SENSORY TESTING -- A STATISTICIAN'S APPROACH 221 perfect (and demonstrably so !) that he is willing to concede that perhaps the difference is not as great as was hitherto thought. It should be noted from Table III, however, that with the same 9/10 probability on the single identi- fication, the likelihood of the single panel member getting 5 out of 6 in a 6+6 test is as high as 0.935. At this point, it is appropriate to make use of wider panels of non- specialist individuals, preferably having nothing to do with the technical a•pects of the product under consideration. It will be possible to establish then whether there are generally detectable differences present it is usual to find that only a proportion of the panel can clearly discriminate. The general finding is that panel members are idiosyncratic in the differences they can detect some panel members can detect certain differences and different sets of panel members are good at other differences. Yet again there are small groups who appear to be good at a very wide range of differences. A hypothesis which would be open to investigation is that these groupings may be genetically determined and could throw light on the mechanisms of the sense of smell. How should the samples be presented? It is clearly of the essence of the logic of the tests described that the differences between the samples should be made manifest only by the sense under test. The test procedure lays down that if the panel member sorts the samples into the correct groups, then there is evidence that the choice was by something other than random selection. It is, however, only by careful attention to the actual mechanism of presenting the samples to the subject that we can be in a position to assume that the only alternative basis for the subject's choice is the use of the relevant test. It we are conducting a test on taste, the food technologist will wish to make sure that we cannot tell Stork from butter by its different colour with an odour test of bath crystals, we shall wish to disguise colour differ- ences in order to be sure that discrimination can only be made by the sense of smell in an auditory test of violins (9), •ve shall wish to conceal the actual Stradivarius from the audience panel by a screen. In the toiletries situation, where we are concerned with perfumes, we find it relatively easy to make comparison between alternative versions of our own products it becomes more difficult if we wish to compare our own product with a competitor. The use of amber-glass jars conceals very minor variations in appearance, of the kind which are sometimes detectable in full light between powders perfumed with different perfume ingredients.

222 JOURNAL OF THE SOCIETY OF COSMETIC CHEMISTS Much more di•cult to conceal are the differences between clear liquids and cloudy suspensions which can arise in making smell comparisons between alternative formulations. In these circumstances it is sometimes necessary to conceal the actual surface of the layer of product by a wire gauze baffle, or a layer of grinding rods or some such inert material. As a final resort, the use of special dim or coloured lighting or even the wearing of dark goggles is effective. If there is any doubt whatsoever, a preliminary test could be to require the subjects to sort the samples wittyout using the sense (taste, smell, etc.) which is really to be tested if the subjects succeed, then the conditions for the test are not appropriate. For the statistician it should not be necessary to draw attention to the fact that the samples under test should be allocated to their code numbers at random. This was forcibly brought home to one statistician who lapsed into a convenient convention of always labelling one group of samples in a 6-•6 test with odd numbers and the other with even numbers. This was convenient because the incorrect items in the sorted groups •vere at once obvious. However, when the test organisers carried out a test in which all twelve samples were poured out of a single bottle of bath additive, it was found that a significant proportion of the panel (at a probability quite beyond the accepted extreme 1 in 1 000 level) were dividing the samples up into odds and evens. It was therefore established that random numbering of samples under test was the only acceptable procedure. Finally, it must be stressed that the form of presentation of the sample to the subject needs very careful consideration. Sensory tests of the kind we are discussing can only handle one aspect of the product at a time we can, for example, carry out tests of the neat smell of a product in the iar, a face cream, a deodorant stick or what you will we may test a shampoo when diluted down in water, or a bath preparation when steaming in hot water, or the residual smell left after using a hair spray, but whatever we test, we must take note of the fact that we are only considering and testing one particular facet of the organoleptic characteristics of the product. One of the more unwelcome tasks of the statistician is to have to draw attention to his clients that they may not generalise from such limited tests, say of the odour of a bath powder in the dry state to the perfume acceptability of the product in general. Too often it proves to be only too convenient to accept the easily obtained answer instead of the relevant one unfortun- ately we so often never get enough information to know how misled we have been.