*4.2. Statistical Analysis*

To avoid bias in the univariate and multivariate analyses, variable categories containing a small number of events, whose statistical results were unreliable, were merged with larger categories. Thus, in accordance with TNM-7, we divided tumor sizes (T) in T1 + T2 and T3 + T4, and tumor stages (S) in S1 + S2 and S3 + S4. The tumor sizes as defined by the TNM-8 were classified in T1 + T2 + T3a and T3b + T4, and the tumor stages in S1 and S2 + S3 + S4. Univariate analyses were executed for each clinical parameter in relation to the presence/absence of recurrence. Statistical comparison between recurrent and non-recurrent patients was carried out with the Chi-square test or the Fisher exact test for categorical variables, and with the Mann–Whitney test for continuous variables (not normally distributed). Where a significance was found in the Chi-square test, the Cramer's V was used as a post-test to determine the strength of association between variables. Kaplan–Meier curves were created to estimate the DFI in patient groups, and differences were statistically evaluated by the log-rank test. Finally, Cox regression was performed to quantify the hazard ratios of several explanatory variables, both continuous and categorical. The selection of covariates was made by including all parameters with a *p*-value < 0.25 in the univariate analysis, together with those of known clinical significance. Proportional hazards assumption and absence of multi-collinearity were preliminarily assessed for these covariates. The backward stepwise approach was used for the model selection. Data analyses were done by using the SPSS software for Windows (SPSS, Inc., Chicago, IL, USA), considering the probability value <0.05 as the threshold limit for statistical significance.
