The costs associated with sensory evaluation increase with the number of panelists to be enrolled. Classical power computation can be used to derive the minimal number of subjects of a sensory panel in order to control both type I (α risk) and type II (β risk) errors. However, this power computation requires estimates of the size of the product effect to be sought and of the residual variability of the ANOVA model used. Generally, both product effect size and residual variability are difficult to estimate a priori by the sensory analyst. This work offers estimations of these two parameters thanks to the analysis of hundreds descriptive andhedonic studies collected respectively in two databases, SensoBase and PrefBase. The meta-analysis of the data allowed to quantify these two parameters and made possible the calculation of the number of panelists. Hence, tables of panel sizes were proposed for 3 levels of respectively product effect size, residual variability and type I and II errors. Of course, this was done independently for descriptive and hedonic tests.Another approach based on resampling in numerous datasets was applied for both descriptive and hedonic studies. The method used to derive adequate panel size consisted in removing k subjects from the N of the original panel and then measuring the loss of information in product comparisons. For descriptive panels, panel size could be reduced by a quarter but this reduction strongly depends on the type of attributes. For hedonic panels, panel sizes varied extremely and depended mainly on the size of the liking differences between products to be compared. We expect that this difference is directly affected by the level of sensory complexity of the products. Finally, the resampling approach was applied to examine the need to replicate with trained sensory panels. Results suggested that replicates are no longer necessary at the testing phase, that is once the panel is trained
