Nutrition & Food Science

Statistics for food science V: comparison of many groups (part A)

John A. Bower
Article information:
To cite this document:
John A. Bower, (1997),"Statistics for food science V: comparison of many groups (part A)", Nutrition & Food Science, Vol. 97
Iss 2 pp. 78 - 84
Permanent link to this document:
http://dx.doi.org/10.1108/00346659710161894
Downloaded on: 22 February 2016, At: 06:58 (PT)
References: this document contains references to 15 other documents.
To copy this document: [email protected]
The fulltext of this document has been downloaded 634 times since 2006*
Users who downloaded this article also downloaded:
John A. Bower, (1998),"Statistics for food science V: ANOVA and multiple comparisons (Part B)", Nutrition & Food
Science, Vol. 98 Iss 1 pp. 41-48 http://dx.doi.org/10.1108/00346659810196309
John A. Bower, (1998),"Statistics for food science - V part C: non-parametric ANOVA", Nutrition & Food Science, Vol. 98
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

Iss 2 pp. 102-108 http://dx.doi.org/10.1108/00346659810201087

John A. Bower, (2000),"Statistics for food science – VI: correlation and regression (Part C)", Nutrition & Food Science,
Vol. 30 Iss 4 pp. 194-199 http://dx.doi.org/10.1108/00346650010330243

Access to this document was granted through an Emerald subscription provided by emerald-srm:126741 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for Authors service
information about how to choose which publication to write for and submission guidelines are available for all. Please
visit www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company manages a portfolio of
more than 290 journals and over 2,350 books and book series volumes, as well as providing an extensive range of online
products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee on Publication
Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive preservation.

*Related content and download information correct at time of download.

 Experiments which compare three or more
Statistics for food treatment groups or conditions cannot usually
be analysed using the methods described in
science V: comparison “Statistics for food science (SFS) IV” where
of many groups (part A) two group comparisons were detailed. Such
experiments demand a wider, more versatile
technique which forms the analysis basis of
many simple to complex experimental
John A. Bower designs. The first question concerns possible
incorrect application of two-sample and
paired tests to the many sample group situa-
tion.

Can two-sample or paired tests be

applied to many sample comparisons?
In practice it is possible to perform this proce-
The author dure – the data set from the many-sample
John A. Bower is a Lecturer in Food Science in the experiment is split into pairs, each group
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

Department of Applied Consumer Studies, Queen paired with each other one. A series of the
Margaret College, Edinburgh, UK.
appropriate independent or related sample
tests can then be used, but from a statistical
Abstract
standpoint this approach is not recommend-
Describes statistical methods applied to three or more
ed, and is usually condemned. There is an
sample groups. Discusses analysis of variance in
increased risk of committing a Type I error
parametric forms and the requirement for experimental
(SFS IIIa) and rejecting a null hypothesis
design control before its application.
which is in fact true, i.e. accepting a result as
significant when it occurred by chance. This
arises because of the increased number of
comparisons being performed and applies to
any “multiple testing” circumstance. A statis-
tical test done with a single pair, at the 5 per
cent level involves a 5 per cent risk of commit-
ting a Type I error. This condition applies to
randomly selected samples and subsequent
randomly selected pairs which are indepen-
dent of any other pair. For experiments which
have three or more sample groups, pairing
would result in many more comparisons, e.g.
five treatment groups demand ten inter-
comparisons, not all independent of one
another, and these would be at an increased
risk level; for seven groups 21 tests would be
required and one of these would commit the
error. The procedure may be justified if the
experiment was organized in a paired or two-
sample manner, but this constitutes a differ-
ent type of design which would usually require
a different form of analysis. The separate
treatment “global” approach has organiza-
tional advantages. It is simpler and easier to
present a series of single treatments for analy-
sis or assessment than to prepare many sets of
Nutrition & Food Science
Number 2 · March/April 1997 · pp. 78–84 pairs, especially for sensory evaluation experi-
© MCB University Press · ISSN 0034-6659 mentation in food science.
78
 Statistics for food science V: comparison of many groups (part A) Nutrition & Food Science
John A. Bower Number 2 · March/April 1997 · 78–84

