/∑w ², where w represents the overlap weight for each individual [16].

2.5. Statistical Analysis

The primary estimand was the difference in restricted mean DFS time (RMST) at 48 months between treatment groups within the overlap‐weighted population. The overlap‐weighted population represents patients for whom both treatment options would be clinically appropriate based on observed baseline characteristics. This target population mimics clinical equipoise and provides estimates most relevant to treatment decision‐making in ambiguous cases. RMST was prespecified as the primary estimand to provide an absolute, patient‐interpretable effect measure in the anticipated setting of few events and potential violations of the proportional hazards assumption. Overlap‐weighted Cox proportional hazards regression with covariate adjustment was used to estimate the HR as a complementary relative measure, using robust sandwich variance estimators for confidence intervals. This combined approach of propensity score weighting with outcome regression adjustment provides enhanced protection against residual confounding.

As sensitivity analyses, we conducted: (1) stabilized inverse probability of treatment weighting (SIPTW) with covariate adjustment and weights truncated at the 1st and 99th percentiles [17]; (2) multivariable Cox regression adjusting for age, pT4, pN2, lymphovascular invasion, and perineural invasion; and (3) unadjusted Cox regression. Direct covariate adjustment in the context of propensity score weighting provides an additional layer of confounding control but requires estimating additional parameters, which may reduce precision in small samples. RMST differences were calculated at τ = 36 months and τ = 48 months [18].

All statistical analyses were performed using Python version 3.11 with the lifelines package (version 0.27) for survival analysis and scikit‐learn (version 1.3) for propensity score estimation. A two‐sided p‐value < 0.05 was considered statistically significant. Missing data were minimal in our cohort (< 5% for all covariates). A complete case analysis was performed without imputation.

3. Results

3.1. Patient Characteristics

A total of 157 patients met the inclusion criteria and were included in the landmark analysis after excluding those with events or loss to follow‐up before 12 months post‐surgery. Of these, 58 patients (36.9%) received IV‐Only treatment (FOLFOX) and 99 patients (63.1%) received PO‐Only treatment (oral fluoropyrimidines). The median follow‐up time from the landmark date was 36 months (interquartile range, 24–48 months).

Baseline characteristics differed substantially between treatment groups (Table 1). Patients in the PO‐Only group were older (mean age 68.2 vs. 60.8 years), more likely to have ECOG performance status ≥ 1 (84.8% vs. 46.6%) and pT4 stage (14.1% vs. 5.2%), and less likely to have high‐risk pathological features including pN2 stage (14.1% vs. 44.8%) and lymphovascular invasion (42.4% vs. 53.4%).

TABLE 1.

Baseline characteristics of Stage III colon cancer patients.

Variable	Raw			OW
	IV‐Only	PO‐Only	SMD	IV‐Only	PO‐Only	SMD
n	58	99		23.9 a	23.9 a
Age	60.8 ± 9.3	68.2 ± 12.7	−0.661	62.8 ± 8.9	63.1 ± 13.3	−0.021
Male	37 (63.8%)	44 (44.4%)	0.393	54.9%	54.3%	0.012
pT4	3 (5.2%)	14 (14.1%)	−0.305	7.3%	7.8%	−0.019
pN2	26 (44.8%)	14 (14.1%)	0.709	32.0%	30.9%	0.022
LNR	0.2 ± 0.1	0.1 ± 0.1	0.722	0.1 ± 0.1	0.1 ± 0.1	0.020
Tumor location	—	—	—	—	—	—
Right	17 (29.3%)	33 (33.3%)	−0.086	31.3%	30.9%	0.008
Left	41 (70.7%)	66 (66.7%)	0.086	68.7%	69.1%	−0.008
LVI	31 (53.4%)	42 (42.4%)	0.220	50.9%	50.3%	0.011
PNI	7 (12.1%)	12 (12.1%)	−0.002	13.0%	13.3%	−0.008
Poorly differentiated	4 (6.9%)	7 (7.1%)	−0.007	8.7%	8.6%	0.002
ECOG ≥ 1	27 (46.6%)	84 (84.8%)	−0.875	64.9%	66.4%	−0.032
Surgery year	—	—	—	—	—	—
2017	5 (8.6%)	17 (17.2%)	−0.256	6.1%	18.3%	−0.381
2018	11 (19.0%)	17 (17.2%)	0.046	20.3%	12.6%	0.211
2019	14 (24.1%)	22 (22.2%)	0.045	22.0%	19.1%	0.072
2020	19 (32.8%)	18 (18.2%)	0.337	35.7%	17.1%	0.431
2021	9 (15.5%)	25 (25.3%)	−0.242	15.9%	33.0%	−0.405

