GROUPING OF WINES ACCORDING TO BODY BY STATISTICAL ANALYSIS OF CHEMICAL AND ORGANOLEPTIC PARAMETERS

Size: px

Start display at page:

Download "GROUPING OF WINES ACCORDING TO BODY BY STATISTICAL ANALYSIS OF CHEMICAL AND ORGANOLEPTIC PARAMETERS"

Horace Payne
6 years ago
Views:

1 GROUPING OF WINES ACCORDING TO BODY BY STATISTICAL ANALYSIS OF CHEMICAL AND ORGANOLEPTIC PARAMETERS P. Lehtonen (1), R. Sairanen (1), J. Alin (2), E. Ali-Mattila (1), J. Meriläinen (1), J. Sopenperä (1) (1) Alko Inc. P.O. Box 33, FI-00181, Helsinki, Finland (2) Numos Ltd. Tekniikantie 12, FI Espoo, Finland (2) Present address: SAS Institute Oy Tekniikantie 14, P.O. Box 85, FI Espoo, Finland ABSTRACT The oakiness, colour and body of about 600 red wines from all over the world were evaluated by sensory analysis on a scale of -5 to +5. The same wines were analysed for ethanol, total acid, volatile acid, glucose, fructose and glycerol concentration as well as density and dry extract. The wines' ph values and absorbances at 280, 420, 520, 540, 610 and 620 nm were also determined. The data were subjected to cumulative logistic regression and principal component analysis (PCA). The simplest model explaining the body of a wine contained four chemical parameters: ethanol concentration, absorbance at 420 nm, total acid concentration and ph (p<0.01). Body also correlated well with the other two organoleptic parameters: oakiness and colour. Red wines sold by the Finnish alcohol retail monopoly, Alko Inc., were placed in four groups according to the body of the wine. Customers are able choose among light-bodied wines, medium-bodied wines, full-bodied wines and extra full-bodied wines. The body of any new red wine in the Alko range will be determined by sensory analysis and verified by chemical and statistical methods. RESUME Nous avons évalué par analyse sensorielle le boisé, la robe et le corps d environ 600 vins rouges du monde entier, sur une échelle de -5 à +5. Pour les mêmes vins ont été analysés : les teneurs en éthanol, acides totaux, acides volatils, glucose, fructose et glycérol, ainsi que la densité et l extrait sec total. Les valeurs de ph et les absorbances des vins à 280, 420, 520, 540, 610 et 620 nm ont également été déterminées. Les données ont été soumises à une régression logistique cumulative et à une analyse en composantes principales (ACP). Le plus simple modèle décrivant le corps d un vin contient quatre paramètres chimiques : la teneur en éthanol, l absorbance à 420 nm, la teneur en acides totaux et le ph (p<0,01). Il y a également une bonne corrélation entre le corps et les deux autres paramètres organoleptiques : le boisé et la robe. Nous avons classé les vins rouges vendus par le monopole finlandais de vente d alcool Alko Inc. en quatre groupes, d après le corps du vin. Les clients sont alors en position de choisir entre des vins légers, moyennement corsés, corsés et très corsés. Le corps de tout vin rouge nouveau dans la gamme Alko sera déterminé par analyse sensorielle et vérifié par les méthodes chimiques et statistiques.

2 INTRODUCTION Until recently, the retail outlets of Alko Inc. placed their red wines in three groups according to body: light-bodied wines, medium-bodied wines and full-bodied wines. Most of the red wines, 65%, were found in the medium-bodied group, with only 3% in the light-bodied group and 32% in the full-bodied group. To improve its customer service, Alko Inc. has now added a fourth group, extra full-bodied wines, to its selection. This has meant that all of the about 600 red wines in the Alko range have had to be re-evaluated and re-grouped according to body. MATERIALS AND METHODS The oakiness, colour and body of 639 red wines from all over the world were evaluated by sensory analysis on a scale of -5 to +5. A value of -5 indicated the lowest level of the measured property and +5 indicated the highest level. In the case of body, for instance, a very light wine may have received a value of -5 and a very full-bodied wine a value of +5. The sensory panel consisted of five to ten professional tasters. The same wines were analysed for ethanol, total acid, volatile acid, glucose and fructose concentration as well as density, ph and dry extract using a FOSS WineScan FT120 instrument (FOSS, Hillerød, Denmark). The sum concentration of glucose and fructose was then subtracted from the dry extract to obtain dry extract without sugars. Glycerol was determined using enzyme kits from Boehringer Mannheim (Indianapolis, Indiana, USA). Glycerol concentrations were expressed in milligrams per litre and also relative to the ethanol concentration (glycrel) in the beverage. Relative glycerol concentration was calculated according to Eq. 1, where c g = absolute glycerol concentration (mg/l), p = ethanol percentage w/w and ρ = density (kg/m 3 ). The limit of quantification of glycerol was 10 mg/l. glycrel = (c g * 10)/(p * ρ) (1) The red wine samples were diluted 1:100 with distilled water and their absorbances at 280, 420, 520, 540, 610 and 620 nm were determined using a Perkin-Elmer LAMBDA 25 spectrophotometer (PerkinElmer Life And Analytical Sciences, Inc., Waltham, Massachusetts, USA). Thus, there were in all 17 chemical parameters to be used in statistical processing. Statistical analyses were carried out to support the task of grouping the red wines according to body. The data were subjected to cumulative logistic regression and principal component analysis (PCA) using SAS Enterprise Guide 4.1 (SAS Institute, Cary, North Carolina, USA) and its PROC GENMOD procedure for modelling.

