Sensory Evaluation of Food

Statistical Methods and Procedures

Michael O'Mahony

• Publisher: CRC Press
• Year: 2017
• ISBN:
  9781351416894
• Edition: 1
• Topic: Technology & Engineering
• Subtopic: Food Science
• Index: Technology & Engineering

Sensory Evaluation of Food: Statistical Methods and Procedure covers all of the basic techniques of sensory testing, from simple discrimination tests to home use placements for consumers. Providing a practical guide to how tests are conducted, the book explores the fundamental psychological and statistical theories that form the basis and rationale for sensory test design. It also demonstrates how statistics used in sensory evaluation can be applied in integrated applications in the context of appropriate sensory methods, as well as in stand-alone material in appendices. Offering a balanced view of diverse approaches, this is an essential guide for industry professionals and students.

1

Before You Begin

1.1 How to Use this Book

This book presents a simplified look at statistics, as do all introductory texts. In the interests of clarity and simplicity some shortcuts will be taken, but wherever possible, these will be indicated. Although the study of mathematical statistics is clear enough, its application to real-life situations is not always so clear. Statistical tests involve, in their derivation, certain assumptions about the data and it is sometimes doubtful whether these assumptions are always fulfilled in behavioral measurement. If they are not fulfilled, it is not always clear how important this may be. For instance, in some experimental situations, should a one- or a two-tailed test be employed? How much do data from intensity scal#s violate the assumptions for analysis of variance, and how important are such violations? In what situations are nonparametric tests more desirable? These matters will be discussed in more detail throughout the text; suffice it to say that in the area of sensory and behavioral measurement such doubts should be taken seriously. As far as possible, we will give recommendations and try to indicate where there is disagreement. However, this text is not a statistical treatise; it is not intended to give an exhaustive account of these controversies—for that, go and get a degree or two in statistics! The aim of this text is to give a working knowledge of statistical analysis with as few frills as possible. Enough background will be given so that the reader can understand what a- statistical test is doing and when and how it should be used. The arguments and logic will always be expressed as nonmathematically as possible.

The mathematical formulas will be expressed simply; formal mathematical notation with its welter of subscripts will not be used here. If it offends the mathematical purist, I apologize, but a simpler, albeit more casual approach makes teaching simpler at this level. So be warned that the notation used here will not always be the standard notation used by statisticians. We will not bother to put all algebra referring to populations in capital letters and that referring to samples in lowercase letters. We will not bother to identify each value with its own subscript; that only terrifies the reader and is bad teaching practice. Furthermore, although most of the statistical tests will be illustrated by examples from sensory measurement and the sensory analysis of food, this principle will not be adhered to slavishly. Sometimes a statistical principle is better illustrated by other examples, and in the interests of clarity and better teaching, these will be adopted. For example, some probability theory is more simply explained using familiar examples such as tossing of dice and coins.

Unfortunately, there is little agreement among the available texts on the best approach to explaining the application of various statistical tests to behavioral measurement; there is even disagreement in such texts over the names to be given to various tests. The babel is best seen by a consideration of all the different names given in behavioral texts to the various analysis-of-variance procedures (Table G.10 in Appendix G). In this text, we will give statistical tests the simplest yet most fully descriptive names possible; we will also keep the use of jargon to a minimum—we are not trying to impress anyone.

Finally, this text is brief, sparse, and not very conversational. It is hoped that the starkness will eliminate distraction and allow the reader to stick strictly to the point. It is hoped that this will speed the process of learning without leaving the reader with a sense of desolation.

1.2 The Use of Statistics in Sensory Evaluation Versus its Use in Psychophysics, Experimental Psychology, and Consumer Testing

The experimental and statistical skills used for the sensory analysis of foods or consumer testing are very similar to those used in experimental psychology. For further study, texts on psychological statistics are recommended. There are, however, some important differences between the application of statistics to sensory evaluation and its application to psychology or consumer testing. These are due to the different goals and approaches of the various disciplines. Often the differences are not important, but sometimes they are crucial because they can alter the actual mathematical operations involved in statistical tests; for example, analysis of variance. So, dear reader, if you are an experimental psychologist or a consumer researcher, about to use this text as a simple reminder, or if you are a sensory analyst about to consult a psychology text, please read this section beforehand to avoid confusion.

First, let us look at differences between sensory evaluation and experimental psychology, of which sensory psychophysics is an important part. Generally, in experimental psychology the goal is to find out something about the behavior of people in general, or even of animals. In sensory psychophysics, specifically, the goal is to examine the functioning of the senses and the brain, and the focus is generally on people. In both these areas, the general strategy is to take a random sample from the population of people under consideration. The sample is examined in a suitable experiment, the appropriate statistical analysis is performed on the sample, and using this analysis, conclusions are drawn about the population. A sample is used because there is generally not enough time or money to examine the whole population. The sample is generally picked as randomly as possible in the hope that it will match the population as closely as possible.