Which analysis method is suitable for which depends on the scale of the experiment,
comparing more than two sample ranging from laboratory scale to full produc-
groups? tion scale. These experimental units are
For such experiments the recommended randomly allocated to the treatments which
statistical design and analysis approach is by are the specific settings or combinations of
analysis of variance (ANOVA). Irrespective of factor level(s). Any source of variation can be
treatment numbers the risk level is main- examined as an experimental factor. A
tained at the specified significance level. Some process condition such as temperature, or an
of the simpler designs where ANOVA is ingredient quantity in a food formulation
applied are given in Table I. Terminology of would be examples of quantitative factors
these designs and tests can vary slightly, where the variable is continuous. The levels of
depending on individual texts. Other impor- these factors would be the individual temper-
tant tests include tests based on chi-square atures or the ingredient weights respectively,
which can be applied to analysis of categorical i.e. different magnitudes of the factor. Factors
data from three or more groups. can also be qualitative, and set at two or more
Although experiments using these tech- levels where magnitude is not considered, e.g.
niques can appear to be straightforward in the type or identity of an ingredient such as
operation, the design features become the type of salt in a formulation. Here the
levels could be sodium chloride and potassi-
increasingly important and form the founda-
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

um chloride. Other possible sources of varia-

tion of many comparative experiments.
tion could include sensory panellists or
panels, individual analysts, instruments,
What does experimental design entail? process machines, laboratories, etc. In certain
experiments involving human subjects as
Experimental design should operate in all
samples, the treatment factors would be in the
tests, trials and experiments within any disci-
form of the different conditions under which
pline where control is required. In its simplest
the groups of subjects were examined. All
form, it is the adoption of an organized
settings for these factors are under control
manner for planning, execution and analysis
and are the independent variables. The
of the work. Initial design steps would be to
response is the dependent variable which is
establish clear objectives for the experiment
measured at the end of the experiment on
and to postulate hypotheses to be tested. In each treatment . One or more responses can
the context of ANOVA, design would include be measured, e.g. the concentration of a
decisions on which factors are being exam- chemical constituent, or the magnitude of a
ined, what levels they are to be set at and the physical, instrumental or sensory parameter.
methodology to be employed for measure- The simplest experiment would involve a
ment of the response. A typical experimental minimum of one factor at two levels but
procedure might begin with division of the beyond this there is theoretically no limit.
food material(s), which is to be experimented Obviously the more factors and levels which
on, into individual lots or batches, the size of are included, the larger the experiment

Table I Some experimental designs analysed by ANOVA (comparison of three or more groups of data)

Some assumptions
Sample Measurement Population Parameter tested/
Design/test nature scale distribution(s) compared
Parametric ANOVA:
Completely randomized Independent Interval Normal Means of groups
design (CRD) 1-way
Randomized block design Related Interval Normal Treatment effect
(RBD) 2-way
Non-parametric:
Kruskal-Wallis 1-way Independent Ordinal Identical except Distribution; median
ANOVA (CRD) for location
Friedman’s 2-way ANOVA Related Ordinal None Distribution; central
(RBD) tendency (median)
79
 Statistics for food science V: comparison of many groups (part A) Nutrition & Food Science
John A. Bower Number 2 · March/April 1997 · 78–84

becomes. Consideration must be given to all As with other statistical tests there are both
other sources of variation and uncontrolled parametric and non-parametric forms of
factors which might affect validity. If possible, ANOVA (Table I). Much of the explanation
these must be controlled or limited using below applies to parametric analysis of vari-
design techniques such as randomization, ance. Parametric ANOVA can also be applied
blocking and replication. Alternatively, there to two data set comparisons in which case it is
are statistical procedures which can circum- equivalent to a t-test, but the commonest
vent these problems or allow some quantifica- application is for three or more groups.
tion of them so that they can be taken into
account in eventual analysis. The importance
What does ANOVA do? – the ANOVA
of replication has been discussed several times
in this series and blocking is dealt with below
method
where design types are examined. Random- The design details above indicate that sources
ization has been proven to reduce error and of variation are assigned under various factors
bias caused by taking a systematic, non- and levels. The technique of ANOVA parti-
random approach to organization of experi- tions this source of total variation into its
mental stages. The technique involves ran- component parts, enabling an assessment of
domization of the order in which stages and the magnitude of effect of each source on the
individual treatments are performed or allo- response(s). Essentially ANOVA compares
cated, e. g. in sensory experiments designs to variability between treatments with that with-
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

