10
Problematic
• The number of subjects needed for sensory and consumer tests is an important and redundant problematic for most companies using sensory analysis routinely.
• Anyone can find norms and tables in literature proposing guidelines in terms of panel size (e.g. AFNOR NF V 09-500, 2012), but these advices are finally empirical and
“a priori” defined. It cannot really be adapted to the specific objectives and products of each company and study.
• In this context, bootstrap (Efron, 1979) can be a very useful resampling technique enabling to extract “a posteriori” information of your data about the panel size.
Example of Results
o Example for steaks: 6 products, 84 consumers, 12 discriminated pairs in the original dataset o Example for dairy products: 12 products, 102 consumers, 39
out of the 15 possible (percentage of discrimination: 80%) discriminated pairs in the original dataset out of the 66 possible (percentage of
discrimination: 59%)
Number of discriminated pairs
Good panel size?
o The mean number of discriminated pairs is not o With 95% confidence intervals:
linear. • A significant loss of information
o With 95% confidence intervals: occurs with a panel size of 40
• A significant loss of information occurs with a consumers.
panel size of 16 consumers only.
• But with less than 40 consumers, we could be o What about 50 or 60 consumers?
afraid to loose to much information… Significant loss
Significant loss
of information
of information
o With around 50 consumers (42% economy), we
would have good chances to reach the same
information in terms of discriminated pairs. If consumer tests are conducted
routinely, meta-analyses are
Product F-value possible to determine your own
guidelines in terms of panel sizes.
Discussion
Advantages Limits Perspectives
The method may be applied to every kind of hedonic Bootstrap is time-consuming and depends on your Adaptation to other sensory and marketing
assessment, emotion measurements and so on… software and hardware protocols (discrimination and similarity tests,
Bootstrap is a non-parametric method Approximately 20 minutes for 1,000 bootstraps sensory profile, temporal methods…)
no assumption on the distribution of the data By now, only one dataset and one descriptor studied Definition and optimization of cost/loss of
(specially in case of consumers segmentation) at each analysis information criteria
Meta-analyses of several datasets and descriptors
R-Package?