In the sensory evaluation of foods or other products, there can be more than one goal. First, in some situations, people may be used for a preliminary analysis to measure the flavor of the food and provide clues for later chemical or physical analysis. They may be used to measure, say, flavor changes due to a change in the processing of the food. The question here is: Does the new processing procedure alter the sensory characteristics of the food? Changes in sensory characteristics can provide clues about changes in physical characteristics. Should judges detect a change in the odor, it would provide a clue to chemists to investigate changes in the volatiles. Should judges detect a change in sourness it would indicate that chemists might first want to investigate changes in acid content or even sugar content. The goal is to seek information about the food, not the people who taste the food. A random sample of people is not selected. A panel comprising of a few people is selected and trained to become judges in the specific task required, whether it be detecting off-flavors in beverages or rating the texture of cakes. Generally, potential judges are screened to see whether they are suitable for the task at hand. This careful selection of judges is in no way a random sampling and is the first reason why the judge could not be regarded as representative of the population in general. The training then given to the selected judges will usually make them more skilled in their task than the general population would be; this is a second reason why they can no longer be regarded as representative of the population from which they were drawn. Our judges are not a sample to be examined so that conclusions can be drawn about a population; they are the population! They are a population of laboratory instruments used to study samples of food. On the other hand, the food samples tested are samples from which conclusions can be drawn about the populations of foods. They may be samples of a food processed in different ways from which conclusions may be drawn about the whole population of that food processed in these ways. So the focus of the study is very much on the food; the judges are merely instruments used for measuring samples of food. Logically, one good judge would be adequate for this purpose, as is one good gas chromatograph or atomic absorption spectrometer. More than one judge is used as a precaution because unlike a gas chromatograph, a judge can become distracted; the cautious use of second opinions can provide a useful fail-safe mechanism.

Having detected a change in flavor under such circumstances, the question becomes: Can this flavor change be detected by the consumers or people other than trained panelists? It is possible that the flavor change is so slight as to be detectable only by highly trained judges. One way of answering this question is to take a sample of untrained people and test them to see whether they can detect the flavor change. Then from this sample, inferences can be made about the population of untrained people in general. The experiment now becomes similar to one in sensory psychophysics. People are sampled, tested, and conclusions are then drawn about people in general. This is the second use of sensory evaluation.

It can thus be seen that in the sensory evaluation of food, depending on the aims of the study, conclusions may be drawn only for those judges actually tested or they may be drawn for the whole population of people from which the judges were sampled. It is reasonable to expect the probabilities and the statistical procedures to vary for these two cases. The difference becomes important for analysis of variance, where different denominators are used for calculating F values. Where the conclusions apply only to the judges tested, the judges are said to be a fixed effect. Where they apply to the whole population from which the judges were drawn, the judges are said to be a random effect. In this text, the analysis-of-variance procedures all use people as fixed effects because they are easier to understand and learn that way. To use people as random effects when the aim is to study people per se, the calculation is altered very slightly. The brief chapter (Chapter 13) on fixed- and random-effects models shows the slight differences required when people are “random effects.” Be sure to read it before you perform an analysis of variance on data designed to give you facts about people in general rather than only those people tested in the experiment. It is worth noting that texts on psychological statistics differ from this text. In psychological texts people are generally treated only as random effects.

The traditions of psychophysics and sensory evaluation vary even more. In psychology or psychophysics, people are the subject of the investigation and thus tend to be called subjects. In sensory evaluation, or sensory analysis as it is also called, people are often specially selected and trained and tend to be referred to as judges. In sensory analysis, judges are tested while isolated in booths; experimenters and judges communicate in writing. In sensory psychophysics, the experimenter and subject often interact and communicate verbally; this requires special training for experimenters so that they do not influence or bias the subjects. The methods of sensory analysis are often derived from those of psychophysics, but care must be taken when adapting psychophysical tests to sensory evaluation. This is because the goals of the two disciplines are different and they can affect the appropriateness of various behavioral methods of measurement to be used, as well as the statistical analysis employed. In psychophysics, the aim is to measure the “natural” functioning of the senses and cognitive processes. Extensive training will not be given if it alters a subject’s mode of responding or recalibrates the person, so obscuring his or her “natural” functioning.

