Tableone

Categorical variables were compared using the Chi squared test. Non-normally distributed data were analyzed using the Mann–Whitney U test. survival was estimated using Kaplan–Meier survival curves and compared using the log-rank test. Multivariable analyses used Cox proportional hazards models. Stratified survival analyses by underlying histology (adenocarcinoma and squamous cell carcinoma) and by response to neoadjuvant therapy classification were performed. Analyses were also performed according to degree of downstaging (> 3 stages, 3 stages, 2 stages, and 1 stage). A p value of < 0.05 was considered to be statistically significant. Data analysis was performed using R Foundation Statistical software (R 3.2.2) with TableOne, ggplot2, Hmisc, Matchit, and survival packages (R Foundation for Statistical Computing, Vienna, Austria) as previously described.17

Categorical variables were compared using the Chi square test, and non-normally distributed data were analyzed using the Mann–Whitney U test. survival was estimated using Kaplan–Meier survival curves and compared using the log-rank test. Multivariable analyses used Cox proportional hazards models (“Appendix 1”). A comparison of outcomes between 5-year periods (1989–1993, 1994–1998, 1999–2003, 2004–2008, 2009–2013, 2014–2018) was also performed. For the final cohort, patients were included up to January 2017 to allow for a minimum 3 years of follow-up. A p value < 0.05 was considered to be statistically significant. Data analysis was performed using R Foundation Statistical software (R 3.2.2) with TableOne, ggplot2, Hmisc, Matchit and survival packages (The R Foundation for Statistical Computing, Vienna, Austria), as previously reported.
As a review of past practice and outcomes, this study was deemed exempt from the need for ethical approval.

Categorical variables were compared using the chi-square test. Non-normally distributed data were analyzed using the Mann-Whitney U test. Survival was estimated using Kaplan-Meier survival curves and compared using the log-rank test. A multilevel logistic regression model was used to produce an adjusted odds ratio (OR) and a 95% confidence interval (95% CI) to determine the association between surgical approach and TO. Multivariable analyses used Cox proportional hazards models.
In all models, patient-level, hospital-level, and tumor-level characteristics were included. Importantly, year of diagnosis was included to adjust for changes in developments in patient selection, diagnostic staging, multimodality treatment, and perioperative care (e.g., prehabilitation). Subset analyses were performed in high-volume centers to assess the impact of open, laparoscopic, and robotic techniques on TO and survival for both esophagectomy and gastrectomy patients. Post hoc analyses also were performed to compare outcomes for laparoscopic and robotic techniques.
A p value lower than 0.05 was considered statistically significant. Data analysis was performed using R Foundation Statistical software (R 3.2.2) with TableOne, ggplot2, Hmisc, Matchit and survival packages (R Foundation for Statistical Computing, Vienna, Austria), as previously reported.¹⁸

We used Cronbach’s α to evaluate consensus quantitatively among the international expert panel; a Cronbach’s α value of at least 0.80 was representative of an acceptable measure of internal reliability.^{20 (link)–23 (link)} Categorical variables were compared using the χ² test. Non-normally distributed data were analyzed using the Mann-Whitney U test for comparisons across 2 groups, and the Kruskal-Wallis test for comparisons of more than 2 groups. Stratified analyses were performed for responses from the second voting round by: annual department volume (≤50, 51–100, ≥101 procedures) and annual surgeon volume (≤20, 21–50, ≥51 procedures). A P value of <0.05 was considered statistically significant and no adjustments were made for multiple comparisons. Heat maps were developed to display the level of consensus (ie, green: ≥80% agreement, yellow: 70%–80%, and red: <70% agreement) across the different research questions.^{24 (link)} Data analysis was performed using R version 3.2.2, with TableOne, ggplot2, Hmisc, Matchit, and survival packages (R Foundation for Statistical Computing, Vienna, Austria).

Categorical variables were compared using the Chi square test. Non-normally distributed data were reported as median and interquartile range (IQR) and were analyzed using the Mann–Whitney U test. Survival was estimated using Kaplan–Meier survival curves and compared using the log-rank test. A multivariable Cox regression model was developed to model clinically relevant variables predictive of long-term survival. The multivariable analysis included those variables deemed potentially clinically relevant. A p value < 0.05 was considered statistically significant. Data analysis was performed using R Foundation Statistical software (R 3.2.2) with TableOne, ggplot2, Hmisc, and survival packages (The R Foundation for Statistical Computing, Vienna, Austria), as previously described.15 (link),16 (link)

Categorical variables were compared using the chi-square test. Non-normally distributed data were analyzed using the Mann-Whitney U test. Survival was estimated using Kaplan-Meier survival curves and compared using the log-rank test. Multivariable analyses used Cox proportional hazards models. The conditional probability of receiving different treatment options (esophagectomy vs gastrectomy), as indicated by the propensity score, was estimated using a multivariable logistic regression model including all the variables listed in Table S2. Next, balanced cohorts using nearest-neighbor propensity score-matching (PSM) without replacement (caliper width 0.1 standard deviation) were developed.23 (link) Balance diagnostics were performed using standardized mean differences, with a value lower than 0.1 indicating good balance.23 (link) The overall survival (OS) of the matched patients who received the aforementioned treatment options was evaluated. A p value of lower than 0.05 was considered to be statistically significant. Data analysis was performed using R Foundation Statistical software (R 3.2.2) with TableOne, ggplot2, Hmisc, Matchit, and survival packages (R Foundation for Statistical Computing, Vienna, Austria), as previously reported.24 (link)

Categorical variables were compared using the Chi-square test, and non-normally distributed data were analyzed using the Mann–Whitney U test. Survival was estimated using Kaplan–Meier survival curves and compared using the log-rank test. Multivariable analyses used Cox proportional hazards models. The conditional probability of receiving AC, i.e. the propensity score, was estimated using a multivariable logistic regression model including all variables listed above. We then created balanced cohorts using 1-to-1 nearest-neighbor PSM without replacement (caliper width 0.1 standard deviations).23 (link) Balance diagnostics were conducted using standardized mean differences, with a value < 0.1 indicating good balance.23 (link) Sensitivity and interaction analyses were performed by nodal status (i.e. N0, N1, N2/3), margin status (i.e. R0, R1), and receipt of NART on long-term survival. A p-value of < 0.05 was considered statistically significant. Data analysis was performed using R Foundation statistical software (R 3.2.2) with TableOne, ggplot2, Hmisc, Matchit, and survival packages (The R Foundation for Statistical Computing, Vienna, Austria) as previously described.24 (link)^,25 (link) This study was exempt from Institutional Review Board approval.