Comparative Effectiveness of Laparoscopic versus Open Colectomy on 5-Year Overall Survival in Non-Metastatic Colon Cancer: A Target Trial Emulation Study Using Retrospective Cohort Data from a Cancer Registry Authors: Semaw Ferede Abera, Gabriele Robers, Anika Kästner, Ulrike Stentzel, Kerstin Weitmann, Wolfgang Hoffmann Published: 2026 ABSTRACT AbstractBackgroundThe clinical acceptance of laparoscopic colectomy (LC) as a standard alternative treatment for colon cancer hinges primarily on evidence of favorable long-term oncological outcome. The aim of this study was to compare the 5-year overall survival (OS) between stage I–III colon cancer patients who underwent either LC or open colectomy (OC).Materials and MethodsA target trial emulation (TTE) design was implemented using statewide retrospective cohort data from the Mecklenburg-Western Pomerania Cancer Registry spanning 2008 to 2018. The causal interpretation of TTE relied on the assumptions of consistency, conditional exchangeability, and positivity. The Royston-Parmar parametric survival model was applied to estimate the hazard ratio with bootstrapped 95% confidence interval (95% CI). This was complemented by differences in estimates of restricted mean survival time (RMST). Sensitivity analyses, including the E-value method, were conducted to assess the robustness of the estimated treatment effect. Three traditional (fully adjusted, change-in-estimate, and Directed Acyclic Graph-guided) Cox models were performed for comparison purposes.ResultsWe included 1265 patients, of whom 30% (n= 380) received LC. The estimated HR for 5-year OS comparing LC with OC was 0.81 (95% CI: 0.60, 1.08). RMST differences for OS favored LC, but the effects were small: 0.12 months (95% CI: −0.10, 0.34), 0.36 months (95% CI: −0.15, 0.87), and 1.20 months (95% CI: −0.20, 2.60) up to 1-, 3-, and 5-years, respectively. Sensitivity analyses generally confirmed the robustness of the results. Traditional Cox models showed a survival advantage for LC, but all were affected by bias.ConclusionThis study found comparable 5-year OS between LC and OC. The traditional Cox models spuriously overestimated the survival advantage of LC. TTE can provide useful real-world evidence for comparative effectiveness in surgical oncology, particularly when randomized evidence is limited, but it cannot replace well-conducted RCTs.Keywords: colon cancer, laparoscopic colectomy, overall survival, target trial emulation, cancer registry, Germany Introduction The acceptability of laparoscopic colectomy (LC) as an alternative to open colectomy (OC) for colon cancer treatment is contingent upon its long-term oncological outcomes.1 Randomized controlled trials (RCTs) have demonstrated that LC improved short-term outcomes and provided equivalent 3-year overall survival (OS) compared to OC.2–6 Long-term OS was studied in multiple additional trials and meta-analyses of some of these RCTs, which generally concluded that the two approaches were comparable.7–14 Although RCTs provide the highest level of evidence for treatment comparisons, the trials that had established a foundation for comparing long-term OS between the two surgical approaches have a number of limitations. These include small sample sizes and power calculation issues;7,8,10–12 inability to prove a true equivalent effect due to non-inferior design,10,11 underpowered subgroup analyses;8,10,14 inconsistencies in tumor stage-specific results;12,14,15 restricted patient populations based on age or tumor stage,11 and combining colon and rectal cancers as a single tumor entity.13 Collectively, the landmark RCTs, including COST, COLOR, CLASICC, and Barcelona, consistently demonstrated comparable long-term OS between LC and OC. Nevertheless, uncertainty remains regarding whether these findings are maintained in routine clinical practice, as these trials were conducted in selected patient populations under controlled conditions, potentially limiting the generalizability of their findings to broader, unselected populations encountered in everyday surgical care.3,7,8,10–14,16 Consequently, whether the results observed in randomized studies are maintained in real-world clinical practice remains an important unanswered question.16,17 High-quality population-based cancer registry data therefore offer an important opportunity to complement trial evidence by evaluating long-term treatment effectiveness in routine clinical care.12 However, it should be noted that observational data are susceptible to serious biases, including immortal time bias and selection bias, which can distort estimates; thus, traditional analysis of observational data often yields misleading results.18–20 For instance, when comparing the two surgical treatment modalities, differences in the baseline profiles of patients (patient selection) or the exclusion of a substantial proportion of patients with missing data can give biased results.21,22 The target trial emulation (TTE) framework addresses these biases and may provide more accurate estimates of causal treatment effects.20,23 TTE is a design framework for causal inference in which investigators explicitly define the protocol of the hypothetical pragmatic randomized trial that would ideally answer the research question and then emulate that protocol as closely as possible using observational data to estimate causal treatment effects.23,24 In contrast to the ideal settings of RCTs, this rigorous study design allows researchers to evaluate the comparative effectiveness of