As studies have underlined the sensitivity of lidar metrics to scan angles, the objective of this study was twofold. Firstly, we further investigated the influence of lidar scan angle on the ABA predictions of stand attributes of riparian (29 field plots), broadleaf (42 field plots), coniferous (31 field plots) and mixed (45 field plots) forest types in France. Secondly, we evaluated the potential of voxelisation approaches to normalise scan angle effects in lidar metrics and mitigate scan angle effects in ABA models. To achieve these objectives, we first selected a model based on four lidar metrics with different sensitivities to lidar scan angle, i.e. mean and variance of canopy height values, gap-fraction, and coefficient of variation of plant area density (PAD) profile. For each plot, we considered the point cloud scanned from one flight line independently and characterised each resulting point cloud by the mean scan angle (MSA) and classified them into one of three classes: A (0° <=MSA < 10°), B (10°<=MSA < 20°) or C (20°<=MSA < 30°). An experimental setup involving nine scenarios was conceived to study the impact of the number of flight lines (scenarios fl1, fl2 and fl3) and predominant scan angle (scenarios A, B or C) or combination of scan angle directions (scenarios A and B, or A and C, or B and C), on area-based approach (ABA) models. We built ABA models for the same forest plots for 5000 resampled datasets in each scenario to predict three forest attributes, i.e., stem and total volume (Vst and Vtot) and basal area (BA). Three goodness-of-fit criteria were computed for each model (coefficient of determination (R2), relative root mean square error (rRMSE) and mean percentage error (MPE). We compared the distributions of the goodness-of-fit criteria between scenarios to assess the behaviour of the predictive models when: 1) the number of flight lines (i.e., scan angles) increases (fl1, fl2 or fl3); 2) lidar datasets comprise specific scan angle (A, B or C) or combination of scan angles (AB, AC or BC); 3) voxelisation is used to compute Pf and CVPAD. The results show that models built with point clouds scanned from multiple flight lines were more robust, with a lower standard deviation of their goodness-of-fit criteria. On average, across all forest types, compared to fl1, the standard deviations of R2 distributions were lower for fl2 and fl3 by 42 % and 77 %, respectively. We also observed that a dataset with a predominantly nadir configuration (i.e., scenario A) did not always result in better predictions (mean R2 higher by 0.08, 0.07, 0.04 for scenario B for broadleaf, coniferous and mixed, respectively). For a set of calibration plots, the resulting forest attribute models depend on the acquisition geometry over the plots, as observed in this study, which could result in unreliable wall-to-wall predictions. The risk is particularly high in acquisitions with low overlapping rates, with many areas covered by only one flight line. Using voxel-based Pf and CVPAD together with the mean and variance of heights helped to mitigate the impacts of changes in scan angles by a) increasing the means of the distributions, thereby improving the accuracy of predictions, or b) reducing the standard deviations, thereby increasing prediction precision, or c) both of the above.< Réduire