Cognitive functioning over time

We conducted linear mixed model (LMM) analyses to investigate the course of cognitive performances over time (one model per cognitive test), initially in the total patient sample. In the longitudinal LMM, time (T0, T3, T12) was level 1 and its measurements were nested in the patients at level 2. Because only three time points were involved, we adopted a linear effect of time for all models. Intercepts were specified as random effects, allowing for individual estimations of the data of each patient. Random slopes were added to those models if they significantly improved model fit (likelihood ratio test, α = 0.05). Among the tested correlation structures (autoregressive, continuous autoregressive, compound symmetry, general correlation matrix, scaled identity), the one providing the best fit based on the Akaike information criterion (AIC) for the majority of the models was adopted uniformly.

First, we created models with only time as predictor to investigate the overall course of performances without any other predictors. In the final models, we included a time*carrier status interaction (non‐carrier as reference group). We also included a time*diagnosis interaction to account for possible differences in performance over time between meningioma and glioma patients (glioma as reference group). Similar models were constructed to investigate the effect of carrier status for the T0−T3 interval and T3−T12 interval separately, using time as factor instead of a continuous variable (no random slopes). Within the group of patients who received adjuvant treatment, regardless of diagnosis, we performed ancillary analyses, again of time*carrier status.

We used the restricted maximum likelihood (REML) algorithm to estimate model parameters. Global fits of the models (models with only time as predictor vs the final models) were compared using AIC, and tested with likelihood ratio tests in case of a significant effect of carrier status. Analyses of the data [44] were performed using SPSS software (version 24) and Rstudio software (lme4 and nlme packages [45, 46]). We adopted a correction for multiple testing (taking into account the 10 tests we performed to investigate all cognitive measures) per main analysis (pretreatment performance, posttreatment change with time only, and posttreatment change with APOE carrier status) using the false discovery rate (FDR) correction procedure of Benjamini and Hochberg (BH) [47] (original α = 0.05).