therapeutic interventions in a real-world setting.24,25 There is evidence that causal effect estimates derived from a carefully specified and emulated target trial are frequently similar to those from RCTs,26 rendering TTE a valuable option to address clinical research questions in a timely manner, particularly when RCT evidence is unavailable. The aim of this study was to compare the 5-year OS between stage I–III colon cancer patients who underwent either LC or OC, using retrospective cohort data from Mecklenburg-Western Pomerania Cancer Registry covering 2008 to 2018. Our work presents the first implementation of TTE on statewide German cancer registry data to compare the effectiveness of LC versus OC on 5-year OS and contrast the results estimated by the traditional Cox regression models. By applying the TTE methodology to address biases inherent in the observational registry data, our study demonstrates unbiased treatment effect estimates that the traditional analytic methods fail to achieve. This approach provides methodological robustness for using existing cancer registry data to conduct high-quality surgical effectiveness research. This is particularly relevant when evidence from RCTs is unavailable or when it needs to be triangulated with real-world data, especially in cases where the external validity of RCTs is in question. Materials and Methods Design of the Target and Emulated Trial A TTE methodology was employed to conduct this observational study. The methodological details of this study were specified a priori in a published protocol.27 Figure 1 presents the flow chart of the emulated study. A summary of how we specified and emulated the hypothetical pragmatic target trial is provided in Table S1. In brief, our study population included adult patients with confirmed diagnosis of solitary, primary, stage I–III colon cancer diagnosed between January 1, 2008, and December 31, 2018, who underwent LC or OC during this period. Eligibility flowchart. Abbreviations: LC, Laparoscopic colectomy; OC, Open colectomy. The treatment strategies in the target trial would be either laparoscopic or open surgery. Our emulated trial included patients who underwent surgery within three months of confirmed colon cancer diagnosis. The target trial would randomize eligible participants to either treatment modality, stratified by tumor stage and hospital, in an unblinded manner, with randomization occurring one day before surgery. In the emulated study, patients were classified into the LC or OC group based on their registry records. Unlike in the target trial, patients in the emulated study were not randomly assigned; instead, randomness was emulated by controlling for baseline confounders using inverse-probability weights of treatment and censoring after cloning.24 In the target trial, patients were followed up from the date of assignment to either treatment modality until death from any cause (the outcome of interest), the last date of follow-up or the administrative censor date (31st December, 2023), whichever occurred first. In the emulated trial, follow-up started at time zero, that is, the time point when the eligibility criteria were met. Time zero was defined as the time point at which patients became eligible for the study, because eligibility, rather than the date of surgery, represents the appropriate time of cohort entry in the emulated target trial. This alignment of eligibility criteria, treatment assignment, and the start of follow-up prevents immortal time bias.23,24 Initiation of treatment could take place within a grace-period of the first three months following diagnosis of colon cancer. Follow-up ended on the date of death, last date of follow-up, or on the specified administrative censor date (31st December, 2023), whichever occurred first. OS was calculated as the time in months from time-zero to death from any cause. For patients who did not die, follow-up was right-censored at the time of last follow-up date or the administrative censoring date, whichever came first. As pre-specified in the study protocol, the emulation of the target trial was implemented using the cloning, censoring, and weighting (CCW) design, following the grace-period approach.27 Therefore, our causal estimand of interest in the emulated study is the observational-analogue of the per-protocol effect of LC versus OC performed in the first three months after a colon cancer diagnosis. Due to the cloning and censoring approach implemented following the grace-period, the observational-analogue of the intention-to-treat effect was not estimable in the emulated study.23,24 The causal interpretation of the emulated per-protocol effect relied on consistency, conditional exchangeability for treatment and censoring given measured covariates, and positivity.24,28 Statistical Analysis Target Trial Emulation A TTE with the CCW method was implemented to estimate the observational-analogue of the per-protocol effect.24 This involved first cloning each patient record to allow each record to enter both treatment arms during the grace-period. Then, clones were censored upon detection of non-adherence to the initially assigned treatment based on the observed data. Prior to cloning, a stabilized inverse-probability of treatment weighting (sIPTW) was estimated to give more weight to patients less likely to undergo LC, thus controlling for confounding.29,30 The following confounders were controlled for in the sIPTW model: age at colon cancer diagnosis (in years), sex, year