counteract order and other bias effects such as in treatments. The between treatment varia-
carry-over should be employed, etc. tion is caused by the nature of the treatments,
Adequate controls or base line measures e.g. the difference in an analytical, instrumen-
should be included so that the test results can tal or sensory measure caused by different
be gauged relative to these. A control treat- food processes or formulations. The within
ment is often represented by an experimental treatment variation is due to the variation in
unit which is actually “untreated”. In an the values for the replicates within each treat-
experiment studying the effect of processing
ment group, caused by several error effects.
conditions the control could be a unit consist-
This error effect is also referred to as the
ing of raw unprocessed food. Space precludes
residual variation, i.e. the “ left over” variation
giving a fuller account of experimental design
which cannot be accounted for by treatment
at present. Some features of an experiment in
food science research[1] which employed a or factor effects. The variance (SFS I) from
number of these design features is illustrated these two sources is used to calculate two
in Table II. different estimates of population variance.
The control for this experiment was a These measures, referred to in ANOVA ter-
formulation which excluded the dairy ingredi- minology as mean squares (MS), are calculat-
ents (the main factor of interest). A large ed for these data groups and their relationship
number of experimental runs were required to is expressed as a variance ratio (F) :
accommodate the many factors and levels
causing the researchers to chose partial MS of treatments
F=
replication, albeit of randomly allocated full MS of error
treatments.

Table II Example of design features of an experiment analysed by ANOVA

Factor Type No. of levels Level identity

Dairy ingredient identity Qualitative 5 Na caseinate, skim-milk,
Na caseinate (high viscosity),
Whey protein,
Demineralized whey
Starch concentration Quantitative 2 2%, 4%
Process temperature Quantitative 2 76°C, 82°C
Experimental units: 5kg batches of smoked meat sausage
Response measures: Sensory properties (15 attributes), instrumental texture (Instron testing instrument),
instrumental colour (Minolta chromometer)
Source:[1]
80
 Statistics for food science V: comparison of many groups (part A) Nutrition & Food Science
John A. Bower Number 2 · March/April 1997 · 78–84

The F ratio for the experiment is compared set there is likely to be more than one source
with tabular F values, calculated for different of variation: one is due to random error at the
levels of significance over a range of experi- measurement stage – end measurement varia-
mental conditions, i.e. the numbers of treat- tion or analytical error; the other is caused by
ments and replicates, expressed as the degrees sampling variation due to inherent variation
of freedom. If the experimental F ratio
within the food material. The magnitude of
exceeds the critical F value (usually at the 5
per cent level or less) then a statistically signif- this latter variation will depend on the food,
icant result is obtained for the data as a whole. e.g. a series of experimental units drawn from
If the null hypothesis is true, i.e. there is no freshly mixed whole milk will be more homo-
evidence to suggest that the population means geneous in chemical composition than those
are not equal, then within the context of the drawn from a day’s harvest of root vegetables.
experiment, the treatments are not demon- Whether or not both these sources of variation
strating a measurable effect. In this case the are measured and hence make an impact on
two estimates of population variance (via the the statistical analysis, depends on procedure.
treatment measures and the background If a treatment is applied once and several
error) will be similar and the F ratio will be
samples are taken for response measurement,
low – theoretically equal to unity.
then the variability of this set of results usually
If the null hypothesis is false and there is a
reflects the analytical variance. Any inherent
real and detectable effect due to the treat-
variation effect will depend on the form of the
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