Note: Values are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables. For overlap‐weighted pseudo‐population, percentages represent weighted proportions.

Abbreviations: ECOG, Eastern Cooperative Oncology Group; LNR, lymph node ratio; LVI, lymphovascular invasion; OW, overlap weighting; PNI, perineural invasion; SMD, standardized mean difference.

^a Expressed as the sum of overlap weights, representing the size of the weighted pseudo‐population.

3.2. Propensity Score Weighting and Covariate Balance

After overlap weighting, all covariates achieved excellent balance with SMD less than 0.1 (Table 1, Figure 2). After overlap weighting, the sum of weights was 47.8 (23.9 per group). The ESS, calculated using Kish's formula, was 97.4 (62.0% of the original sample), indicating reasonable retention of statistical information despite the weighting.

FIGURE 2.

Propensity score distribution and covariate balance assessment. (A) Propensity score distribution between IV‐Only and PO‐Only groups before and after overlap weighting, with overlap bounds indicated by vertical dashed lines. (B) Love plot demonstrating standardized mean differences (SMD) for baseline covariates before (filled circles) and after (open circles) overlap weighting. Variables with SMD < 0.1 (vertical dashed line) indicate adequate balance.

3.3. Primary Analysis

During a median follow‐up of 36 months from the landmark date, 22 DFS events occurred (3 in the IV‐Only group and 19 in the PO‐Only group). The Kaplan–Meier estimated 3‐year landmark DFS rate was 93.1% (95% CI, 86.5%–99.7%) in the IV‐Only group and 76.5% (95% CI, 67.6%–85.4%) in the PO‐Only group (Figure 3).

FIGURE 3.

Disease‐free survival by chemotherapy regimen. Kaplan–Meier curves comparing disease‐free survival between IV‐Only and PO‐Only groups after overlap weighting. ^†Number at risk represents pseudo‐population size under overlap weighting.

The primary analysis using overlap weighting with covariate adjustment showed differences in disease‐free time favoring FOLFOX across multiple time horizons. At 36 months, the FOLFOX group had an additional 2.4 months of disease‐free time (RMST difference: 2.4 months; 95% CI, 0.2–4.7; p = 0.024). This benefit increased over longer follow‐up: at 48 months, the FOLFOX group had 4.0 additional months of disease‐free time compared with oral fluoropyrimidines (RMST difference: 4.0 months; 95% CI, 0.6–7.9; p = 0.014; Figure 4). The hazard ratio from overlap‐weighted Cox regression with covariate adjustment was 0.37 (95% CI, 0.09–1.45; p = 0.154); this was directionally consistent with the RMST findings but did not reach statistical significance.

FIGURE 4.

Restricted mean survival time analysis. Comparison of disease‐free survival between IV‐Only and PO‐Only groups using restricted mean survival time (RMST) approach at 48 months, with difference in RMST (ΔRMST) and corresponding 95% CI shown.

3.4. Sensitivity Analyses

Effect estimates were consistent across all analytical approaches (Table 2, Figure 5). In the unadjusted analysis, FOLFOX was associated with significantly better outcomes (HR, 0.23; 95% CI, 0.07–0.79; p = 0.019). SIPTW with covariate adjustment yielded a similar result (HR, 0.19; 95% CI, 0.04–0.97; p = 0.046). When using multivariable Cox regression with direct covariate adjustment, the point estimate remained in the same direction but did not reach statistical significance (HR, 0.44; 95% CI, 0.11–1.79; p = 0.249). The attenuation of effect size and widening of confidence intervals observed with traditional multivariable regression compared with propensity score methods is consistent with the challenges of direct covariate adjustment in small samples with pronounced baseline imbalances.