of treatment, tumor laterality, hospital type (colorectal cancer center certified by the German Cancer Society or not, based on OncoMap data),31 performance status [the first ECOG score within 3 months after diagnosis], tumor grade, tumor stage, local tumor size, time to surgery, and the minimum number of lymph nodes harvested intraoperatively. Selection bias from this artificial censoring was also addressed using inverse-probability censoring weighting (IPCW), conditional on baseline covariates. The IPTW and IPCW weights were then combined (by multiplication) and applied in the Royston-Parmar parametric survival model (RPpsm) to estimate the causal effect of LC versus OC on OS via hazard ratio (HR). The RPpsm has several advantages over the traditional Cox proportional hazards model, especially its flexible feature to directly and smoothly model the baseline hazard function using restricted cubic splines.30,32 To account for the uncertainty during the weighting procedure, point estimates were accompanied by 95% confidence intervals (95% CI) obtained via non-parametric bootstrapping with 500 replicates. The HR was complemented by differences in restricted mean survival time (RMST) at 1-, 3-, and 5-years. A weighted Kaplan-Meier method was used to compare time-to-event survival curves according to surgical treatment modality over a 60-months period. Multiple imputation by chained equations (MICE) method was performed to impute missing values in ECOG, tumor stage, and tumor grade under the missing at random (MAR) assumption.33 A total of 10 imputed datasets (m = 10) were generated. Comparator Models Three traditional comparator analyses were conducted. Initially, the influence of the emulation process on the treatment effect was assessed by performing a multivariable Cox regression on the naïve (unemulated) study population, adjusting for the four key confounders (age, tumor grade, ECOG performance status, and tumor laterality) identified through the Directed acyclic graphs (DAGs) tool (Figures S1 and S2). The DAGs causal diagram, informed by literature review and expert discussion, was used to identify confounders, mediators, colliders, and prognostic factors (Table S2). Next, we modeled the outcome in the emulated and imputed data using two traditional (unweighted) Cox regression approaches: (1) a fully adjusted model that included all covariates and (2) a backward elimination strategy based on the 10% change-in-estimate (CIE) criterion approach (Tables S3 and S4). We used Stata’s coldiag2 program, which provides condition indices and variance-decomposition proportions (VDPs) for each variable, to assess multicollinearity (MC) in these traditional Cox proportional hazards models.34 As recommended by Belsley et al, a condition index of <10 and VDPs of <0.50 were considered indicative of acceptable levels of MC.35 These three traditional comparator methods, which are commonly used in the analysis of observational epidemiological data, were chosen to illustrate their performance in estimating the outcome in contrast to the TTE, thereby highlighting the methodological advantages of the more rigorous causal framework. They were also contrasted with two other trial emulation models, the sIPTW-only model and the doubly-robust (DR) model, in addition to the main TTE model described above.36 Similarity in causal effect estimate from the sIPTW-only model with the causal effect estimate from the DR model helps to assess the adequacy of the propensity score model specification.29 Subgroup and Sensitivity Analyses Subgroup analyses were performed by tumor stage and treatment period. We also conducted three sensitivity analyses: 1) excluding patients with missing data, 2) including patients with locally advanced colon cancer (stage T4N0-2M0) cases, and 3) using the E-value method37 to quantify bias from any unmeasured confounders like comorbidity. Because the E-value is typically calculated on the relative risk scale, we first converted the hazard ratio from the main TTE model to an approximate relative risk using VanderWeele’s formula, RR ≈ (1 − 0.5^√HR)/(1 − 0.5^√(1/HR)), before computing the E-value.37 We used the Cochran-Armitage test to assess the annual trend of LC over treatment years. Moreover, Pearson’s chi-squared test was used to assess associations between treatment group and other categorical covariates, and Fisher’s exact test was applied for categories with expected cell frequencies less than 5. We used a t-test to assess group differences in continuous covariates. All analyses were performed in Stata version 18.0 (StataCorp LLC, College Station, TX) and statistical significance was defined as a two-tailed p-value less than 0.05. This study was registered retrospectively at ResearchRegistry.com (UIN: researchregistry11770) and the German Clinical Trials Register (DRKS00038934). It has been prepared in line with the STROCSS 2025 guideline for reporting surgical observational studies.38 Results Study Population and Baseline Characteristics Figure 1 illustrates the eligibility criteria. A total of 13,594 solitary colon adenocarcinoma cases diagnosed between 2008 and 2024 were initially identified, with 191 cases of multiple tumors excluded, leaving 13,403 cases. To assess 5-year OS, the study focused on patients diagnosed and undergoing their first surgery between 2008 and 2018, excluding 6593 more recent cases. The most common reasons for exclusion during the study period were ineligible or unspecified tumor locations (n= 2036), ineligible