Results

It remains very difficult in sensory analysis to eliminate a judge with an answer outside the panel in a descriptive profile. Under the assumption of a normal or Gaussian distribution, the results of a panelist must not be more than two standard deviations by simplification, below or above the values of the mean of the panel.

The regular evaluation of the judges is very similar to the procedures for calibrating the measuring instruments used in a laboratory. This evaluation therefore proves to be necessary to carry out fairly regularly in order to ensure that the panelists ensuring the sensory measurement of the company remain appropriate, that is to say discriminating, and repeatable in their evaluation of the products to be described.

The distribution of statistical scores is essential in selecting the type of statistics to use. These statistics are broadly classified into two groups according to the distribution of the scores, either the distribution is said to be 'normal' (similar to a bell curve or even of the Gaussian type), and then it is possible to use the classic statistics based on means and standard deviations, such as Student's tests and analysis of variance; either the distribution is identified as 'non-normal' (di-symmetric, multimodal, crushed, etc.), and then particular 'non-Gaussian' type hypothesis tests are to be used, such as Willcoxon tests or Friedman tests, based essentially on rows or schedules.

Statistical processing highlights avenues for analysis and interpretation, and particularly by using multidimensional methods; However, it will always be necessary to check that these first conclusions or lines of analysis are not due to one or more "aberrant" individuals who have misinterpreted a question or who have under- or over-scored incoherently, and which, by this fact, would have shifted the results very strongly in one direction or another.

A missing value is often due to an error of attention of a judge or to an impossibility for him to give his answer.

It can be tempting to invent values in the absence of judges or because of the cost of the sessions or even because of a problem of product availability while trying to “stick” to the previous results. This is of course contrary to all the rules of professionalism and ethics, but you should also know that this exercise remains very difficult to carry out without other results counteracting and highlighting these suspicious data.

A classification makes it possible to grade or evaluate products using ranks and therefore with a precision linked to the number of possible response modalities, ie the number of products. Thus, products evaluated within a group of three products in classification will be evaluated much more coarsely than some of these products evaluated within a group of 5 or 6 products. A comparison of these two results therefore has very little meaning and meaning.

It is easy to think of being able to repeat for example a series of a triangular test to the judges present in order to be able to use a statistic based on the double of judges. The statistical test must then be adapted with a coefficient of reduction of the power of the test, but this is not recommended, because in the extreme, asking the same question to ten people will never give the same result as asking this same question ten times, to the same person.

Yes, of course, because sensory descriptors are specific questions put to experts in specific conditions, and which have been explained during evaluation sessions – It would thus be difficult to think that a general management audience , marketing, or technical management, is fully aware of the meaning of descriptors. The latter may, in some respects, seem simple to understand, but they can also turn out to be true false friends in the interpretation and analysis of the results.

About sixty panelists constitute a basic element in the standards to be able to begin to interpret the results, and thus be able to draw conclusions on the assessments and preferences of the population or the target studied. This is why redividing this sample into several groups to try to distinguish segmentations of consumption and appreciation seems very tricky due to the too small number resulting for each of the classes constituted.

It is often quite easy and quick to conclude from the differences between two products, that one, for example, is more appreciated than another because its preference rating remains higher, and this indicates a notable trend.

The result of a statistical test is often surrounded beforehand by hypotheses which are often ignored or little taken into account when applying such and such a test; It can also be hypotheses on the structure and distribution of the real population.

A statistical test or a graphic result is often a focus of observation of the data, it remains necessary to consider all of these points of observation to be able to draw robust conclusions on the products – An apparent opposition of the results can result from the applying inappropriate statistics. Consequently, a selective choice of one statistic in favor of another, for the sole reason of convergence with expected results, could turn out to be a serious error or worse ethically, it would be to choose the statistic which would make it possible to obtain any desired result.