TABLE 2.

Treatment effect estimates for disease‐free survival.

Outcome	Estimate	95% CI	p
HR (unadjusted)	0.23	0.07–0.79	0.019*
HR (OW + covariates) a	0.37	0.09–1.45	0.154
HR (SIPTW + covariates)	0.19	0.04–0.97	0.046*
HR (multivariable)	0.44	0.11–1.79	0.249
ΔRMST (τ = 36.0 months)	+2.4 months	0.2–4.7	0.024*
ΔRMST (τ = 48.0 months)	+4.0 months	0.6–7.9	0.014*

Note: Hazard ratios (HR) are reported for Cox proportional hazards models. The primary analysis (OW + Covariates) used overlap weighting with covariate adjustment in a combined weighting‐and‐regression framework. ΔRMST represents the difference in restricted mean survival time (RMST) between IV‐Only and PO‐Only groups, calculated at twotime horizons (τ = 36 and 48 months). Positive values indicate longer disease‐free time in the IV‐Only group. All analyses compared IV‐Only (FOLFOX, n = 58) versus PO‐Only (oral fluoropyrimidines, n = 99) in the 12‐month landmark cohort (N = 157). HR < 1.0 favors IV‐Only treatment.

Abbreviations: CI, confidence interval; OW, overlap weighting; RMST, restricted mean survival time; SIPTW, stabilized inverse probability of treatment weighting.

^a Primary analysis.

^* p < 0.05.

FIGURE 5.

Hazard ratios for disease‐free survival across different analytical approaches. Forest plot comparing hazard ratios (HR) and 95% confidence intervals (CI) for IV‐Only versus PO‐Only regimens using unadjusted analysis, overlap weighting with covariates (primary analysis), SIPTW with covariates, and multivariable Cox regression.

4. Discussion

In this retrospective cohort study of patients with stage III colon cancer, we compared the DFS between oral fluoropyrimidines and FOLFOX among patients for whom both options were clinically reasonable. The RMST analysis showed that patients receiving FOLFOX had 4.0 additional months of disease‐free time at 48 months (2.4 months at 36 months) (p = 0.014). The hazard ratio from overlap‐weighted Cox regression with covariate adjustment was 0.37 (95% CI 0.09–1.45; p = 0.154), directionally favoring FOLFOX but not reaching statistical significance. Hazard ratio point estimates across all analytical approaches ranged from 0.19 to 0.44, consistently favoring FOLFOX, though not all reached statistical significance given the limited number of events.

The observed RMST difference reflects a conditional estimand specific to the overlap‐weighted population, namely patients who were clinically suitable for either regimen and remained recurrence‐free at 12 months. While the overlap‐weighted population emphasizes patients in the region of propensity score overlap, substantial baseline differences remained before weighting, indicating that treatment selection was strongly influenced by age and functional status. After weighting, these differences were balanced (all SMD < 0.1), creating a pseudo‐population where both treatments would be appropriate based on observed covariates. This target population differs fundamentally from the intent‐to‐treat population in MOSAIC, which demonstrated a 23% relative risk reduction in DFS [1] and significant overall survival benefit [2, 19] with FOLFOX. In the presence of treatment effect heterogeneity, conditional effects in selected subgroups can differ substantially from marginal effects in the overall population. Furthermore, our comparison involved oral fluoropyrimidines rather than intravenous fluorouracil, and while the X‐ACT trial established non‐inferiority of capecitabine to bolus fluorouracil regimens [6], direct comparison with infusional fluorouracil in the FOLFOX era remains limited. The pronounced baseline differences between groups before weighting confirm that treatment selection was strongly influenced by patient characteristics. Overlap weighting identified a pseudo‐population in which both treatments were clinically reasonable, and the successful covariate balance after weighting (all SMD < 0.1) supports the validity of comparisons within this target population.