surgical procedure (n= 1561), and metastatic diseases at diagnosis (n= 1107). Finally, the emulated study population comprised 1265 cases (9.3% of the initially identified patient population). The study population had a median age of 72 years (interquartile range [IQR]: 64–78), and the proportion of women was 46.4% (Table 1). Laparoscopic colectomy (LC) was performed in 380 patients (30%). The proportion of patients who underwent LC increased modestly over time, from 23.2% in 2008 to 35.6% in 2018, with a strong annual trend across treatment years (Table 1). LC was performed more frequently in younger patients, in cases with left-sided tumors, and in smaller tumor size and less advanced tumor stages compared to open colectomy (OC). The proportion of cases with fewer than 21 harvested lymph nodes was higher in the LC group than in the OC group. A total of 302 all-cause deaths were recorded during the follow-up (Table 1). Table 1Baseline Characteristics Between Eligible Patients in the Open and Laparoscopic Groups, 2008–2018 (N= 1265)CharacteristicsN (%)Open Colectomy (n= 885)Laparoscopic Colectomy (n= 380) ap-valuebSex0.512 Male678 (53.6)469 (53.0)209 (55.0) Female587 (46.4)416 (47.0)171 (45.0)Age at diagnosis, median (interquartile range)72 (64–78)73 (65–79)69 (62–76)<0.001Age category (in years)<0.001 18–4943 (3.4)28 (3.2)15 (3.9) 50–59141 (11.1)91 (10.3)50 (13.2) 60–69327 (25.8)199 (22.5)128 (33.7) 70–79482 (38.1)358 (40.5)124 (32.6) ≥80272 (21.5)209 (23.6)63 (16.6)Residence, at diagnosis0.322c In MV1234 (97.5)866 (97.9)368 (96.8) Outside MV31 (2.5)19 (2.1)12 (3.2)Diagnosis year<0.001 2008–2009153 (12.1)118 (13.3)35 (9.2) 2010–2014534 (42.2)391 (44.2)143 (37.6) 2015–2018578 (45.7)376 (42.5)202 (53.2)Treatment period<0.001d 2008–2009152 (12.0)117 (13.2)35 (9.2) 2010–2014531 (42.0)389 (44.0)142 (37.4) 2015–2018582 (46.0)379 (42.8)203 (53.4)Treating hospital0.377 Uncertified center693 (54.8)492 (55.6)201 (52.9) Certified center572 (45.2)393 (44.4)179 (47.1)Tumor laterality<0.001 Right685 (54.2)542 (61.2)143 (37.6) Left580 (45.8)343 (38.8)237 (62.4)ECOG Performance status0.960 0162 (42.1)106 (41.5)56 (43.1) 1160 (41.6)107 (42.0)53 (40.8) 263 (16.3)42 (16.5)21 (16.1) Missing880 (69.6)630 (71.2)250 (65.8)Grade0.672 Low161 (13.0)108 (12.5)53 (14.2) Intermediate907 (73.3)635 (73.6)272 (72.9) High168 (13.7)120 (13.9)48 (12.9) Missing28 (2.3)22 (2.5)6 (1.8)Tumor stage<0.001 I357 (28.2)215 (24.3)142 (37.4) II494 (39.1)371 (41.9)123 (32.4) III414 (32.7)299 (33.8)115 (30.3)Residual tumor0.171e No (R0)1239 (98.7)863 (98.4)376 (99.5) Yes (R1, R2, RX)16 (1.3)14 (1.6)2 (0.5) Missing10 (0.8)8 (0.9)2 (0.6)N-stage0.221 N0851 (67.3)586 (66.2)265 (69.7) N+414 (32.7)299 (33.8)115 (30.3)T-stage<0.001 T1156 (12.3)77 (8.7)79 (20.8) T2262 (20.7)180 (20.3)82 (21.6) T3847 (67.0)628 (71.0)219 (57.6)Harvested lymph nodes0.007 <21634 (51.3)422 (48.8)212 (57.1) ≥21602 (48.7)443 (51.2)159 (42.9) Missing29 (2.4)20 (2.3)9 (2.6)Time to surgery (days), median (interquartile range)10 (1–21)10 (1–21)12 (1–23)0.100Notes: Percentages of missing values were calculated based on all cases, whereas percentages of non-missing values were calculated using only the observed cases; statistical tests did not include the missing values; a40 patients underwent converted colectomy; bp-value calculated from the non-missing values; cp-value from Fisher’s exact test; dp-value calculated for linear test using the Cochran-Armitage test for annual treatment trend; ep-value from Fisher’s exact test; Median (interquartile range): p-value from a t-test; Frequency (%): p-value from Pearson test; MV refers to the German federal state of Mecklenburg-Western Pomerania. Baseline Characteristics Between Eligible Patients in the Open and Laparoscopic Groups, 2008–2018 (N= 1265) Notes: Percentages of missing values were calculated based on all cases, whereas percentages of non-missing values were calculated using only the observed cases; statistical tests did not include the missing values; a40 patients underwent converted colectomy; bp-value calculated from the non-missing values; cp-value from Fisher’s exact test; dp-value calculated for linear test using the Cochran-Armitage test for annual treatment trend; ep-value from Fisher’s exact test; Median (interquartile range): p-value from a t-test; Frequency (%): p-value from Pearson test; MV refers to the German federal state of Mecklenburg-Western Pomerania. Traditional Comparator Models: Analysis on the Naïve Data DAGs-guided Cox proportional hazards model was performed on the naïve (unemulated) data, adjusting for the minimally sufficient adjustment set of four confounders (Figures S1 and S2). This model included age, tumor grade, ECOG performance status, and tumor laterality in addition to the intervention variable. The analysis showed a strong protective effect of LC on the hazard of mortality compared to OC (HR= 0.66, 95% CI: 0.51–0.85). Traditional Comparator Models: Analyses on the Emulated Data Different traditional (unweighted) Cox proportional hazards models were used to analyze the treatment effect, comparing LC to OC. All analyses were conducted using imputed data on the emulated population (Table 2). The unadjusted (crude) Cox model indicated that LC was associated with a strongly lower hazard of all-cause death compared to OC (HR = 0.58, 95% CI: 0.44–0.76). The fully adjusted model, which excluded tumor grade and stage due to multicollinearity, indicated that LC was associated with a 32% lower hazard of mortality (HR = 0.68, 95% CI: 0.51–0.92). After