Statistical validation of sensory data: a study on wine Angela Carlucci and Erminio Monteleone* Dipartimento di Biologia, Difesa e Biotecnologie Agro Forestali, Universita ` degli Studi della Basilicata, C da Macchia Romana, I-85100 Potenza, Italy Abstract: A methodological procedure involving an appropriate statistical validation of sensory data was de®ned in order to describe the typical sensory pro®le of a young red wine Aglianico) destined to aging. A trained panel of eight assessors rated the intensity of nine attributes on 16 products. Sensory data were submitted to statistical validation using a procedure organised in three main steps, namely ®xed and mixed models of analysis of variance ANOVA) and data standardisation. Results of the ®xed ANOVA model computed on raw data showed signi®cant differences between the products for all the attributes except cherry aroma and pungent mouthfeel. However, the results showed inconsistency between assessors. For many attributes, interaction effects were found. When individual differences between assessors were minimised or eliminated using scaled data, results of the ®xed and mixed models showed no signi®cant differences p  0.05) amongst the products for any of the sensory attributes. Therefore the differences amongst the products before data handling and before applying the mixed model of ANOVA were due to the assessors' variability. The results after statistical validationshowedthatthewinesareverysimilarandthatalltheattributesusedtodescribethesensory characteristics of the products can be considered, at a qualitative and a quantitative level, as typical descriptors of the Aglianico wine destined to aging. # 2001 Society of Chemical Industry Keywords: sensory analysis; panel performance; ANOVA; wine INTRODUCTION Descriptiveanalysishasbeenwidelyusedbyresearch- ers to describe the sensory characteristics of food products.Itprovidesqualitativeaswellasquantitative measures of a product's characteristics. The qualita- tivecomponentcomprisesthedescriptiveterms,called attributes, which de®ne the sensory pro®le of the product. The quantitative component measures the degree or intensity of each characteristic perceived to be present. 1,2 Therefore descriptive analysis was proposed in order to obtain an objective characterisa- tion of food products by means of selected sensory descriptors. 3,4 Members of a well-trained laboratory panel are supposed to give small variation in their analytical evaluations. 5 The variations between assessors should be minimised or removed as much as possible before testing,duringtheselectionofpanelmembersorinthe training period, but it is very dif®cult to eliminate the differences completely. 6 When analysing sensory pro®ling data, several problems occur. Even when panellists have been screened and trained well, variationsbetweenassessorsexist.Therearetwomain types of variation that in¯uence the total perception: individual differences in the use of the scale and individual differences in sensitivity, motivation and culture. 7±10 Sensory data always contain information about both these variations. The use of scale is most often of little interest for the experimenter. It is considered merely a nuisance effect telling nothing abouttheproductoraboutthemoreimportantpartof perception. 11 Forinstance,twoassessorscanhavethe sameinterpretationofaproductandstilldecidetouse different scores. In order to validate the information aboutproductperceptionandproductdifferences,itis necessary to know the cause of variability between assessors. Nevertheless, separation of the two types of variation between assessors is generally very dif®cult, because they are always intermingled. 11 There are, however, ways of arranging or preparing pro®le data to take the two effects into consideration. Onesimplewaytoremovethedifferencesintheuseof scaleisdatastandardisation,inwhicheachvariablefor each assessor is scaled to unit variance or any other ®xed value). 11 After removing the effect of scale, it is possibletostartwithathoroughdataanalysis.Analysis of variance ANOVA) is commonly used to analyse sensory data, but in many research papers a descrip- tion of the model of the analysis of variance is often ignored or only very brie¯y mentioned. 5 There are insteadtwomainmodelsofANOVA:the®xedmodel, in which the assessors are considered as ®xed effects, (Received 13 November 2000; accepted 27 February 2001) * Correspondence to: Erminio Monteleone, Dipartimento di Biologia, Difesa e Biotecnologie Agro Forestali, Universita ` degli Studi della Basilicata, C de Macchia Romana, I-85100 Potenza, Italy Contract/grant sponsor: European Union; contract/grant number: POM B35 # 2001 Society of Chemical Industry. J Sci Food Agric 0022±5142/2001/$30.00 751 Journal of the Science of Food and Agriculture J Sci Food Agric 81:751±758 online: 2001) DOI: 10.1002/jsfa.879