*2.3. Statistical Analyses*

Data were summarized by mean ± standard deviation (SD) for continuous variables and through natural frequencies and percentages for categorical ones.

Treatment responses were reported as objective response rate (ORR) and disease control rate (DCR). The time-to-event endpoints examined were progression-free survival (PFS) and overall survival (OS). PFS was defined as the time from start of first line treatment (or from start of osimertinib for the subgroup analysis) to disease progression or death for any cause, whichever occurred first. Patients who were alive and progression-free at 30 June 2018, the last follow-up update, were censored at that date.

OS was defined as the time from start of first line treatment to death for any cause. Alive patients were censored at the date of the last follow-up update. PFS and OS functions were estimated using the Kaplan–Meier method, and the log-rank test was used to assess differences between groups. Median PFS and OS were reported as point estimates and 95% confidence intervals (CI) in round brackets. The Cox proportional hazards regression model was used to quantify the association between specific covariates and the time-to-event endpoints. Results are reported as HR and 95% CI in round brackets. To assess the association between mutations and the duration of response to TKIs, patients were divided into short-term responders (PFS less than 6 months), intermediate-term responders (PFS ≥6 months and ≤24 months) or long-term responders (PFS >24 months). The association between categorical variables was tested by the Pearson's χ<sup>2</sup> test or Fisher exact test, when appropriate, whereas those between a continuous variable and a categorical one was tested by means of the Student *t*-test or *F* test for more than two categories. To evaluate the independent role of *TP53* mutations in a multivariate analysis and to obtain more accurate estimates of their prognostic effect, a combined analysis including

data of the present work with those from our previous one, was performed. Follow-up of our previous cohort was updated on 30 June 2018. A multivariable model was obtained using backward stepwise variable selection, setting the significance level for variable removal from the model equal to 0.10. In a perspective of parsimonious modelling, when appropriate, categories of some study variables were grouped. The proportional hazards assumption was evaluated using a statistical test based on Schoenfeld residuals. In case of non-proportional hazards for a specific variable, a Cox model with time-dependent coefficient, β(t), was fitted. To simplify model interpretation, a step function for β(t) was used, dividing the follow-up period in three time periods since treatment started: the first 6 months, 6–12 months, and greater than 12 months.

Overall and when not otherwise specified, a two-sided *p*-value (*p*) <0.05 was considered statistically significant. All statistical analyses were performed using STATA 15.0 software (College Station, TX, USA) and R version 3.6.1.