applying the 10%-CIE method, the adjusted hazard ratio became 0.71 (95% CI: 0.53–0.96), suggesting a strong protective effect. Results from both models based on emulated complete-case data (n= 371) are also presented in Tables S5 and S6. Table 2Univariable and Multivariable Cox Proportional Hazards Model Using the Traditional Analytic Approaches (N= 1265)VariablesUnivariable Cox ModelMultivariable Cox ModelCrude HR(95% CI)Model 1 (Fully Adjusted):HR (95% CI)Model 2 (10% CIE Model):HR (95% CI)Surgical type Open1.001.001.00 Laparoscopic0.58 (0.44, 0.76)0.68 (0.51, 0.92)0.71 (0.53, 0.96)Mean age, in years1.79 (1.57, 2.05)1.55 (1.32, 1.82)1.47 (1.26, 1.72)Sex Male1.001.00 Female0.88 (0.70, 1.11)0.67 (0.52, 0.85)Hospital type Non-certified1.001.00 Certified0.89 (0.71, 1.11)0.98 (0.76, 1.25)Tumor laterality Right1.001.00 Left0.77 (0.61, 0.97)0.86 (0.66, 1.12)Treatment period 2008–20091.03 (0.71, 1.50)0.98 (0.65, 1.49) 2010–20141.19 (0.94, 1.52)1.17 (0.91, 1.51) 2015–20181.001.00Tumor stage I1.001.00 II1.99 (1.44, 2.76)1.70 (1.21, 2.39) III2.33 (1.68, 3.23)2.23 (1.59, 3.12)Grade Low1.00 Intermediate1.20 (0.83, 1.74) High1.48 (0.95, 2.31)Performance status 01.001.001.00 12.01 (1.22, 3.31)1.70 (1.01, 2.85)1.66 (0.98, 2.80) 24.26 (2.49, 7.30)3.04 (1.66, 5.56)2.98 (1.61, 5.50)Harvested lymph nodes <211.001.00 ≥210.99 (0.79, 1.24)1.06 (0.83, 1.35)Notes: Model 1= fully adjusted model after removing grade and stage due to multicollinearity; Model 2= 10% change-in-estimate criterion model with backward elimination strategy.Abbreviations: CI, confidence interval; CIE, Change-in-estimate; HR, Hazard ratio. Univariable and Multivariable Cox Proportional Hazards Model Using the Traditional Analytic Approaches (N= 1265) Notes: Model 1= fully adjusted model after removing grade and stage due to multicollinearity; Model 2= 10% change-in-estimate criterion model with backward elimination strategy. Abbreviations: CI, confidence interval; CIE, Change-in-estimate; HR, Hazard ratio. Target Trial Emulation Balance of Covariates Using Inverse Probability Treatment Weighting We applied stabilized inverse probability treatment weighting to the multiply imputed data of the emulated population (N= 1265) using baseline covariates. Our analysis showed evidence of adequate covariate balance between treatment groups after weighting, with all weighted standardized mean differences (SMDs) falling below 10% (Figure 2). This indicates successful mitigation of imbalances in confounders at baseline between the treatment groups and bestows solid justification for subsequent causal treatment effect analyses.29,39 A visual summary of the effectiveness of the weighting is provided in Figure S3. Balance of confounders based on SMDs for baseline characteristics. Treatment Effect: Results from the Main Target Trial Emulation Model The 5-year survival rates were 78.4% in the LC group and 74.5% in the OC group (Figure 3). In the main analysis, using the RPpsm with stabilized IPTW and IPCW, the estimated hazard of death was 19% lower with LC than with OC (HR= 0.81, 95% CI: 0.60–1.08), with the confidence interval ranging from a 40% lower to an 8% higher hazard. Weighted Kaplan-Meier curve comparing the 5-year OS by treatment strategy. Treatment Effect: Differences in RMSTs by Treatment Strategy The restricted mean survival time (RMST) at 1, 3, and 5 years of follow-up for laparoscopic versus open colectomy groups is presented in Table 3. The estimated RMST differences were in favor of the LC group at all time points, modestly increasing from 0.12 months at 1 year to 1.20 months at 5 years. However, the extent of these differences in effect was small and the respective confidence intervals were compatible with little or no difference in long-term survival between the two surgical treatment strategies. Table 3Estimates of Differences in Restricted Mean Survival Times in MonthsEffect MeasureLaparoscopic Surgery (95% CI)Open Surgery (95% CI)Difference Estimate (95% CI)1-year RMST (months)11.70 (11.50, 11.90)11.58 (11.40, 11.80)0.12 (−0.10, 0.34)3-year RMST (months)33.30 (32.70, 33.90)32.94 (32.30, 33.60)0.36 (−0.15, 0.87)5-year RMST (months)53.40 (52.30, 54.50)52.20 (51.00, 53.40)1.20 (−0.20, 2.60)Abbreviation: 95% CI, 95% confidence interval. Estimates of Differences in Restricted Mean Survival Times in Months Abbreviation: 95% CI, 95% confidence interval. Treatment Effect: Subgroup Analyses Subgroup analyses generally showed consistent treatment effects across tumor stages and treatment periods (Table 4). The hazard ratio estimates for tumor stages I–III were all close to the overall treatment effect, with confidence intervals that substantially overlapped the null value, indicating no meaningful evidence of variation in the treatment effect by UICC tumor stage. Similarly, treatment-period analyses yielded largely consistent findings across calendar periods, although patients treated during 2015–2018 showed a more favorable estimated effect for LC than those treated in earlier periods. Weighted Kaplan-Meier curves stratified by tumor stage are presented in Figure S4. Table 4Subgroup Analyses of Comparing Laparoscopic versus Open Colectomy by Tumor Stage and Treatment PeriodVariablesHR (95% CI) from the RPpsm*Tumor Stage I1.11 (0.55, 2.15) II0.66 (0.35, 1.07) III0.84 (0.47, 1.31)Treatment period 2008 - 20091.33 (0.49, 3.16) 2010 - 20141.08 (0.67, 1.59) 2015 - 20180.45 (0.26, 0.66)Note: *Bootstrapped 95% confidence interval (95% CI) estimates with 500 replicates.Abbreviation: HR, Hazard