4.2. Data Analysis

The primary objective of the meta-analysis was to explore the influence of seven commonly available baseline clinical, demographic or genetic variables (gender, age, ethnicity, smoking habit, ECOG PS, brain metastasis, EGFR mutation) on PFS-HR and/or OS-HR (efficacy outcomes) reported in RCTs of EGFR-TKIs in NSCLC patients with activating EGFR mutations. PFS-HRs and OS-HRs for each modality of the predictor variables were reported for each trial. We evaluated heterogeneity among studies using the χ² Q test and I² statistics. For the Q test, significant heterogeneity was declared if p < 0.05, while I² values > 50% were considered to indicate evident heterogeneity. Pooled PFS-HR and pooled OS-HR were calculated using random-effects models. The pooled ratio of the PFS-HRs and OS-HRs were reported together with their corresponding 95% CI. Results were also graphically displayed as a forest plot.

To explore whether the predictor variables may influence PFS and OS, an interaction test was performed following the approach reported by Fisher et al. [13]. This approach avoids the risk of ecological bias in testing heterogeneity among groups, by computing within-trial interaction as the ratio of the reported HRs in the two groups, and then these trial-specific interaction HRs are pooled across trials using a random-effects model.

The secondary objective of the meta-analysis was to explore whether the interactions associated with the primary objective were influenced by setting, EGFR-TKI generation, and type of comparator arm. These three variables identifying subgroups of RCTs as specified above were included in the model as a moderator to test if some heterogeneity among the trial-specific interaction HRs may be due to their influence.

Reporting bias was evaluated by assessing visual asymmetry on funnel plots of global HRs against standard errors. To examine whether the association between effect sizes and the related standard errors was greater than expected to occur by chance, the regression test for funnel plot asymmetry was carried out. Since tests for funnel plot asymmetry typically have low power, results must be interpreted with caution.

As regards PFS-HR, funnel plots and corresponding regression tests were also performed considering setting, EGFR-TKI generation, and type of comparator arm as moderator variable. For models involving moderators, the residuals were analyzed instead of effect sizes. The statistical software R version 3.2.5 (13) was used for all statistical analyses. Meta-analysis was performed using metafor package, version 2.1–0, with p < 0.05 considered as statistically meaningful. Prof. Dolce, a biostatistician at the Department of Public Health of University Federico II of Naples, was responsible for the statistical analysis, which was internally reviewed for accuracy by a senior biostatistician of the Department.