ments, then depending on the design and

experimental unit at the point of end assess-
precision of the experiment, the treatment
ment. If the unit is made more homogeneous,
variance will be greater in magnitude than the
by the treatment itself or by the end analysis
error variance. Thus, it can be seen that a
maximum for F is obtained when the variabil- method, any inherent variability is likely to be
ity between treatments is high combined with lessened. Inherent variation and “unit to unit”
low variability within treatments (i.e. good variation can be more appropriately estimated
agreement for replicates ). If there is high by comparison with a repeat treatment on a
variability within replicates then F falls; simi- similar experimental unit. The data (Table
larly F falls if there is low variability between III) for an analytical experiment[2] illustrate
treatments, i.e. the treatments may not be this. Replicate end analyses were performed
demonstrating sufficient difference for signifi- on duplicate samples drawn from macerated
cance or detection. An insignificant F ratio single experimental units of raw tissue from
does not prove that the treatments are having different potato tubers. The variability,
no effect – it may be that the experiment was expressed as the percentage coefficient of
too small to detect the differences. A larger variation (%CV; SFS I) is compared with that
experiment would possibly achieve this, but between several experimental units from the
the practical significance of such a smaller same source.
difference would have to be scrutinized – the The treatment per cent CV is much greater
scientific or commercial importance must be due to the relatively large differences in potas-
considered. sium content between different lots of raw
material. Thus, with raw material of this
nature, even before application of treatments
What is the difference between within the experimental units may differ consider-
and between treatment variance? ably and these differences have the potential
It is important to clarify the exact nature of to swamp any treatment effect. In cases of a
replication which gives rise to the within single application of treatments, such circum-
treatment variation. It is possible with some stances may give a false impression of the
forms of design to apply each treatment once experiment’s precision. Apparent significant
and to perform a single response measure- effects between treatments could be due to
ment on each. This procedure would be high inherent variability. Additionally, there is
limited in the extent of analysis due to insuffi- a danger in cases where only one full treat-
cient degrees of freedom. Additionally, within ment is performed: it must be truly represen-
treatment variation cannot be determined tative, i.e. if it were done again would a similar
without replication. Another approach is to result be obtained? If a gross error occurs it
apply each treatment once and draw replicate may not be detected but a duplicate treatment
samples for end analysis. For such a sample could show up the error. To overcome such
81
 Statistics for food science V: comparison of many groups (part A) Nutrition & Food Science
John A. Bower Number 2 · March/April 1997 · 78–84

problems the experimenter must ensure What if the ANOVA assumptions are
adequate sample size and some degree of broken and how can the data be
full treatment repetition within the assessed?
experiment.
As stated previously (SFS IIIa) such assump-
In sensory experiments this procedure is
tions need only be approximate, and proce-
often not followed and it is common practice
dures such as transformation can be applied
to draw multiple servings from single product to allow data to “qualify” in this respect[4] .
batches for panellist assessment. Although The experimenter is recommended to estab-
more work and planning would be involved, it lish some awareness of the status of the data,
is recommended[3] that full replication of especially as they may reveal additional infor-
each production run is included in the design. mation – spurious data, data entry errors,
hidden trends, etc. Visual examination can
often show trends if the data are of small size,
What assumptions does ANOVA
e.g. inconsistency between replicates, obvious
make? differences in treatment means and even
The assumptions for this analysis method are suspicion of non-additivity. Suggested checks
similar to those of the t-test (SFS IV) in terms would include inspection by graphical means,
of normality of distribution and equality of to allow an impression of distribution shape
and characteristics (if sample size permits), as
variance between groups, etc. If the same
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

well as indication of any outliers. Performance

measurement system is used on each group
of some analysis on features such as homo-
(same method, same analyst or same sensory
geneity of variance and normality of distribu-
panel, etc.) then the homogeneity of variance
tion is useful. A measure of agreement within
is a reasonable assumption. Normality of replicates via standard deviation or coefficient
distribution is assumed mainly in the error of variation can quickly ascertain whether this
term, with a random, independent or uncor- variability is within expected ranges. Mea-
related nature to the error. The treatment sures which reveal the uncertainty of the
effect is assumed to be independent of level – sample estimate, such as confidence intervals
this can be infringed in analytical experiments and standard errors, are also recommended.
where the variance may depend on the con- After ANOVA, diagnostic checking by exami-
centration of the analyte. nation of residuals in terms of distribution
A more complex assumption lies in that of and sequence can reveal additional informa-
additivity of factor effects. At a fundamental tion on adherence to assumptions. Signifi-
level the ANOVA method commences by cance tests are available for several of these
specification of a model which describes, in procedures to aid in decision making.
mathematical terms, the effect that treatments In practice, research publications in food
have on the response measure in the experi- science rarely include all of the above checks
in the published material, but such proce-
mental units – the ANOVA model. Essential-
dures may improve publication potential[5].
ly, ANOVA assumes that as treatments are
Modern software packages allow rapid perfor-
applied to the experimental units the eventual
mance of many such techniques.
result on the response will be modification of
the population mean by a linear addition of
these effects plus error. The specific magni- Which form of ANOVA and experimental
tudes of the effects (or coefficients) of the design is applicable?
model can be derived from the experimental If doubt exists regarding the above assump-
data by the process of ANOVA. tions then the results obtained by parametric