Consumer testing must also be considered. Whereas judges are used in sensory analysis to assess the flavor of a food per se, consumer testing selects samples of consumers from the marketplace to find out whether they like or will buy the food. For consumer testing, the people must be as representative as possible of the potential buying population so that they do not give a misleading indication of potential sales. To perform consumer tests on members of a panel used for sensory analysis would be to use a sample that was not representative of the potential buying public. The panel would be comprised of a few people who had been specially selected for their skill in sensory evaluation and who may also have been made atypical as a result of their training. There is another important point: The judges on a sensory panel are usually recruited from the company; they are often laboratory or office workers who take time from their regular duties to participate in the panel. These judges may have a knowledge of the product which may bias their judgments of whether they like it; they may even select responses that they feel might impress their bosses. Thus, to use a sensory panel for consumer testing would seem unwise. Despite this, some food companies blindly adopt this research strategy.

The sample of consumers tested should be as representative as possible of the potential buying population. It is a sample from which conclusions are drawn about the population and so, as with sensory psychophysics, the people tested are considered to be random effects! Psychophysicists tend to pick a sample as randomly as possible in the hope that it will turn out to be representative of the population. The strategy of random sampling is used in consumer testing, but another strategy is also adopted. Here the demographics of the population are studied and a sample carefully chosen so as to match the population demo-graphically as closely as possible. If the population has 80% males, so will the sample. If the population has 50% Hispanics, the sample will also. This approach to making the sample as representative as possible of the population is called stratified sampling. Sampling will be discussed further in the section on the perils of sampling (Section 1.4).

As you read through the text, it will become apparent that statistical tests are set up to show that differences exist between sets of data. However, should the reader be using statistics for quality assurance, trying to show that a reformulated product has the same flavor as the regular product, the aim will be to show that differences do not exist. This is using statistical analysis in a rather different manner, and readers are advised to consult Appendix F before analyzing their data.

1.3 Why use Statistics Anyway?

In Section 1.2 we mentioned the various disciplines that use statistics. We discussed their sampling techniques and whether inferences made from samples were carried to populations. To summarize, psychophysicists, experimental psychologists, and consumer testers are interested primarily in the behavior or likes and dislikes of people. They make their measurements on a sample of people, for reasons of time and economy, and use a statistical analysis to make inferences about the population from which the sample was taken. Naturally, the larger the sample, the more representative it is of the population. The more powerful the statistical test applied to the sample, the better the information obtained about the population. This is true, of course, as long as all the conditions required for the proper application of the test are fulfilled. There are procedures for estimating how large a sample should be taken for a given test of a given power, but these will not be dealt with here. Sensory analysts select and train a small panel to assess samples of foods. Conclusions resulting from the statistical analysis may or may not be generalized further beyond the actual panelists tested; the panelists may be random or fixed effects. In the same way, conclusions may or may not be generalized beyond the actual food servings tested. Replicate servings of a given food processed in a given way (often called “replicates” or “reps”) can themselves be seen as a fixed or a random effect, depending on how far the conclusions are to be generalized.

Statistical tests may be inferential; inferences are made about populations from samples. Statistical procedures may also be merely descriptive; the data in the sample may merely be summarized in some way with no inferences being made about the population. Thus a mean is a summary of the data in a given sample; it gives a central summary value of the data. In this way, it is descriptive. However, the best estimate of the mean of a population is also the mean of the sample. In this way, it is inferential.

It is important when taking a sample or designing an experiment to remember that no matter how powerful the statistics used, the inferences made from a sample are only as good as the data in that sample. Furthermore, the inferences made from a sample apply only to the population from which the sample was taken. This can be crucial in medical or consumer-orientated research.

Some Definitions and Jargon

Descriptive statistics are used to describe the data obtained: graphs, tables, averages, ranges, etc.

Inferential statistics are used to infer, from the sample, facts about the population.

A parameter is a fact concerning a population.

A statistic is a fact concerning a sample.

Data is plural; the singular is datum.

1.4 Some Perils of Sampling

Generally, statistical tests require that a sample be selected randomly from the population it is representing. This may be a complex matter but there are additional points of design which are worth mentioning here.

When sampling, you must select your sample in an unbiased manner. If you wish to take a sample of university students and examine the sample to see whether there are more men than women in the university, it is no use standing outside the men’s toilet. You will only get the occasional brave, or lost, lady and so have a biased sample. This type of thing happened when in 1936 the Literary Digest polled 10 million Digest readers and phone subscribers as to who would win the U.S. presidential election. The poll indicated Landon, but in fact Roosevelt was elected. The Literary Digest had taken a biased sample. Only the richer people read the Literary Digest or had phones at that time. The poorer Americans tended not to be included in the poll; it was their vote which elected Roosevelt.

So take care how you collect your data. If the sample is chosen so as to be representative of the population, the conclusions drawn f...