DISCUSSION II ON FITTING EQUATIONS TO SENSORY DATA

In his article " On Fitting Equations to Sensory Data." Moskowitz suggests many strategies for model fitting which depart from current statistical methodology. Four areas discussed by Moskowitz are addressed here: (1) Forcing terms into a model; (2) The use of hold-out samples; (3) The use of aggregate data (averaging across people, suppressing the person-to-person variation); and (4) The use of random data as a predictor variable in a regression equation. All four of these points will be examined within this article.     

ON FITTING EQUATIONS TO SENSORY DATA: A POINT OF VIEW, AND A PARADOX IN MODELING AND OPTIMIZING

When fitting equations to data relating ingredients or factor scores to subjective ratings, there are at least two methods to create the equations. One method, (linear) forces in linear terms and allows additional quadratic and interaction terms. The other method, (quadratic) forces in linear and quadratic terms, and then permits cross-terms to enter. The two methods produce contradictory results. The first (and expected) contradiction is that the quadratic model shows optimum levels in the middle range of the levels tested for some, but not all, ingredients. The second (and unexpected) contradiction is that the linear method (which usually does not incorporate many additional terms) generates better validation predictions for hold-out samples than does the quadratic method. The differences between the optimum generated by the linear model and the optimum generated by the quadratic model can be quite substantial in terms of expected liking, sensory profile, and image profile.     

RELATING SENSORY AND CHEMICAL DATA TO UNDERSTAND FLAVOR

Flavor cannot be measured directly by instruments, it is an interaction of consumer and product. A gas chromatogram, even with 250 peaks on it, does not say anything about the flavor, only about the volatiles. In some cases a character impact compound might be present, in a second group of products a small number of compounds may be responsible for the majority of the flavor, while a third class contains more complex foods, which have generally been processed in some way. A range of methods has been, more or less, successfully used to try to link composition data to flavor data, including the calculation of odor units, fractionation of chromatographic effluent, “nasal appraisal” of chromatographic effluent, and a range of multivariate statistical procedures. However, defects in understanding of the mechanism of operation of the chemical senses has limited the success which has been achieved.     

A SIMPLE ALTERNATIVE TO GENERALIZED PROCRUSTES ANALYSIS: APPLICATION TO SENSORY PROFILING DATA

A statistical method for analyzing sensory profiling data obtained by means of fixed vocabulary or free choice profiling is discussed. The most interesting feature of this method is that it involves only simple statistical treatment and can therefore be performed using standard software packages. The outcomes of this method are compared to those of Generalized Procrustes Analysis on the basis of two data sets obtained, respectively, by means of fixed vocabulary and free choice profiling. A significance test is also discussed in order to assess whether the overall configuration of the products is meaningful. This significance test is based upon a simulation study involving the permutation procedure.     

NOTE ON COMPUTERIZED DATA COLLECTION IN CONSUMER SENSORY EVALUATION

Panelists with computer experience (n = 24) and without (n = 26) rated three foods representing liked, neutral and disliked categories using a 9-point hedonic scale on computer and paper ballots. In paired t-test results between computer and ballot ratings, no differences were found in overall acceptance, flavor and appearance attributes. Regression analysis indicated a significant (p<0.01) linear relationship between computer and paper ballot ratings. Computer ballot ratings for overall acceptance were comparable to paper ballot ratings in a consumer sensory evaluation test. Slight differences in overall acceptance were found between computer and paper ballot responses from panelists with and without computer experience. GLM indicated the group without computer experience rated overall acceptance and flavor attributes for cereal, and appearance and flavor attributes for potato chips significantly higher (p<0.05) than panelists with computer experience.     

NONPARAMETRIC ANALYSIS FOR TRINARY SENSORY DATA

Sensory evaluation data is sometimes collected using trinary category scales. Nonparametric data analysis for such data is discussed and a homogeneity statistic for trinary data is proposed which is simple enough to calculate with a pocket calculator. A statistic for identifying market segmentation in trinary data is also suggested.     

REDUCING THE NOISE CONTAINED IN DESCRIPTIVE SENSORY DATA

A data reduction protocol was designed to minimize distortion inherent in sensory data. Following removal of nonexistent attributes and treatment levels, extreme value analysis and distribution comparisons combined with graphical respresentations, facilitated elimination of inconsistent (with respect to overall consensus) panelists. Application of a calibration factor showed superresponsive panelists (those with intensity values consistently higher than other panelists) were among the most accurate and thus were retained in spite of their tendency to produce extreme value data. Panelists that consistently produced a narrow variance around the overall mean and rarely produced extreme values were classified as noncomittal and removed. Analysis of variance calls for a split plot design; blocks (sessions) and treatments in main plot, and panelists in subplot. In general, the subplot can be ignored. These methods are suggested for evaluating panelists' training needs; and for eliminating data that distorts the statistical analysis.     

A FORTRAN PROGRAM, 'SIGMPLOT', FOR FITTING SIGMOID CURVES TO THRESHOLD DATA AND FOR PLOTTING THE RESULT

A computer program is presented for determining position andslope of a symmetrical sigmoid curve which fits to experimentaldata obtained in experiments with absolute or discriminationthresholds. If a suitable plotter is available, the programprovides a plotting of experimental data and the fitted curve.50 % and 75 % threshold values and the difference between these(in certain cases equal to the ‘jnd’) are alwaysprinted. Available options and necessary limitations are described.     

A NON-PARAMETRIC RANKING METHOD FOR THE STATISTICAL EVALUATION OF SENSORY DATA

Sensory data are rarely normally distributed and should, therefore,be statistically analyzed by non-parametric techniques. A computerprogram was devised for generating tabular data by which upto twenty samples can be compared following evaluation by asmany as seventy-five panelists. When samples are initially ranked,rank sums for each sample may be compared to appropriate entriesin the tables directly. If scalar values are assigned by panelists,the technique is still useful, but the values must first beconverted to ranks. ^*Scientific Article No. A1857, Contribution No. 4767 of theMaryland Agricultural Experiment Station (Food Science Program).     

SENSORY EVALUATION BASED ON VERBAL JUDGMENTS

Studies of the repeatability and the homogeneity of expert panel scores in sensory profiling show that lasting and reliable evaluations of food products are difficult to obtain: strong inter- and intra-individual differences are commonly observed. Our hypothesis is that this variability is due to quantification methods that consist of asking panelists to furnish quantitative values (by attributing a numerical point to perceived intensity) and that using natural language in the form of verbal judgements in a hierarchical tree would allow improving the reliability of sensory evaluations. This hypothesis was tested by comparing a numerical value scale and a specific hierarchical semantic scale that subjects devised themselves prior to the experiment. The products (chocolate dessert creams) were evaluated using each of the two methods. The results show that the capacity to discriminate between products is better when verbal judgements are used. We discuss the advantages of hierarchical semantic scales for sensory profiling.