Table III Potassium content (mg/100g wet basis) of raw potato for replicate end analyses (atomic absorption
spectroscopy) and replicate treatments (raw controls)

Experimental unit (500g) Replicate end analysis (5g) Mean % CV

Replicate treatment 1 426,422 424 0.67 Average %CV
Replicate treatment 2 389,383 386 1.10 for end analysis
Replicate treatment 3 409,415 412 1.03 = 0.93
Notes: Overall mean = 407
Average % CV for repeated full treatments = 4.77
82
 Statistics for food science V: comparison of many groups (part A) Nutrition & Food Science
John A. Bower Number 2 · March/April 1997 · 78–84

ANOVA can only be viewed as indicative of give more focused attention on the main
differences, and non-parametric ANOVA treatment factor. Thus, on analysis the proce-
should be used if an appropriate design is dure for “2-way ANOVA” is applicable. The
available, otherwise the parametric form block effect can be any factor which is likely to
(which is more powerful) is applicable. Other have a potentially uncontrolled effect on the
important points include the general design response – e.g. if the experiments require to
features of the experiment in respect of be done over several days, then each day could
whether the samples are related or indepen- constitute a block. The blocks could also be a
dent (SFS IV). This depends on the experi- series of different process machines, different
mental objectives and any limits which are laboratories, analysts or indeed, as above,
imposed and needs to be decided at the plan- panellists. In all these cases the variation due
ning stage. to different days, machines, or panellists, etc.
If independent sample groups are warrant- would be calculated and removed from the
ed then a completely randomized design error term. An improved estimate of treat-
(CRD) is applicable with each experimental ment effects is obtained and the RCB design
unit being randomly allocated to a group. In is one of the commonest design types, possi-
its simplest form this design deals with varia- bly because of its relatively simple manner of
tion of one factor and the analysis method is increasing experimental precision.
referred to as “1-way ANOVA”. It is com- Beyond these simple designs the number of
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

monly applied to simple experiments in food factors which can be examined is theoretically
science, involving chemical, physical or unlimited but practical considerations limit
instrumental measure changes caused by most experiments to a selected few. Thus
factors such as processing, etc. For effective there are 3-way, 4-way designs, etc., the com-
use the experimental units must be homoge- plexity of calculation increasing with the
neous or “equally variable” within limits of number of factors. There are distinct advan-
inherent variation. Food experiments where tages in examining the effect of several factors
the response measurement is by objective or together rather than one at a time. This
instrumental techniques can often meet this approach is ultimately more economical of
requirement. In the case of sensory experi- experimental time and resources. In addition
ments, the CRD would require separate to an assessment of main effects, such designs
groups of panellists for each treatment and which include replication allow detection of
they would be assumed to be uniform in their interaction between factors. The presence of
ratings[6]. This is less likely and limits the use significant interaction indicates that factors
of the CRD for much sensory work. Such a are not operating independently but depend
design could apply in a consumer study where on the levels of other factors. More complex
each consumer was only able to be tested designs are available to cover a wide variety of
once, or possibly in storage studies where the experimental demands[7,8]. For non-para-
original panellists were not available for each metric analysis the options are more limited,
assessment stage during storage. The problem and for the more complex designs there may
is avoided by use of a related group design be no readily available non-parametric equiv-
where the panellist effect is viewed as a second alent.
factor. If a CRD is used in sensory studies it is The analysis method and the information
recommended[6] that in order to counteract obtained in all these designs also depends on
panellist variation a large number(> 100) of the exact nature of the treatments. If these are
consumers is recruited. The CRD does not under direct control and are selected by the
require equal numbers within each treatment experimenter they are described as fixed
group which could be a useful feature in the effects and subsequent inferences only apply
above sensory studies. to those factors and levels. A random-effects
The related groups circumstance of a model, with random selection of factors and
single sensory panel assessing several treat- levels, allows inference to a wider population
ments requires a block design such as the of possible factors and levels. The ANOVA
randomized complete block (RCB), with data calculation differs for these two cases and
items in each treatment group having corre- looks at means and variances respectively . A
sponding members in each other group. The mixed-effects model is also possible with both
variation in the blocks could be considered as fixed and random factors. The fixed-effect
another factor but is usually removed so as to model is the more common. Random effects
83
 Statistics for food science V: comparison of many groups (part A) Nutrition & Food Science
John A. Bower Number 2 · March/April 1997 · 78–84