The consistency of hazard ratios across different analytical approaches (point estimates ranging from 0.19 to 0.44, all favoring FOLFOX) provides converging evidence for the treatment effect direction. However, the varying width of confidence intervals reflects the challenges of hazard ratio estimation in small samples with limited events. This underscores the value of RMST as the primary estimand, which provides a clinically interpretable absolute benefit measure that is robust to proportional hazards violations and does not rely on asymptotic properties that may not hold in small samples. The discrepancy between the statistically significant RMST differences and the non‐significant hazard ratios warrants discussion. With only 22 events, estimation of the log‐hazard ratio is highly imprecise, resulting in wide confidence intervals. In contrast, RMST integrates the entire survival curve up to the restriction time without assuming proportional hazards, providing a more stable estimate in small samples. Furthermore, the asymptotic properties underpinning Cox regression may not hold well with so few events, whereas RMST‐based inference tends to be more robust in this setting. This methodological difference further supports our prespecified choice of RMST as the primary estimand.

In studies with limited events and potential violations of the proportional hazards assumption, RMST offers advantages as a primary estimand over the hazard ratio. Given the limited number of events in our study, RMST provides a more stable and clinically interpretable summary of treatment effect. RMST directly quantifies the area between survival curves up to a clinically relevant time horizon, providing an absolute measure of treatment benefit that does not rely on the proportional hazards assumption and is robust to late crossings or converging survival curves. Furthermore, RMST yields a clinically intuitive interpretation: in our study, patients receiving FOLFOX had an average of 4.0 additional months of disease‐free time over 48 months compared with those receiving oral fluoropyrimidines. This absolute benefit is arguably more meaningful to patients and clinicians than a hazard ratio, particularly when confidence intervals are wide. The consistency of the effect direction across all Cox‐based approaches, combined with the statistically significant RMST difference, provides converging evidence supporting a treatment benefit in the direction favoring FOLFOX, though larger studies are needed to confirm its magnitude.

Our overlap‐weighted 3‐year DFS estimates (93.1% for FOLFOX and 76.5% for oral fluoropyrimidines) appear higher than those reported in the MOSAIC trial (72.2% for FOLFOX and 65.3% for FL in stage III patients) [1]. This difference likely reflects our landmark design, which conditions on patients remaining recurrence‐free at 12 months and thus excludes early recurrences. Era effects (improved surgical techniques and surveillance), differences in comparator regimens, and the overlap‐weighted target population may also contribute to the higher observed survival rates.

Baseline characteristics showed that FOLFOX was preferentially administered to younger patients with higher‐risk disease despite better functional status. Several clinical factors may contribute to the observed differences. Patients selected for FOLFOX were younger, had better functional status, and had more adverse pathological features, consistent with clinicians reserving the more intensive regimen for those most likely to tolerate and benefit from it. This pattern of treatment selection reflects guideline‐concordant care, wherein oxaliplatin‐based therapy is recommended for patients who can tolerate combination chemotherapy, while oral fluoropyrimidines serve as an alternative for elderly patients or those with comorbidities precluding oxaliplatin use [20, 21]. The IDEA collaboration demonstrated that among patients receiving CAPOX, 3 months of therapy was non‐inferior to 6 months for low‐risk stage III disease, supporting the broader principle that adjuvant chemotherapy decisions, including both the choice of regimen and its duration, should be tailored to individual patient and tumor characteristics, which is the central clinical question addressed by our study [22]. The proportion of patients receiving FOLFOX varied across the study period (2017–2021), potentially reflecting evolving clinical practice and the influence of emerging evidence such as the IDEA collaboration results. We included year of surgery as a continuous covariate in the propensity score model to account for these secular trends, though residual confounding from period effects cannot be excluded.

Our study has several methodological strengths. The 12‐month landmark design eliminated immortal time bias, ensuring that treatment comparisons were not confounded by the time required to complete adjuvant therapy. Overlap weighting optimized covariate balance while preserving ESS, and the combination of propensity score weighting with outcome regression adjustment provided enhanced control for measured confounding. The RMST analysis complemented hazard ratio estimation by providing a clinically interpretable absolute measure of treatment effect that does not rely on proportional hazards assumptions. The consistency of results across multiple analytical approaches supports the robustness of our findings.