ratio. Subgroup Analyses of Comparing Laparoscopic versus Open Colectomy by Tumor Stage and Treatment Period Note: *Bootstrapped 95% confidence interval (95% CI) estimates with 500 replicates. Abbreviation: HR, Hazard ratio. Sensitivity and Additional Comparator Analyses for the TTE Method Multiple sensitivity and comparator analyses were conducted to evaluate the treatment effect of LC versus OC on the overall mortality hazard (Figure 4). Sensitivity analyses based on complete-case data, as well as those using imputed data that included the T4 cases, yielded attenuated effects consistent with the main model, with hazard ratios approaching the null value of 1 (Figure 4). The result of the third sensitivity analysis using the E-value method is presented in Figure S5. Results of sensitivity and comparator analyses. Blue text and lines indicate the main TTE model and the TTE sensitivity analyses; black text and lines indicate the comparator models, including the traditional and alternative TTE comparator analyses. Abbreviations: CIE, Change-in-estimate; DAGs, Directed acyclic graphs; pop, Population; RPpsm, Royston-Parmar parametric survival model; TTE, Target trial emulation. Additional comparator analyses for the TTE method, including the sIPTW-only model and the DR model based on RPpsm, yielded findings that were generally consistent with the main TTE results (Figure 4). Discussion Studying OS of two surgical treatment modalities in non-metastatic colon cancer, this research evaluated the performance of traditional observational analyses in contrast to the advanced TTE framework. Using the statewide data (2008–2018) from the Mecklenburg-Western Pomerania Cancer Registry, our study shows that the more rigorous TTE models give more accurate causal estimates compared to the traditional analytic techniques. Our main contribution is to illustrate how advances in transparent design and analytical methodology may address the shortcomings of traditional analytic methods. The results position TTE as a credible causal inference method for preventing avoidable bias in surgical effectiveness research based on observational data. The estimated causal effect of LC versus OC in this study generally accords with the long-term findings of landmark RCTs, which have reported similar survival following the two surgical procedures.3,7–14 For clinical practice, this suggests that the choice of the surgical approach should not be based on the expectation of superior long-term survival alone. Clinically more important are the established short-term advantages of LC alongside patient characteristics, surgeon expertise, and local care resources.2,4–6 Population-based studies from Germany and the United States have reported larger survival advantages for LC, but these estimates may partly reflect treatment selection and other biases that traditional observational analyses cannot fully address.21,22 The consistency of our findings with evidence from RCTs conducted across different countries and healthcare systems supports their potential relevance beyond Germany, although transferability may vary with case mix (ie, patient and tumor characteristics), surgical experience, and access to minimally invasive surgery.3,5–14,21,22 TTE can help prevent avoidable design-related biases but cannot overcome limitations of the underlying data, including bias from unmeasured confounding.23 Accordingly, our results should be viewed as complementary real-world evidence supporting comparable long-term OS in routine care, rather than as independent evidence that either procedure is universally preferable.23,24,26 We began our analysis with the simplest traditional analytical method, applying the DAGs-guided Cox regression model on the naïve (unemulated and unweighted) data (Figures 4 and S2). This model serves a dual methodological purpose: on the one hand, it helps to assess the influence of our emulation process by providing a benchmark using the unemulated data, and on the other hand, it controls for overadjustment bias through principled covariate selection (Figure S2). By identifying colliders and mediators that should not be controlled for and identifying the minimum sufficient set of confounders required for unbiased estimation of the total effect of surgical treatment modality on 5-year OS, the DAGs tool provides a critical advantage in helping researchers avoid adjustment-induced bias.40 However, this model gave an estimate that differed substantially from the main TTE estimate, indicating possible bias from other sources (Figure 4). In the next levels of analyses using both fully adjusted and CIE-based traditional methods, we performed multivariable Cox regression modeling on the emulated (complete-case and imputed) data (Table 2, Figure 4, and Tables S3 to S6). These traditional analytic techniques are common in epidemiological literature and provided us with baseline estimates for methodological comparison. However, estimates from these two traditional models still differed substantially from the main TTE-based estimate, again suggesting bias. This underscores the importance of combining a valid trial emulation design with robust analytic methods to address biases commonly encountered in observational data analysis.23 In summary, the traditional methods consistently showed a strong protective effect of LC on all-cause mortality compared to OC. These findings were in agreement with nationwide population-based observational studies, which reported higher survival associated