would apply to a sensory study where a 3 Meilgaard, M., Civille, G. V. and Carr, B.T., Sensory
randomly selected panel of consumers was Evaluation Techniques, 2nd ed., CRC Press, Boca
enlisted to measure product preference. The Raton, FL, 1991, pp. 257-62.
random-effects model would allow inferences 4 Steel, R. G. D. and Torrie, J. H., Principles and
Procedures of Statistics – a Biometrical Approach, 2nd
to be made about the consumer population’s
ed. McGraw-Hill International, Singapore, 1981,
preference attitudes. pp.167-71.
Irrespective of the number of factors, 5 Huck, S. W. and Cormier, W. H., Reading Statistics and
replicates, etc. one or more F ratio values are Research, 2nd ed., HarperCollins College Publishers,
obtained, each stemming from a source of New York, NY, 1996.
variation (factor effects). If the F ratio is 6 Gacula, M. C., Design and Analysis of Sensory Opti-
sufficiently large and its P value is equal to or mization, Food & Nutrition Press, Trumbull, CT, 1993,
less than the stated α, then a significant effect pp. 29-34.
has been detected for the experiment as a 7 Box, G. E. P., Hunter, W. G. and Hunter, J.S., Statistics
whole. for Experimenters, John Wiley & Sons, New York, NY,
1978.
8 Cochran, W. G. and Cox, G. M., Experimental Designs,
Does ANOVA identify individual 2nd ed., John Wiley & Sons, New York, NY, 1957.
treatment differences?
Downloaded by Auburn University At 06:58 22 February 2016 (PT)

A significant variance ratio indicates that at Further reading

least two of the treatments differ in respect of
Bender, F.E., Douglass, L.W. and Kramer, A. (Eds), Statistics
the hypothesis being tested. It does not indi- for Food and Agriculture, Food Products Press, New
cate where the significant difference(s) lie and York, NY, 1988.
a pair-wise comparison must be performed. Chatfield, C., Statistics for Technology, 3rd ed., Chapman
This will be detailed in part B along with & Hall, London, 1992.
illustrations of parametric ANOVA and data Cohen, S.S., Practical Statistics, Edward Arnold, Sevenoaks,
evaluation procedures. 1988.
Gacula, M.C. and Singh, J., Statistical Methods in Food and
Consumer Research, Academic Press, Orlando, FL,
References 1984.
1 Baardseth, P., Naes, T., Mielnik, J., Skrede, G., Holland, Miller, J.C. and Miller, J.N., Statistics for Analytical Chem-
S. and Eide, O., “Dairy ingredient effects on sausage istry, 3rd ed., Ellis Horwood, Chichester, 1993.
sensory properties studied by principal component O’Mahony, M., Sensory Evaluation of Food – Statistical
analysis”, Journal of Food Science, Vol. 57 No. 4, Methods and Procedures, Marcel Dekker Inc., New
1992, pp. 822-8. York, NY, 1986.
2 Bower, J. A., “Cooking for restricted potassium diets in Smith, G.L., “Statistical analysis of sensory data”, in
dietary treatment of renal patients”, Journal of Piggot, J.R. (Ed.), Sensory Analysis of Foods, 2nd ed.,
Human Nutrition and Diet., Vol. 2, 1989, pp. 31-8. Elsevier Applied Science, London, 1988, pp. 335-79.

84
 This article has been cited by:

1. Experimental Design 155-178. [CrossRef]

2. Types of Significance Test 114-140. [CrossRef]
3. M.Manuela Hernández-Herrero, Guillaume Duflos, Pierre Malle, Stéphane Bouquelet. 2003. Collagenase activity and protein
hydrolysis as related to spoilage of iced cod (Gadus morhua). Food Research International 36, 141-147. [CrossRef]
4. M.M. Hernandez-Herrero, A.X. Roig-Sagues, E.I. Lopez-Sabater, J.J. Rodriguez-Jerez, M.T. Mora-Ventura. 2002. Influence
of Raw Fish Quality on Some Physicochemical and Microbial Characteristics as Related to Ripening of Salted Anchovies
(Engraulis encrasicholus L). Journal of Food Science 67:10.1111/jfds.2002.67.issue-7, 2631-2640. [CrossRef]
Downloaded by Auburn University At 06:58 22 February 2016 (PT)