This study has important limitations. First, as a single‐center retrospective study with 157 patients and only 22 events, it was underpowered to detect treatment effects with precision, as reflected in the wide confidence intervals. In underpowered studies, statistically significant results tend to overestimate the true effect size (Type M error) [23]; accordingly, the observed RMST difference of 4.0 months may overestimate the true treatment effect. Furthermore, RMST estimates at longer time horizons (48 vs. 36 months) are more susceptible to unmeasured time‐varying confounders accumulating over the extended observation period. Second, microsatellite instability (MSI) status was available but could not be reliably incorporated into the models because the number of MSI‐high cases was small (only seven cases in the analyzed cohort) and completely collinear with right‐sided tumor location, which would lead to model instability. Other molecular markers, such as circulating tumor DNA, were not routinely available [24, 25, 26, 27]. Third, detailed information on chemotherapy dose modifications, treatment‐related toxicities, and reasons for treatment discontinuation was not systematically captured. Fourth, cumulative neurotoxicity, a key determinant of oxaliplatin dose modification and discontinuation, was not systematically captured in our database and could not be included as a covariate or outcome. Similarly, specific comorbidities that may influence treatment selection, such as pre‐existing neuropathy, renal insufficiency, and cardiovascular disease, were not available at the individual level for inclusion in the propensity score model. Given the modest effect sizes and wide confidence intervals observed, residual confounding from these unmeasured variables is a meaningful concern that tempers the strength of our conclusions. Fifth, treatment strategies and patient status may evolve over time, but time‐varying exposures and covariates were not explicitly modeled, and the estimated effects should therefore be interpreted as conditional on the landmark design. Sixth, overall survival was not analyzed because the number of deaths in the landmark cohort was limited (n = 19, with only 2 in the IV‐Only group), insufficient for meaningful between‐group comparison; additionally, the follow‐up duration may be insufficient to capture long‐term survival differences that emerge years after adjuvant therapy [28]. Finally, residual confounding from unmeasured variables cannot be excluded despite propensity score adjustment. The absence of systematic neurotoxicity outcome data also prevented assessment of the efficacy and tolerability trade‐off between FOLFOX and oral fluoropyrimidines. Future studies with CTCAE‐graded adverse event data should incorporate RMST‐based analyses of both efficacy and tolerability endpoints to provide a more complete picture for shared decision‐making.

In conclusion, in this real‐world cohort of patients with stage III colon cancer, oral fluoropyrimidines were associated with 4.0 fewer months of disease‐free time compared with FOLFOX at 48 months among patients for whom both treatment options were clinically appropriate. The 12‐month landmark design ensured that treatment comparisons were free of immortal time bias; however, findings are conditional on patients remaining recurrence‐free at 12 months post‐surgery, and generalizability is limited to this landmark‐eligible population. As with all observational studies, residual confounding from unmeasured variables may remain; our estimates should be interpreted as adjusted associations within the overlap population defined by measured covariates. Effect estimates were consistent across all analytical approaches, demonstrating the robustness of the observed association. This quantitative estimate may inform shared decision‐making when clinicians and patients weigh treatment options in the region of clinical equipoise. As a single‐center study with limited events, these findings should be confirmed in larger multicenter cohorts. The combination of RMST with overlap weighting provides a methodological framework for comparative effectiveness research in observational studies with limited events and substantial confounding by indication.

Author Contributions

Shih‐Feng Huang: conceptualization, methodology, software, formal analysis, data curation, writing – original draft, writing – review and editing, visualization, project administration. Chao‐Wen Hsu: supervision, resources, validation, writing – review and editing. Chih‐Chien Wu: investigation, data curation, writing – review and editing. Yu‐Hsun Chen: methodology, validation, writing – review and editing. All authors have read and approved the final manuscript.

Funding

The authors have nothing to report.

Ethics Statement

This study was approved by the Institutional Review Board of Kaohsiung Veterans General Hospital (IRB No. 251201‐2). The requirement for informed consent was waived due to the retrospective nature of the study.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request.