Statistical Analysis
StataTM Statistical Software v. 14 (StataCorp,
College Station, Texas) was used for consolidation, uniform recording
and basic descriptive statistics. Survival analyses and output graphics
were conducted with R Statistical Environment v. 3.5.2 (R Core
Team, Vienna, Austria). The Kaplan-Meier procedure was used to estimate
overall survival (OS) curves and survival at specific time points (1, 2
and 5 years) with associated 95% confidence intervals (26, 27). The
log-rank test was applied to assess the existence of significant
differences between survival curves in different subgroups. Multivariate
Cox regression models were fitted to analyse the association of the OS
and the event-free survival (EFS) with a series of potential prognostic
factors (28). Non-linear effects were analysed by introducing the
relevant predictors as penalised smoothing splines (29). This method was
employed to assess whether there may be an age threshold beyond which
patients had a less favourable outcome and also to establish a similar
threshold for serum AFP level at diagnosis, which influenced OS or EFS.
The Spearman test was used to establish an association between ordinal
variables, in this case SES and stage. The proportional hazards
assumption was tested by computing the scaled Schoenfeld Residuals test
(30). A p-value <5% was used as a cut-off to define
statistical significance. P-values for relative hazards were calculated
in a post-hoc analysis comparing patients with AFP levels above and
below the identified thresholds in a Cox model controlling for all other
potential confounding factors.