with LC compared to OC.21,22 It might be tempting to dismiss these findings as spurious, given the strong consistency across the models and with the literature based on big observational data.21,22 However, the improved survival for the laparoscopic group in these observational studies was most likely due to selection bias, as acknowledged by the authors.22 Indeed, traditional analyses of observational data can yield misleading results due to their susceptibility to various biases, including immortal time bias and selection bias.18–20 Such risks of bias in any observational data justify the use of the more robust TTE framework.23 Building on the traditional fully-adjusted and CIE models, which were both based on emulated data in this study, we performed two more rigorous causal inference models within the TTE framework, the sIPTW-only and the DR models (Figure 4). Both models found comparable effects of LC and OC on 5-year survival. The DR model combines the outcome modeling with the treatment (propensity score) modeling simultaneously and provides an unbiased causal estimate as long as either model is correctly specified.36 Its property indeed gives “two chances” to obtain an unbiased causal effect estimate of the treatment if one of its models is correctly specified.28 In this study, the results obtained from the sIPTW-only and DR models were nearly identical. This confirms adequate propensity score specification in the sIPTW-only model, indicating successful mitigation of model misspecification bias as well as effective control of selection bias.28,29,36 Based on the emulated data, the main RPpsm model and the traditional models differed in the magnitude and precision of their treatment-effect estimates, demonstrating the substantial impact of appropriate weighting on the treatment effect estimation (Figure 4). This effect difference is in line with previous findings that traditional regression-based covariate adjustment often yields different results than weighting-based methods.41 This shows that, although the traditional analytical approaches address some of the existing confounding, they alone may not fully account for selection bias and immortal time bias. Additionally, the observed differences in mortality hazard between the DAGs-guided model using naïve data and the main RPpsm TTE model using emulated data in this study align with the explanation that explicit emulation of a hypothetical pragmatic target trial can reduce avoidable biases inherent in real-world data.24,42 Meticulous adjustment for confounding and bias mitigation through well-specified weighting as well as transparent and valid emulation procedures, which approximate random treatment assignment in observational data analysis, are hallmarks of trial emulation. This entails combining the successful emulation of the design elements of an explicitly specified hypothetical pragmatic target trial with rigorous causal inference techniques, such as inverse probability weighting.24,43 As Hernán and Robins have emphasized, this approach helps reduce the preventable biases that plague observational studies and enables the estimation of causal effects with validity close to that of RCTs.24 Common limitations, such as missing values for important confounders (for instance, ECOG was absent in 69.6% of cases in this study) or the complete absence of measurement for pertinent confounders like comorbidity and surgeon’s expertise, frequently impede epidemiological investigations of oncological outcomes based on German cancer registry data.21 Our study has apparently addressed the first challenge through MICE.33 Handling missing values through multiple imputation, under the MAR scenario, provides more robust and precise estimates.44 This approach allowed us to increase the database by more than three-fold; without any imputation, the analysis would have been limited to only 371 cases, leading to a reduced statistical power and greater imprecision in the estimated treatment effect (Tables S5 and S6). However, multiple imputation cannot recover confounders that were never measured. The absence of comorbidity and patient-level information on surgeon expertise may have influenced both the selection of surgical approach and 5-year OS, because patients’ underlying health status may affect their suitability for LC, whereas surgeon experience and operative volume may influence procedure selection and subsequent outcomes.21,45,46 We therefore evaluated the sensitivity of the findings to unmeasured confounding using the E-value method. VanderWeele and Ding have underscored the importance of reporting the E-value as a standard practice in observational studies, as it enables evaluation of the robustness of a study’s findings.37 An E-value of 1.58 for the point estimate indicates that an unmeasured confounder, or a combination of unmeasured confounders acting jointly, would need to be associated with both procedure selection and 5-year OS by risk ratios of at least 1.58 each, conditional on the measured covariates, to move the observed point estimate to the null (Figure S5). An influence of this magnitude on both the procedure selection and the outcome is unlikely. This is particularly so because our adjustment for age, ECOG performance status, and hospital certification status in this study probably has captured some differences in underlying health status and surgeon’s expertise. However, residual bias cannot be fully excluded. This could be better addressed in future studies by linking cancer