3 RESULTS AND DISCUSSION As the aim of the statistical analyses was to allow the red wines to be divided into four groups instead of the former three, the analyses were performed more on practical than scientific grounds. It was first studied whether a cumulative regression model with chemical parameters as independent variables could explain the subjective body grouping made by human wine tasters. Cumulative regression modelling was chosen because it has quite good interpretability and allows the modelling of cumulative categories, for which the body of red wine is well suited. The first models were constructed starting from a model where the estimate of wine body from the subjective tasting is the response variable and all the chemical parameters constitute independent variables. Two-level interactions of chemical parameters were also included in the first model. The least significant variables and interactions were removed from the model by manual backward elimination, finally arriving at a very simple model of just four independent variables, each significant at the 0.05 level. Based on normal model validation statistics, the models created from the initial set of data fitted the data very well with an explanatory power of close to 80%. The simplest model explaining the body of a wine contained the following four chemical parameters: ethanol concentration, absorbance at 420 nm, total acid concentration and ph (p<0.01). The statistical model can be simplified to Eq. 2: body of red wine = ethanol concentration + total acids + ph + absorbance at 420 nm (2) Body also correlated well with the other two organoleptic parameters: oakiness and colour. The grouping obtained by the simplest statistical model consisting of four chemical parameters closely matched that derived from the organoleptic analysis, as can be seen in Tab. 1 and Fig. 1. Since data gathering was ongoing when the model was chosen, and new data will be arriving from tasting of new wines and new vintages, the modelling process has been automated to predict the body category from new observations of chemical parameters, utilising the historical linkage between organoleptic wine body and chemical parameters. This predicted wine body is then compared with the actual organoleptically observed wine body, and if the predicted and observed wine bodies are too far apart, a new tasting may be arranged. When the body category for the new wine has been settled, the observation data are added to the history data, thus further strengthening the modelling and verification process. The process is implemented to provide more stable and justified categorisation of red wines according to body. The initial re-tasting resulted in about half the wines being re-grouped according to the chemical evaluation of wine body. After the procedure, 7%, 43%, 42% and 7% of the red wines in the Alko range were characterised as light-bodied, medium-bodied, full-bodied and extra fullbodied, respectively.

4 Table 1. Statistical analysis of 629 red wines for body. The model was based on the determination of ethanol concentration, absorbance at 420 nm, total acid concentration and ph. LB = light-bodied wines, MB = medium-bodied wines, FB = full-bodied wines, EFB = extra fullbodied wines Statistical analysis of chemical Organoleptic parameters estimation LB MB FB EFB LB n % MB n % FB n % EFB n % % LB MB FB EFB Organoleptic grouping LB MB FB EFB Fig. 1. Distribution of 629 red wines according to body by statistical analysis of ethanol concentration, absorbance at 420 nm, total acid concentration and ph. LB = light-bodied wines, MB = medium-bodied wines, FB = full-bodied wines, EFB = extra full-bodied wines

5 CONCLUSIONS Red wines sold by the Finnish alcohol retail monopoly, Alko Inc., were placed in four groups according to the body of the wine. Customers will now be able choose among light-bodied wines, medium-bodied wines, full-bodied wines and extra full-bodied wines. The body of any new red wine in the Alko range will be determined by sensory analysis and verified by the abovementioned chemical and statistical methods. If the group indicated by chemical methods differs significantly from that obtained by sensory analysis, the sensory analysis will be repeated. BIBLIOGRAPHY Allison P., Logistic Regression Using SAS : Theory and Application. Cary: SAS Publishing. Ribereau-Gayon P, Glories Y, Maujean A, Dubourdieu D, Handbook of Enology: Volume 2: The chemistry of wine stabilization and treatments, 2nd edition. Chichester: John Wiley & Sons.

Wine-Tasting by Numbers: Using Binary Logistic Regression to Reveal the Preferences of Experts

Wine-Tasting by Numbers: Using Binary Logistic Regression to Reveal the Preferences of Experts When you need to understand situations that seem to defy data analysis, you may be able to use techniques