registry data with statutory health insurance data and, where available, detailed surgical data.47 This study has also assessed the robustness and comparability of the treatment effect through several sensitivity and comparator modeling approaches. The TTE-based sensitivity analyses also supported the robustness of the main finding (Figure 4). The traditional comparator models showed a spurious survival advantage of LC over OC for 5-year OS, whereas the more rigorous comparator models (sIPTW-only and DR) yielded results generally consistent with the main TTE model, which also aligns with many RCTs on this topic. The transparent principles and rigorous methodological approach of the TTE can be employed to reliably answer pressing surgical research questions using the rich data available in the German cancer registries. Therefore, we recommend the use of the TTE design in cancer registry data analyses to provide more reliable answers for comparative effectiveness research questions in surgical oncology, especially when evidence from RCTs is lacking. This may help guide clinical practice and improve patients’ long-term oncological outcomes. Conclusion In conclusion, this study found broadly comparable 5-year OS among patients with stage I–III colon cancer undergoing LC and OC. This finding is consistent with the available RCT evidence and suggests that LC is as safe as OC. The traditional observational analyses however indicated more favorable effect estimates for LC than the TTE analyses, illustrating how flaws in study design and adjustment for selection and immortal time biases can practically distort treatment-effect estimates. In this population-based study, the TTE framework provided complementary real-world evidence on the comparative effectiveness of LC and OC for 5-year OS. This causal design framework may strengthen effectiveness research in surgical oncology using routinely collected data, particularly when RCTs are unavailable, infeasible, or have limited generalizability. However, causal interpretation still depends on the quality of the underlying data, how well the target trial is emulated, and whether the identifying assumptions are sufficiently met. TTE should therefore complement, rather than replace, well-conducted RCTs. Acknowledgments The authors thank cancer patients and physicians who contributed to the cancer registry by reporting patient data. Funding Statement Support for the publication fee was provided by the University of Greifswald’s Publication Fund. Study Registration The study was retrospectively registered with both the Research Registry (UIN: researchregistry11770) and the German Clinical Trials Register (DRKS00038934). Abbreviations adjHR, adjusted Hazard Ratio; CI, Confidence Interval; CIE, Change-In-Estimate; DAGs, Directed Acyclic Graphs; DR, Doubly-Robust; ECOG, Eastern Cooperative Oncology Group; E-value, Evidence-value; HR, Hazard Ratio; IPTW, Inverse Probability Treatment Weighting; IPCW, Inverse Probability Censoring Weighting; IQR, Interquartile Range; MC, Multicollinearity; MAR, Missing at Random; MICE, Multiple Imputation by Chained Equations; MV, Mecklenburg-Western Pomerania (Mecklenburg-Vorpommern); OS, Overall Survival; RCTs, Randomized Controlled Trials; RR, Relative Risk; RPpsm, Royston‒Parmar parametric survival model; RMST, Restricted Mean Survival Time; SMD, Standardized Mean Differences; sIPTW, Stabilized Inverse Probability Treatment Weighting; TTE, Target Trial Emulation; UICC, Union for International Cancer Control; VDPs, Variance-decomposition Proportions. Data Sharing Statement The data underlying the findings of this study were obtained from the Mecklenburg-Western Pomerania Cancer Registry. The data use agreement does not permit the authors to share the data. However, the data may be obtained from the corresponding author upon reasonable request and with the consent of the Mecklenburg-Western Pomerania Cancer Registry. Ethical Approval and Consent to Participate The German Guidelines for Good Epidemiological Practice (Leitlinien und Empfehlungen zur Sicherung von Guter Epidemiologischer Praxis, GEP) and the Guidelines for Good Practice in Secondary Data Analysis (Leitlinien und Empfehlungen zur Guten Praxis Sekundärdatenanalyse, GPS) state that formal ethical approval is generally not required for analyses conducted exclusively on routinely collected secondary data, provided no patient contact, clinical intervention, or access to directly identifying information occurs. This study fulfils all of these conditions. The legal basis for the use of cancer registry data for this research was the Mecklenburg-Western Pomerania Cancer Registration Act (Krebsregistrierungsgesetz Mecklenburg-Vorpommern; KrebsRG M-V). No individual informed consent was required, as data processing and secondary use of pseudonymized data for scientific research are explicitly permitted under this law. The data-use application for the present study was approved in advance by the Advisory Board of Mecklenburg-Western Pomerania Cancer Registry. Disclosure All authors declare that they have no competing interests. References Associated Data Data Availability Statement The data underlying the findings of this study were obtained from the Mecklenburg-Western Pomerania Cancer Registry. The data use agreement does not permit the authors to share the data. However, the data may be obtained from the corresponding author upon reasonable request and with the consent of the Mecklenburg-Western Pomerania Cancer Registry.